diff --git a/.github/workflows/check-vendor.yml b/.github/workflows/check-vendor.yml index b9e8ac7658..1671ed7b8b 100644 --- a/.github/workflows/check-vendor.yml +++ b/.github/workflows/check-vendor.yml @@ -19,7 +19,7 @@ on: jobs: check-vendor: - runs-on: ubuntu-latest + runs-on: ubuntu-slim steps: - name: Checkout diff --git a/.github/workflows/close-issue.yml b/.github/workflows/close-issue.yml index 8fb5310d0b..ec3df08b2d 100644 --- a/.github/workflows/close-issue.yml +++ b/.github/workflows/close-issue.yml @@ -10,7 +10,7 @@ permissions: jobs: close-issues: - runs-on: ubuntu-latest + runs-on: ubuntu-slim permissions: issues: write pull-requests: write diff --git a/.github/workflows/editorconfig.yml b/.github/workflows/editorconfig.yml index a5cd590017..702dc89f5b 100644 --- a/.github/workflows/editorconfig.yml +++ b/.github/workflows/editorconfig.yml @@ -20,7 +20,7 @@ concurrency: jobs: editorconfig: - runs-on: ubuntu-latest + runs-on: ubuntu-slim steps: - uses: actions/checkout@v6 - uses: editorconfig-checker/action-editorconfig-checker@v2 diff --git a/.github/workflows/gguf-publish.yml b/.github/workflows/gguf-publish.yml index 5bdab0f157..0e95766459 100644 --- a/.github/workflows/gguf-publish.yml +++ b/.github/workflows/gguf-publish.yml @@ -21,7 +21,7 @@ on: jobs: deploy: - runs-on: ubuntu-latest + runs-on: ubuntu-slim steps: - uses: actions/checkout@v6 diff --git a/.github/workflows/labeler.yml b/.github/workflows/labeler.yml index 42f00c0cd8..eab20c6881 100644 --- a/.github/workflows/labeler.yml +++ b/.github/workflows/labeler.yml @@ -7,7 +7,7 @@ jobs: permissions: contents: read pull-requests: write - runs-on: ubuntu-latest + runs-on: ubuntu-slim steps: - uses: actions/checkout@v6 with: diff --git a/.github/workflows/pre-tokenizer-hashes.yml b/.github/workflows/pre-tokenizer-hashes.yml index 8120df0e36..7126b62b69 100644 --- a/.github/workflows/pre-tokenizer-hashes.yml +++ b/.github/workflows/pre-tokenizer-hashes.yml @@ -12,7 +12,7 @@ on: jobs: pre-tokenizer-hashes: - runs-on: ubuntu-latest + runs-on: ubuntu-slim steps: - name: Checkout repository diff --git a/.github/workflows/python-check-requirements.yml b/.github/workflows/python-check-requirements.yml index 08cdcb9d01..1219b87459 100644 --- a/.github/workflows/python-check-requirements.yml +++ b/.github/workflows/python-check-requirements.yml @@ -20,7 +20,7 @@ concurrency: jobs: python-check-requirements: - runs-on: ubuntu-latest + runs-on: ubuntu-slim name: check-requirements steps: - name: Check out source repository diff --git a/.github/workflows/python-lint.yml b/.github/workflows/python-lint.yml index 91dc4d78a4..8d1dd7a7d5 100644 --- a/.github/workflows/python-lint.yml +++ b/.github/workflows/python-lint.yml @@ -15,7 +15,7 @@ concurrency: jobs: flake8-lint: - runs-on: ubuntu-latest + runs-on: ubuntu-slim name: Lint steps: - name: Check out source repository diff --git a/.github/workflows/python-type-check.yml b/.github/workflows/python-type-check.yml index 54d5fab5ba..e801a9f42e 100644 --- a/.github/workflows/python-type-check.yml +++ b/.github/workflows/python-type-check.yml @@ -29,9 +29,7 @@ jobs: uses: actions/setup-python@v6 with: python-version: "3.11" - - name: Install Python dependencies - # TODO: use a venv - run: pip install -r requirements/requirements-all.txt + pip-install: -r requirements/requirements-all.txt - name: Type-check with Pyright uses: jakebailey/pyright-action@v2 with: diff --git a/.github/workflows/update-ops-docs.yml b/.github/workflows/update-ops-docs.yml index 40447db4e4..2ab06eb981 100644 --- a/.github/workflows/update-ops-docs.yml +++ b/.github/workflows/update-ops-docs.yml @@ -14,7 +14,7 @@ on: jobs: update-ops-docs: - runs-on: ubuntu-latest + runs-on: ubuntu-slim steps: - name: Checkout repository diff --git a/.github/workflows/winget.yml b/.github/workflows/winget.yml index 7506091647..2047c276f8 100644 --- a/.github/workflows/winget.yml +++ b/.github/workflows/winget.yml @@ -28,16 +28,17 @@ jobs: owner: context.repo.owner, repo: context.repo.repo, }); - console.log("Latest release:", releases[0].tag_name); - return releases[0].tag_name; + const { tag_name: version, assets: assets } = releases.find(({assets}) => assets.find(asset => asset.name.includes('win-vulkan'))); + const { browser_download_url: asset_url } = assets.find(asset => asset.name.includes('win-vulkan')); + console.log("Latest release:", version); + core.setOutput('VERSION', version); + core.setOutput('ASSETURL', asset_url); - name: Update manifest - env: - VERSION: ${{ steps.find_latest_release.outputs.result }} run: | echo "Updating manifest..." - komac update --version ${{ env.VERSION }} \ - --urls "https://github.com/ggml-org/llama.cpp/releases/download/${{ env.VERSION }}/llama-${{ env.VERSION }}-bin-win-vulkan-x64.zip" \ + komac update --version ${{ steps.find_latest_release.outputs.VERSION }} \ + --urls "${{ steps.find_latest_release.outputs.ASSETURL }}" \ --token ${{ secrets.WINGET_GITHUB_TOKEN }} \ --submit \ ggml.llamacpp diff --git a/CODEOWNERS b/CODEOWNERS index 55f5011dfa..e573a3d2e6 100644 --- a/CODEOWNERS +++ b/CODEOWNERS @@ -18,6 +18,7 @@ /common/jinja/ @ngxson @CISC @aldehir /common/llguidance.* @ggerganov /common/log.* @ggerganov +/common/ngram-map.* @srogmann /common/peg-parser.* @aldehir /common/sampling.* @ggerganov /common/speculative.* @ggerganov @@ -67,6 +68,7 @@ /ggml/src/ggml-rpc/ @rgerganov /ggml/src/ggml-threading.* @ggerganov /ggml/src/ggml-vulkan/ @0cc4m +/ggml/src/ggml-virtgpu/ @kpouget /ggml/src/ggml-webgpu/ @reeselevine /ggml/src/ggml-zdnn/ @taronaeo @Andreas-Krebbel @AlekseiNikiforovIBM /ggml/src/ggml.c @ggerganov diff --git a/README.md b/README.md index 91a8f25d1c..0783e43e5c 100644 --- a/README.md +++ b/README.md @@ -132,6 +132,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo - [x] [FalconMamba Models](https://huggingface.co/collections/tiiuae/falconmamba-7b-66b9a580324dd1598b0f6d4a) - [x] [Jais](https://huggingface.co/inceptionai/jais-13b-chat) - [x] [Bielik-11B-v2.3](https://huggingface.co/collections/speakleash/bielik-11b-v23-66ee813238d9b526a072408a) +- [x] [RWKV-7](https://huggingface.co/collections/shoumenchougou/rwkv7-gxx-gguf) - [x] [RWKV-6](https://github.com/BlinkDL/RWKV-LM) - [x] [QRWKV-6](https://huggingface.co/recursal/QRWKV6-32B-Instruct-Preview-v0.1) - [x] [GigaChat-20B-A3B](https://huggingface.co/ai-sage/GigaChat-20B-A3B-instruct) diff --git a/common/CMakeLists.txt b/common/CMakeLists.txt index ae02c0bd77..3bc7bc6210 100644 --- a/common/CMakeLists.txt +++ b/common/CMakeLists.txt @@ -73,6 +73,8 @@ add_library(${TARGET} STATIC log.h ngram-cache.cpp ngram-cache.h + ngram-map.cpp + ngram-map.h peg-parser.cpp peg-parser.h preset.cpp diff --git a/common/arg.cpp b/common/arg.cpp index 163c9b71b0..218418f070 100644 --- a/common/arg.cpp +++ b/common/arg.cpp @@ -6,6 +6,7 @@ #include "json-schema-to-grammar.h" #include "log.h" #include "sampling.h" +#include "speculative.h" #include "preset.h" // fix problem with std::min and std::max @@ -579,14 +580,14 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context params.mmproj = res.mmproj; } // only download mmproj if the current example is using it - for (auto & ex : mmproj_examples) { + for (const auto & ex : mmproj_examples) { if (ctx_arg.ex == ex) { common_params_handle_model(params.mmproj, params.hf_token, params.offline); break; } } - common_params_handle_model(params.speculative.model, params.hf_token, params.offline); - common_params_handle_model(params.vocoder.model, params.hf_token, params.offline); + common_params_handle_model(params.speculative.mparams_dft, params.hf_token, params.offline); + common_params_handle_model(params.vocoder.model, params.hf_token, params.offline); } // model is required (except for server) @@ -1216,21 +1217,25 @@ common_params_context common_params_parser_init(common_params & params, llama_ex {"-lcs", "--lookup-cache-static"}, "FNAME", "path to static lookup cache to use for lookup decoding (not updated by generation)", [](common_params & params, const std::string & value) { - params.lookup_cache_static = value; + params.speculative.lookup_cache_static = value; } - ).set_examples({LLAMA_EXAMPLE_LOOKUP})); + ).set_examples({LLAMA_EXAMPLE_LOOKUP, LLAMA_EXAMPLE_SERVER})); add_opt(common_arg( {"-lcd", "--lookup-cache-dynamic"}, "FNAME", "path to dynamic lookup cache to use for lookup decoding (updated by generation)", [](common_params & params, const std::string & value) { - params.lookup_cache_dynamic = value; + params.speculative.lookup_cache_dynamic = value; } - ).set_examples({LLAMA_EXAMPLE_LOOKUP})); + ).set_examples({LLAMA_EXAMPLE_LOOKUP, LLAMA_EXAMPLE_SERVER})); add_opt(common_arg( {"-c", "--ctx-size"}, "N", string_format("size of the prompt context (default: %d, 0 = loaded from model)", params.n_ctx), [](common_params & params, int value) { params.n_ctx = value; + if (value == 0) { + // disable context reduction in llama_params_fit if the user explicitly requests the full context size: + params.fit_params_min_ctx = UINT32_MAX; + } } ).set_env("LLAMA_ARG_CTX_SIZE")); add_opt(common_arg( @@ -1291,11 +1296,12 @@ common_params_context common_params_parser_init(common_params & params, llama_ex ).set_env("LLAMA_ARG_CACHE_RAM").set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI})); add_opt(common_arg( {"-kvu", "--kv-unified"}, + {"-no-kvu", "--no-kv-unified"}, "use single unified KV buffer shared across all sequences (default: enabled if number of slots is auto)", - [](common_params & params) { - params.kv_unified = true; + [](common_params & params, bool value) { + params.kv_unified = value; } - ).set_env("LLAMA_ARG_KV_UNIFIED").set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_PERPLEXITY, LLAMA_EXAMPLE_BATCHED})); + ).set_env("LLAMA_ARG_KV_UNIFIED").set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_PERPLEXITY, LLAMA_EXAMPLE_BATCHED, LLAMA_EXAMPLE_BENCH})); add_opt(common_arg( {"--context-shift"}, {"--no-context-shift"}, @@ -1573,7 +1579,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex ).set_sparam()); add_opt(common_arg( {"--temp"}, "N", - string_format("temperature (default: %.1f)", (double)params.sampling.temp), + string_format("temperature (default: %.2f)", (double)params.sampling.temp), [](common_params & params, const std::string & value) { params.sampling.temp = std::stof(value); params.sampling.temp = std::max(params.sampling.temp, 0.0f); @@ -1590,7 +1596,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex ).set_sparam().set_env("LLAMA_ARG_TOP_K")); add_opt(common_arg( {"--top-p"}, "N", - string_format("top-p sampling (default: %.1f, 1.0 = disabled)", (double)params.sampling.top_p), + string_format("top-p sampling (default: %.2f, 1.0 = disabled)", (double)params.sampling.top_p), [](common_params & params, const std::string & value) { params.sampling.top_p = std::stof(value); params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_TOP_P; @@ -1598,7 +1604,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex ).set_sparam()); add_opt(common_arg( {"--min-p"}, "N", - string_format("min-p sampling (default: %.1f, 0.0 = disabled)", (double)params.sampling.min_p), + string_format("min-p sampling (default: %.2f, 0.0 = disabled)", (double)params.sampling.min_p), [](common_params & params, const std::string & value) { params.sampling.min_p = std::stof(value); params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIN_P; @@ -1606,14 +1612,14 @@ common_params_context common_params_parser_init(common_params & params, llama_ex ).set_sparam()); add_opt(common_arg( {"--top-nsigma"}, "N", - string_format("top-n-sigma sampling (default: %.1f, -1.0 = disabled)", params.sampling.top_n_sigma), + string_format("top-n-sigma sampling (default: %.2f, -1.0 = disabled)", params.sampling.top_n_sigma), [](common_params & params, const std::string & value) { params.sampling.top_n_sigma = std::stof(value); } ).set_sparam()); add_opt(common_arg( {"--xtc-probability"}, "N", - string_format("xtc probability (default: %.1f, 0.0 = disabled)", (double)params.sampling.xtc_probability), + string_format("xtc probability (default: %.2f, 0.0 = disabled)", (double)params.sampling.xtc_probability), [](common_params & params, const std::string & value) { params.sampling.xtc_probability = std::stof(value); params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_XTC_PROBABILITY; @@ -1621,7 +1627,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex ).set_sparam()); add_opt(common_arg( {"--xtc-threshold"}, "N", - string_format("xtc threshold (default: %.1f, 1.0 = disabled)", (double)params.sampling.xtc_threshold), + string_format("xtc threshold (default: %.2f, 1.0 = disabled)", (double)params.sampling.xtc_threshold), [](common_params & params, const std::string & value) { params.sampling.xtc_threshold = std::stof(value); params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_XTC_THRESHOLD; @@ -1629,7 +1635,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex ).set_sparam()); add_opt(common_arg( {"--typical"}, "N", - string_format("locally typical sampling, parameter p (default: %.1f, 1.0 = disabled)", (double)params.sampling.typ_p), + string_format("locally typical sampling, parameter p (default: %.2f, 1.0 = disabled)", (double)params.sampling.typ_p), [](common_params & params, const std::string & value) { params.sampling.typ_p = std::stof(value); } @@ -1648,7 +1654,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex ).set_sparam()); add_opt(common_arg( {"--repeat-penalty"}, "N", - string_format("penalize repeat sequence of tokens (default: %.1f, 1.0 = disabled)", (double)params.sampling.penalty_repeat), + string_format("penalize repeat sequence of tokens (default: %.2f, 1.0 = disabled)", (double)params.sampling.penalty_repeat), [](common_params & params, const std::string & value) { params.sampling.penalty_repeat = std::stof(value); params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_PENALTY_REPEAT; @@ -1656,21 +1662,21 @@ common_params_context common_params_parser_init(common_params & params, llama_ex ).set_sparam()); add_opt(common_arg( {"--presence-penalty"}, "N", - string_format("repeat alpha presence penalty (default: %.1f, 0.0 = disabled)", (double)params.sampling.penalty_present), + string_format("repeat alpha presence penalty (default: %.2f, 0.0 = disabled)", (double)params.sampling.penalty_present), [](common_params & params, const std::string & value) { params.sampling.penalty_present = std::stof(value); } ).set_sparam()); add_opt(common_arg( {"--frequency-penalty"}, "N", - string_format("repeat alpha frequency penalty (default: %.1f, 0.0 = disabled)", (double)params.sampling.penalty_freq), + string_format("repeat alpha frequency penalty (default: %.2f, 0.0 = disabled)", (double)params.sampling.penalty_freq), [](common_params & params, const std::string & value) { params.sampling.penalty_freq = std::stof(value); } ).set_sparam()); add_opt(common_arg( {"--dry-multiplier"}, "N", - string_format("set DRY sampling multiplier (default: %.1f, 0.0 = disabled)", (double)params.sampling.dry_multiplier), + string_format("set DRY sampling multiplier (default: %.2f, 0.0 = disabled)", (double)params.sampling.dry_multiplier), [](common_params & params, const std::string & value) { params.sampling.dry_multiplier = std::stof(value); } @@ -1751,14 +1757,14 @@ common_params_context common_params_parser_init(common_params & params, llama_ex ).set_sparam()); add_opt(common_arg( {"--dynatemp-range"}, "N", - string_format("dynamic temperature range (default: %.1f, 0.0 = disabled)", (double)params.sampling.dynatemp_range), + string_format("dynamic temperature range (default: %.2f, 0.0 = disabled)", (double)params.sampling.dynatemp_range), [](common_params & params, const std::string & value) { params.sampling.dynatemp_range = std::stof(value); } ).set_sparam()); add_opt(common_arg( {"--dynatemp-exp"}, "N", - string_format("dynamic temperature exponent (default: %.1f)", (double)params.sampling.dynatemp_exponent), + string_format("dynamic temperature exponent (default: %.2f)", (double)params.sampling.dynatemp_exponent), [](common_params & params, const std::string & value) { params.sampling.dynatemp_exponent = std::stof(value); } @@ -1774,7 +1780,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex ).set_sparam()); add_opt(common_arg( {"--mirostat-lr"}, "N", - string_format("Mirostat learning rate, parameter eta (default: %.1f)", (double)params.sampling.mirostat_eta), + string_format("Mirostat learning rate, parameter eta (default: %.2f)", (double)params.sampling.mirostat_eta), [](common_params & params, const std::string & value) { params.sampling.mirostat_eta = std::stof(value); params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT_ETA; @@ -1782,7 +1788,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex ).set_sparam()); add_opt(common_arg( {"--mirostat-ent"}, "N", - string_format("Mirostat target entropy, parameter tau (default: %.1f)", (double)params.sampling.mirostat_tau), + string_format("Mirostat target entropy, parameter tau (default: %.2f)", (double)params.sampling.mirostat_tau), [](common_params & params, const std::string & value) { params.sampling.mirostat_tau = std::stof(value); params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT_TAU; @@ -1916,28 +1922,28 @@ common_params_context common_params_parser_init(common_params & params, llama_ex ).set_env("LLAMA_ARG_YARN_ORIG_CTX")); add_opt(common_arg( {"--yarn-ext-factor"}, "N", - string_format("YaRN: extrapolation mix factor (default: %.1f, 0.0 = full interpolation)", (double)params.yarn_ext_factor), + string_format("YaRN: extrapolation mix factor (default: %.2f, 0.0 = full interpolation)", (double)params.yarn_ext_factor), [](common_params & params, const std::string & value) { params.yarn_ext_factor = std::stof(value); } ).set_env("LLAMA_ARG_YARN_EXT_FACTOR")); add_opt(common_arg( {"--yarn-attn-factor"}, "N", - string_format("YaRN: scale sqrt(t) or attention magnitude (default: %.1f)", (double)params.yarn_attn_factor), + string_format("YaRN: scale sqrt(t) or attention magnitude (default: %.2f)", (double)params.yarn_attn_factor), [](common_params & params, const std::string & value) { params.yarn_attn_factor = std::stof(value); } ).set_env("LLAMA_ARG_YARN_ATTN_FACTOR")); add_opt(common_arg( {"--yarn-beta-slow"}, "N", - string_format("YaRN: high correction dim or alpha (default: %.1f)", (double)params.yarn_beta_slow), + string_format("YaRN: high correction dim or alpha (default: %.2f)", (double)params.yarn_beta_slow), [](common_params & params, const std::string & value) { params.yarn_beta_slow = std::stof(value); } ).set_env("LLAMA_ARG_YARN_BETA_SLOW")); add_opt(common_arg( {"--yarn-beta-fast"}, "N", - string_format("YaRN: low correction dim or beta (default: %.1f)", (double)params.yarn_beta_fast), + string_format("YaRN: low correction dim or beta (default: %.2f)", (double)params.yarn_beta_fast), [](common_params & params, const std::string & value) { params.yarn_beta_fast = std::stof(value); } @@ -2194,18 +2200,15 @@ common_params_context common_params_parser_init(common_params & params, llama_ex add_opt(common_arg( {"--mmap"}, {"--no-mmap"}, - string_format("whether to memory-map model. Explicitly enabling mmap disables direct-io. (if mmap disabled, slower load but may reduce pageouts if not using mlock) (default: %s)", params.use_mmap ? "enabled" : "disabled"), + string_format("whether to memory-map model. (if mmap disabled, slower load but may reduce pageouts if not using mlock) (default: %s)", params.use_mmap ? "enabled" : "disabled"), [](common_params & params, bool value) { params.use_mmap = value; - if (value) { - params.use_direct_io = false; // disable direct io when mmap is explicitly enabled - } } ).set_env("LLAMA_ARG_MMAP")); add_opt(common_arg( {"-dio", "--direct-io"}, {"-ndio", "--no-direct-io"}, - string_format("use DirectIO if available. Takes precedence over --mmap (default: %s)", params.use_direct_io ? "enabled" : "disabled"), + string_format("use DirectIO if available. (default: %s)", params.use_direct_io ? "enabled" : "disabled"), [](common_params & params, bool value) { params.use_direct_io = value; } @@ -2561,7 +2564,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex {"-hfd", "-hfrd", "--hf-repo-draft"}, "/[:quant]", "Same as --hf-repo, but for the draft model (default: unused)", [](common_params & params, const std::string & value) { - params.speculative.model.hf_repo = value; + params.speculative.mparams_dft.hf_repo = value; } ).set_env("LLAMA_ARG_HFD_REPO")); add_opt(common_arg( @@ -3331,14 +3334,14 @@ common_params_context common_params_parser_init(common_params & params, llama_ex ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_LOOKUP, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}).set_env("LLAMA_ARG_DRAFT_MIN")); add_opt(common_arg( {"--draft-p-split"}, "P", - string_format("speculative decoding split probability (default: %.1f)", (double)params.speculative.p_split), + string_format("speculative decoding split probability (default: %.2f)", (double)params.speculative.p_split), [](common_params & params, const std::string & value) { params.speculative.p_split = std::stof(value); } ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}).set_env("LLAMA_ARG_DRAFT_P_SPLIT")); add_opt(common_arg( {"--draft-p-min"}, "P", - string_format("minimum speculative decoding probability (greedy) (default: %.1f)", (double)params.speculative.p_min), + string_format("minimum speculative decoding probability (greedy) (default: %.2f)", (double)params.speculative.p_min), [](common_params & params, const std::string & value) { params.speculative.p_min = std::stof(value); } @@ -3382,7 +3385,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex {"-md", "--model-draft"}, "FNAME", "draft model for speculative decoding (default: unused)", [](common_params & params, const std::string & value) { - params.speculative.model.path = value; + params.speculative.mparams_dft.path = value; } ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}).set_env("LLAMA_ARG_MODEL_DRAFT")); add_opt(common_arg( @@ -3392,6 +3395,66 @@ common_params_context common_params_parser_init(common_params & params, llama_ex params.speculative.replacements.push_back({ tgt, dft }); } ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI})); + add_opt(common_arg( + {"--spec-type"}, "[none|ngram-cache|ngram-simple|ngram-map-k|ngram-map-k4v]", + string_format("type of speculative decoding to use when no draft model is provided (default: %s)\n", + common_speculative_type_to_str(params.speculative.type).c_str()), + [](common_params & params, const std::string & value) { + if (value == "none") { + params.speculative.type = COMMON_SPECULATIVE_TYPE_NONE; + } else if (value == "ngram-cache") { + params.speculative.type = COMMON_SPECULATIVE_TYPE_NGRAM_CACHE; + } else if (value == "ngram-simple") { + params.speculative.type = COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE; + } else if (value == "ngram-map-k") { + params.speculative.type = COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K; + } else if (value == "ngram-map-k4v") { + params.speculative.type = COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V; + } else { + throw std::invalid_argument("unknown speculative decoding type without draft model"); + } + } + ).set_examples({LLAMA_EXAMPLE_SERVER})); + add_opt(common_arg( + {"--spec-ngram-size-n"}, "N", + string_format("ngram size N for ngram-simple/ngram-map speculative decoding, length of lookup n-gram (default: %d)", params.speculative.ngram_size_n), + [](common_params & params, int value) { + if (value < 1 || value > 1024) { + throw std::invalid_argument("ngram size N must be between 1 and 1024 inclusive"); + } + params.speculative.ngram_size_n = value; + } + ).set_examples({LLAMA_EXAMPLE_SERVER})); + add_opt(common_arg( + {"--spec-ngram-size-m"}, "N", + string_format("ngram size M for ngram-simple/ngram-map speculative decoding, length of draft m-gram (default: %d)", params.speculative.ngram_size_m), + [](common_params & params, int value) { + if (value < 1 || value > 1024) { + throw std::invalid_argument("ngram size M must be between 1 and 1024 inclusive"); + } + params.speculative.ngram_size_m = value; + } + ).set_examples({LLAMA_EXAMPLE_SERVER})); + add_opt(common_arg( + {"--spec-ngram-check-rate"}, "N", + string_format("ngram check rate for ngram-simple/ngram-map speculative decoding (default: %d)", params.speculative.ngram_check_rate), + [](common_params & params, int value) { + if (value < 1) { + throw std::invalid_argument("ngram check rate must be at least 1"); + } + params.speculative.ngram_check_rate = value; + } + ).set_examples({LLAMA_EXAMPLE_SERVER})); + add_opt(common_arg( + {"--spec-ngram-min-hits"}, "N", + string_format("minimum hits for ngram-map speculative decoding (default: %d)", params.speculative.ngram_min_hits), + [](common_params & params, int value) { + if (value < 1) { + throw std::invalid_argument("ngram min hits must be at least 1"); + } + params.speculative.ngram_min_hits = value; + } + ).set_examples({LLAMA_EXAMPLE_SERVER})); add_opt(common_arg( {"-ctkd", "--cache-type-k-draft"}, "TYPE", string_format( @@ -3618,8 +3681,8 @@ common_params_context common_params_parser_init(common_params & params, llama_ex [](common_params & params) { params.model.hf_repo = "ggml-org/Qwen2.5-Coder-7B-Q8_0-GGUF"; params.model.hf_file = "qwen2.5-coder-7b-q8_0.gguf"; - params.speculative.model.hf_repo = "ggml-org/Qwen2.5-Coder-0.5B-Q8_0-GGUF"; - params.speculative.model.hf_file = "qwen2.5-coder-0.5b-q8_0.gguf"; + params.speculative.mparams_dft.hf_repo = "ggml-org/Qwen2.5-Coder-0.5B-Q8_0-GGUF"; + params.speculative.mparams_dft.hf_file = "qwen2.5-coder-0.5b-q8_0.gguf"; params.port = 8012; params.n_ubatch = 1024; params.n_batch = 1024; @@ -3634,8 +3697,8 @@ common_params_context common_params_parser_init(common_params & params, llama_ex [](common_params & params) { params.model.hf_repo = "ggml-org/Qwen2.5-Coder-14B-Q8_0-GGUF"; params.model.hf_file = "qwen2.5-coder-14b-q8_0.gguf"; - params.speculative.model.hf_repo = "ggml-org/Qwen2.5-Coder-0.5B-Q8_0-GGUF"; - params.speculative.model.hf_file = "qwen2.5-coder-0.5b-q8_0.gguf"; + params.speculative.mparams_dft.hf_repo = "ggml-org/Qwen2.5-Coder-0.5B-Q8_0-GGUF"; + params.speculative.mparams_dft.hf_file = "qwen2.5-coder-0.5b-q8_0.gguf"; params.port = 8012; params.n_ubatch = 1024; params.n_batch = 1024; diff --git a/common/chat.cpp b/common/chat.cpp index aba26e97a1..eeb38ad06a 100644 --- a/common/chat.cpp +++ b/common/chat.cpp @@ -2659,6 +2659,10 @@ static common_chat_params common_chat_params_init_translate_gemma(const common_c templates_params inputs_new = inputs; json & messages = inputs_new.messages; + // default to chat_template_kwargs, or en-GB if not specified + std::string default_src_lang = inputs.extra_context.value("source_lang_code", "en-GB"); + std::string default_tgt_lang = inputs.extra_context.value("target_lang_code", "en-GB"); + GGML_ASSERT(messages.is_array()); for (auto & message : messages) { if (message.contains("role") && message["role"].get() != "user") { @@ -2670,8 +2674,10 @@ static common_chat_params common_chat_params_init_translate_gemma(const common_c if (message.contains("content") && !message["content"].is_array()) { auto content_str = message["content"].get(); // default to en-GB if not specified (to make common_chat_format_example works) - auto src_lang = message.contains("source_lang_code") ? message["source_lang_code"].get() : "en-GB"; - auto tgt_lang = message.contains("target_lang_code") ? message["target_lang_code"].get() : "en-GB"; + auto src_lang = message.contains("source_lang_code") + ? message["source_lang_code"].get() : default_src_lang; + auto tgt_lang = message.contains("target_lang_code") + ? message["target_lang_code"].get() : default_tgt_lang; message["content"] = json::array({ json{ {"type", "text"}, diff --git a/common/common.cpp b/common/common.cpp index 26250abb6c..3aa396127c 100644 --- a/common/common.cpp +++ b/common/common.cpp @@ -1097,7 +1097,10 @@ common_init_result::common_init_result(common_params & params) : if (params.fit_params) { LOG_INF("%s: fitting params to device memory, for bugs during this step try to reproduce them with -fit off, or provide --verbose logs if the bug only occurs with -fit on\n", __func__); llama_params_fit(params.model.path.c_str(), &mparams, &cparams, - params.tensor_split, params.tensor_buft_overrides.data(), params.fit_params_target.data(), params.fit_params_min_ctx, + params.tensor_split, + params.tensor_buft_overrides.data(), + params.fit_params_target.data(), + params.fit_params_min_ctx, params.verbosity >= 4 ? GGML_LOG_LEVEL_DEBUG : GGML_LOG_LEVEL_ERROR); } @@ -1208,10 +1211,6 @@ std::vector & common_init_result::lora() { return pimpl->lora; } -void common_init_result::free_context() { - pimpl->context.reset(); -} - common_init_result_ptr common_init_from_params(common_params & params) { common_init_result_ptr res(new common_init_result(params)); diff --git a/common/common.h b/common/common.h index 96c990c05d..fd3ab8cd18 100644 --- a/common/common.h +++ b/common/common.h @@ -164,6 +164,16 @@ enum common_params_sampling_config : uint64_t { COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT_ETA = 1 << 11, }; +enum common_speculative_type { + COMMON_SPECULATIVE_TYPE_NONE, // no speculative decoding + COMMON_SPECULATIVE_TYPE_DRAFT, // draft model + COMMON_SPECULATIVE_TYPE_EAGLE3, // eagle draft model + COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE, // simple self-speculative decoding + COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K, // self-speculative decoding with n-gram keys only + COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V, // self-speculative decoding with n-gram keys and 4 m-gram values + COMMON_SPECULATIVE_TYPE_NGRAM_CACHE, // self-speculative decoding with 3-level n-gram cache + COMMON_SPECULATIVE_TYPE_COUNT // number of types, unknown type +}; // sampling parameters struct common_params_sampling { @@ -243,16 +253,35 @@ struct common_params_model { }; struct common_params_speculative { - std::vector devices; // devices to use for offloading + common_speculative_type type = COMMON_SPECULATIVE_TYPE_NONE; // type of speculative decoding - int32_t n_ctx = 0; // draft context size - int32_t n_max = 16; // maximum number of tokens to draft during speculative decoding - int32_t n_min = 0; // minimum number of draft tokens to use for speculative decoding - int32_t n_gpu_layers = -1; // number of layers to store in VRAM for the draft model (-1 - use default) - float p_split = 0.1f; // speculative decoding split probability - float p_min = 0.75f; // minimum speculative decoding probability (greedy) - std::vector> replacements; // main to speculative model replacements - std::vector tensor_buft_overrides; + // general-purpose speculative decoding parameters + + int32_t n_max = 16; // maximum number of tokens to draft during speculative decoding + int32_t n_min = 0; // minimum number of draft tokens to use for speculative decoding + float p_split = 0.1f; // speculative decoding split probability + float p_min = 0.75f; // minimum speculative decoding probability (greedy) + + // ngram-based speculative decoding + + uint16_t ngram_size_n = 12; // ngram size for lookup + uint16_t ngram_size_m = 48; // mgram size for speculative tokens + uint16_t ngram_check_rate = 1; // check rate for ngram lookup + uint16_t ngram_min_hits = 1; // minimum hits at ngram/mgram lookup for mgram to be proposed + + std::string lookup_cache_static; // path of static ngram cache file for lookup decoding // NOLINT + std::string lookup_cache_dynamic; // path of dynamic ngram cache file for lookup decoding // NOLINT + + // draft-model speculative decoding + + struct common_params_model mparams_dft; + + llama_model * model_dft = nullptr; // a llama_model that can be shared by multiple speculative contexts + + llama_context_params cparams_dft; // these are the parameters for the draft llama_context + + int32_t n_ctx = 0; // draft context size + int32_t n_gpu_layers = -1; // number of layers to store in VRAM for the draft model (-1 - use default) ggml_type cache_type_k = GGML_TYPE_F16; // KV cache data type for the K ggml_type cache_type_v = GGML_TYPE_F16; // KV cache data type for the V @@ -260,7 +289,14 @@ struct common_params_speculative { struct cpu_params cpuparams; struct cpu_params cpuparams_batch; - struct common_params_model model; + std::vector devices; // devices to use for offloading + + std::vector> replacements; // main to speculative model replacements + std::vector tensor_buft_overrides; + + bool has_dft() const { + return !mparams_dft.path.empty() || !mparams_dft.hf_repo.empty(); + } }; struct common_params_vocoder { @@ -378,8 +414,6 @@ struct common_params { std::string path_prompt_cache = ""; // path to file for saving/loading prompt eval state // NOLINT std::string input_prefix = ""; // string to prefix user inputs with // NOLINT std::string input_suffix = ""; // string to suffix user inputs with // NOLINT - std::string lookup_cache_static = ""; // path of static ngram cache file for lookup decoding // NOLINT - std::string lookup_cache_dynamic = ""; // path of dynamic ngram cache file for lookup decoding // NOLINT std::string logits_file = ""; // file for saving *all* logits // NOLINT // llama-debug specific options @@ -438,7 +472,7 @@ struct common_params { bool input_prefix_bos = false; // prefix BOS to user inputs, preceding input_prefix bool use_mmap = true; // enable mmap to use filesystem cache - bool use_direct_io = true; // read from disk without buffering for faster model loading + bool use_direct_io = false; // read from disk without buffering bool use_mlock = false; // use mlock to keep model in memory bool verbose_prompt = false; // print prompt tokens before generation bool display_prompt = true; // print prompt before generation @@ -575,10 +609,6 @@ struct common_params { // return false from callback to abort model loading or true to continue llama_progress_callback load_progress_callback = NULL; void * load_progress_callback_user_data = NULL; - - bool has_speculative() const { - return !speculative.model.path.empty() || !speculative.model.hf_repo.empty(); - } }; // call once at the start of a program if it uses libcommon @@ -714,8 +744,6 @@ struct common_init_result { std::vector & lora(); - void free_context(); - private: struct impl; std::unique_ptr pimpl; diff --git a/common/http.h b/common/http.h index 7c683aafcf..e8ed56f952 100644 --- a/common/http.h +++ b/common/http.h @@ -60,10 +60,10 @@ static std::pair common_http_client(const std: #ifndef CPPHTTPLIB_OPENSSL_SUPPORT if (parts.scheme == "https") { throw std::runtime_error( - "HTTPS is not supported. Please rebuild with:\n" + "HTTPS is not supported. Please rebuild with one of:\n" " -DLLAMA_BUILD_BORINGSSL=ON\n" " -DLLAMA_BUILD_LIBRESSL=ON\n" - "or ensure dev files of an OpenSSL-compatible library are available when building." + " -DLLAMA_OPENSSL=ON (default, requires OpenSSL dev files installed)" ); } #endif diff --git a/common/jinja/runtime.cpp b/common/jinja/runtime.cpp index e3e4ebf1ec..f234d9284f 100644 --- a/common/jinja/runtime.cpp +++ b/common/jinja/runtime.cpp @@ -44,6 +44,12 @@ static std::string get_line_col(const std::string & source, size_t pos) { return "line " + std::to_string(line) + ", column " + std::to_string(col); } +static void ensure_key_type_allowed(const value & val) { + if (!val->is_hashable()) { + throw std::runtime_error("Type: " + val->type() + " is not allowed as object key"); + } +} + // execute with error handling value statement::execute(context & ctx) { try { @@ -95,20 +101,10 @@ value identifier::execute_impl(context & ctx) { value object_literal::execute_impl(context & ctx) { auto obj = mk_val(); for (const auto & pair : val) { - value key_val = pair.first->execute(ctx); - if (!is_val(key_val) && !is_val(key_val)) { - throw std::runtime_error("Object literal: keys must be string or int values, got " + key_val->type()); - } - std::string key = key_val->as_string().str(); + value key = pair.first->execute(ctx); value val = pair.second->execute(ctx); - JJ_DEBUG("Object literal: setting key '%s' with value type %s", key.c_str(), val->type().c_str()); + JJ_DEBUG("Object literal: setting key '%s' with value type %s", key->as_string().str().c_str(), val->type().c_str()); obj->insert(key, val); - - if (is_val(key_val)) { - obj->val_obj.is_key_numeric = true; - } else if (obj->val_obj.is_key_numeric) { - throw std::runtime_error("Object literal: cannot mix numeric and non-numeric keys"); - } } return obj; } @@ -127,9 +123,9 @@ value binary_expression::execute_impl(context & ctx) { value right_val = right->execute(ctx); JJ_DEBUG("Executing binary expression %s '%s' %s", left_val->type().c_str(), op.value.c_str(), right_val->type().c_str()); if (op.value == "==") { - return mk_val(value_compare(left_val, right_val, value_compare_op::eq)); + return mk_val(*left_val == *right_val); } else if (op.value == "!=") { - return mk_val(!value_compare(left_val, right_val, value_compare_op::eq)); + return mk_val(!(*left_val == *right_val)); } auto workaround_concat_null_with_str = [&](value & res) -> bool { @@ -230,7 +226,7 @@ value binary_expression::execute_impl(context & ctx) { auto & arr = right_val->as_array(); bool member = false; for (const auto & item : arr) { - if (value_compare(left_val, item, value_compare_op::eq)) { + if (*left_val == *item) { member = true; break; } @@ -265,10 +261,9 @@ value binary_expression::execute_impl(context & ctx) { } } - // String in object - if (is_val(left_val) && is_val(right_val)) { - auto key = left_val->as_string().str(); - bool has_key = right_val->has_key(key); + // Value key in object + if (is_val(right_val)) { + bool has_key = right_val->has_key(left_val); if (op.value == "in") { return mk_val(has_key); } else if (op.value == "not in") { @@ -465,14 +460,8 @@ value for_statement::execute_impl(context & ctx) { JJ_DEBUG("%s", "For loop over object keys"); auto & obj = iterable_val->as_ordered_object(); for (auto & p : obj) { - auto tuple = mk_val(); - if (iterable_val->val_obj.is_key_numeric) { - tuple->push_back(mk_val(std::stoll(p.first))); - } else { - tuple->push_back(mk_val(p.first)); - } - tuple->push_back(p.second); - items.push_back(tuple); + auto tuple = mk_val(p); + items.push_back(std::move(tuple)); } if (ctx.is_get_stats) { iterable_val->stats.used = true; @@ -602,11 +591,13 @@ value set_statement::execute_impl(context & ctx) { auto rhs = val ? val->execute(ctx) : exec_statements(body, ctx); if (is_stmt(assignee)) { + // case: {% set my_var = value %} auto var_name = cast_stmt(assignee)->val; JJ_DEBUG("Setting global variable '%s' with value type %s", var_name.c_str(), rhs->type().c_str()); ctx.set_val(var_name, rhs); } else if (is_stmt(assignee)) { + // case: {% set a, b = value %} auto tuple = cast_stmt(assignee); if (!is_val(rhs)) { throw std::runtime_error("Cannot unpack non-iterable type in set: " + rhs->type()); @@ -625,6 +616,7 @@ value set_statement::execute_impl(context & ctx) { } } else if (is_stmt(assignee)) { + // case: {% set ns.my_var = value %} auto member = cast_stmt(assignee); if (member->computed) { throw std::runtime_error("Cannot assign to computed member"); @@ -767,22 +759,22 @@ value member_expression::execute_impl(context & ctx) { } JJ_DEBUG("Member expression on object type %s, property type %s", object->type().c_str(), property->type().c_str()); + ensure_key_type_allowed(property); value val = mk_val("object_property"); if (is_val(object)) { JJ_DEBUG("%s", "Accessing property on undefined object, returning undefined"); return val; + } else if (is_val(object)) { - if (!is_val(property)) { - throw std::runtime_error("Cannot access object with non-string: got " + property->type()); - } auto key = property->as_string().str(); - val = object->at(key, val); + val = object->at(property, val); if (is_val(val)) { val = try_builtin_func(ctx, key, object, true); } JJ_DEBUG("Accessed property '%s' value, got type: %s", key.c_str(), val->type().c_str()); + } else if (is_val(object) || is_val(object)) { if (is_val(property)) { int64_t index = property->as_int(); @@ -806,6 +798,7 @@ value member_expression::execute_impl(context & ctx) { auto key = property->as_string().str(); JJ_DEBUG("Accessing %s built-in '%s'", is_val(object) ? "array" : "string", key.c_str()); val = try_builtin_func(ctx, key, object, true); + } else { throw std::runtime_error("Cannot access property with non-string/non-number: got " + property->type()); } diff --git a/common/jinja/runtime.h b/common/jinja/runtime.h index dc7f4e471c..17a6dff5aa 100644 --- a/common/jinja/runtime.h +++ b/common/jinja/runtime.h @@ -79,18 +79,18 @@ struct context { } value get_val(const std::string & name) { - auto it = env->val_obj.unordered.find(name); - if (it != env->val_obj.unordered.end()) { - return it->second; - } else { - return mk_val(name); - } + value default_val = mk_val(name); + return env->at(name, default_val); } void set_val(const std::string & name, const value & val) { env->insert(name, val); } + void set_val(const value & name, const value & val) { + env->insert(name, val); + } + void print_vars() const { printf("Context Variables:\n%s\n", value_to_json(env, 2).c_str()); } @@ -344,9 +344,19 @@ struct array_literal : public expression { } }; -struct tuple_literal : public array_literal { - explicit tuple_literal(statements && val) : array_literal(std::move(val)) {} +struct tuple_literal : public expression { + statements val; + explicit tuple_literal(statements && val) : val(std::move(val)) { + for (const auto& item : this->val) chk_type(item); + } std::string type() const override { return "TupleLiteral"; } + value execute_impl(context & ctx) override { + auto arr = mk_val(); + for (const auto & item_stmt : val) { + arr->push_back(item_stmt->execute(ctx)); + } + return mk_val(std::move(arr->as_array())); + } }; struct object_literal : public expression { diff --git a/common/jinja/string.cpp b/common/jinja/string.cpp index 21ebde39e3..8087e15b35 100644 --- a/common/jinja/string.cpp +++ b/common/jinja/string.cpp @@ -61,6 +61,12 @@ size_t string::length() const { return len; } +void string::hash_update(hasher & hash) const noexcept { + for (const auto & part : parts) { + hash.update(part.val.data(), part.val.length()); + } +} + bool string::all_parts_are_input() const { for (const auto & part : parts) { if (!part.is_input) { diff --git a/common/jinja/string.h b/common/jinja/string.h index 78457f9e41..c4963000ad 100644 --- a/common/jinja/string.h +++ b/common/jinja/string.h @@ -4,6 +4,8 @@ #include #include +#include "utils.h" + namespace jinja { // allow differentiate between user input strings and template strings @@ -37,6 +39,7 @@ struct string { std::string str() const; size_t length() const; + void hash_update(hasher & hash) const noexcept; bool all_parts_are_input() const; bool is_uppercase() const; bool is_lowercase() const; diff --git a/common/jinja/utils.h b/common/jinja/utils.h index 1e9f2a12a1..de6947fc28 100644 --- a/common/jinja/utils.h +++ b/common/jinja/utils.h @@ -3,6 +3,8 @@ #include #include #include +#include +#include namespace jinja { @@ -46,4 +48,102 @@ static std::string fmt_error_with_source(const std::string & tag, const std::str return oss.str(); } +// Note: this is a simple hasher, not cryptographically secure, just for hash table usage +struct hasher { + static constexpr auto size_t_digits = sizeof(size_t) * 8; + static constexpr size_t prime = size_t_digits == 64 ? 0x100000001b3 : 0x01000193; + static constexpr size_t seed = size_t_digits == 64 ? 0xcbf29ce484222325 : 0x811c9dc5; + static constexpr auto block_size = sizeof(size_t); // in bytes; allowing the compiler to vectorize the computation + + static_assert(size_t_digits == 64 || size_t_digits == 32); + static_assert(block_size == 8 || block_size == 4); + + uint8_t buffer[block_size]; + size_t idx = 0; // current index in buffer + size_t state = seed; + + hasher() = default; + hasher(const std::type_info & type_inf) noexcept { + const auto type_hash = type_inf.hash_code(); + update(&type_hash, sizeof(type_hash)); + } + + // Properties: + // - update is not associative: update(a).update(b) != update(b).update(a) + // - update(a ~ b) == update(a).update(b) with ~ as concatenation operator --> useful for streaming + // - update("", 0) --> state unchanged with empty input + hasher& update(void const * bytes, size_t len) noexcept { + const uint8_t * c = static_cast(bytes); + if (len == 0) { + return *this; + } + size_t processed = 0; + + // first, fill the existing buffer if it's partial + if (idx > 0) { + size_t to_fill = block_size - idx; + if (to_fill > len) { + to_fill = len; + } + std::memcpy(buffer + idx, c, to_fill); + idx += to_fill; + processed += to_fill; + if (idx == block_size) { + update_block(buffer); + idx = 0; + } + } + + // process full blocks from the remaining input + for (; processed + block_size <= len; processed += block_size) { + update_block(c + processed); + } + + // buffer any remaining bytes + size_t remaining = len - processed; + if (remaining > 0) { + std::memcpy(buffer, c + processed, remaining); + idx = remaining; + } + return *this; + } + + // convenience function for testing only + hasher& update(const std::string & s) noexcept { + return update(s.data(), s.size()); + } + + // finalize and get the hash value + // note: after calling digest, the hasher state is modified, do not call update() again + size_t digest() noexcept { + // if there are remaining bytes in buffer, fill the rest with zeros and process + if (idx > 0) { + for (size_t i = idx; i < block_size; ++i) { + buffer[i] = 0; + } + update_block(buffer); + idx = 0; + } + + return state; + } + +private: + // IMPORTANT: block must have at least block_size bytes + void update_block(const uint8_t * block) noexcept { + size_t blk = static_cast(block[0]) + | (static_cast(block[1]) << 8) + | (static_cast(block[2]) << 16) + | (static_cast(block[3]) << 24); + if constexpr (block_size == 8) { + blk = blk | (static_cast(block[4]) << 32) + | (static_cast(block[5]) << 40) + | (static_cast(block[6]) << 48) + | (static_cast(block[7]) << 56); + } + state ^= blk; + state *= prime; + } +}; + } // namespace jinja diff --git a/common/jinja/value.cpp b/common/jinja/value.cpp index d2ed824269..17d43826c0 100644 --- a/common/jinja/value.cpp +++ b/common/jinja/value.cpp @@ -114,6 +114,18 @@ static T slice(const T & array, int64_t start, int64_t stop, int64_t step = 1) { return result; } +template +static value empty_value_fn(const func_args &) { + if constexpr (std::is_same_v) { + return mk_val(0); + } else if constexpr (std::is_same_v) { + return mk_val(0.0); + } else if constexpr (std::is_same_v) { + return mk_val(false); + } else { + return mk_val(); + } +} template static value test_type_fn(const func_args & args) { args.ensure_count(1); @@ -128,6 +140,13 @@ static value test_type_fn(const func_args & args) { JJ_DEBUG("test_type_fn: type=%s or %s result=%d", typeid(T).name(), typeid(U).name(), is_type ? 1 : 0); return mk_val(is_type); } +template +static value test_type_fn(const func_args & args) { + args.ensure_count(1); + bool is_type = is_val(args.get_pos(0)) || is_val(args.get_pos(0)) || is_val(args.get_pos(0)); + JJ_DEBUG("test_type_fn: type=%s, %s or %s result=%d", typeid(T).name(), typeid(U).name(), typeid(V).name(), is_type ? 1 : 0); + return mk_val(is_type); +} template static value test_compare_fn(const func_args & args) { args.ensure_count(2, 2); @@ -163,7 +182,7 @@ static value selectattr(const func_args & args) { args.ensure_vals(true, true, false, false); auto arr = args.get_pos(0)->as_array(); - auto attr_name = args.get_pos(1)->as_string().str(); + auto attribute = args.get_pos(1); auto out = mk_val(); value val_default = mk_val(); @@ -173,7 +192,7 @@ static value selectattr(const func_args & args) { if (!is_val(item)) { throw raised_exception("selectattr: item is not an object"); } - value attr_val = item->at(attr_name, val_default); + value attr_val = item->at(attribute, val_default); bool is_selected = attr_val->as_bool(); if constexpr (is_reject) is_selected = !is_selected; if (is_selected) out->push_back(item); @@ -217,7 +236,7 @@ static value selectattr(const func_args & args) { if (!is_val(item)) { throw raised_exception("selectattr: item is not an object"); } - value attr_val = item->at(attr_name, val_default); + value attr_val = item->at(attribute, val_default); func_args test_args(args.ctx); test_args.push_back(attr_val); // attribute value test_args.push_back(extra_arg); // extra argument @@ -347,8 +366,8 @@ const func_builtins & global_builtins() { {"test_is_integer", test_type_fn}, {"test_is_float", test_type_fn}, {"test_is_number", test_type_fn}, - {"test_is_iterable", test_type_fn}, - {"test_is_sequence", test_type_fn}, + {"test_is_iterable", test_type_fn}, + {"test_is_sequence", test_type_fn}, {"test_is_mapping", test_type_fn}, {"test_is_lower", [](const func_args & args) -> value { args.ensure_vals(); @@ -741,6 +760,7 @@ const func_builtins & value_array_t::get_builtins() const { args.ensure_count(1, 4); args.ensure_vals(true, true, false, false); + auto val = args.get_pos(0); auto arg0 = args.get_pos(1); auto arg1 = args.get_pos(2, mk_val()); auto arg2 = args.get_pos(3, mk_val()); @@ -762,10 +782,8 @@ const func_builtins & value_array_t::get_builtins() const { if (step == 0) { throw raised_exception("slice step cannot be zero"); } - auto arr = slice(args.get_pos(0)->as_array(), start, stop, step); - auto res = mk_val(); - res->val_arr = std::move(arr); - return res; + auto arr = slice(val->as_array(), start, stop, step); + return is_val(val) ? mk_val(std::move(arr)) : mk_val(std::move(arr)); }}, {"selectattr", selectattr}, {"select", selectattr}, @@ -785,15 +803,14 @@ const func_builtins & value_array_t::get_builtins() const { } const int64_t attr_int = attr_is_int ? attribute->as_int() : 0; const std::string delim = val_delim->is_undefined() ? "" : val_delim->as_string().str(); - const std::string attr_name = attribute->is_undefined() ? "" : attribute->as_string().str(); std::string result; for (size_t i = 0; i < arr.size(); ++i) { value val_arr = arr[i]; if (!attribute->is_undefined()) { if (attr_is_int && is_val(val_arr)) { val_arr = val_arr->at(attr_int); - } else if (!attr_is_int && !attr_name.empty() && is_val(val_arr)) { - val_arr = val_arr->at(attr_name); + } else if (!attr_is_int && is_val(val_arr)) { + val_arr = val_arr->at(attribute); } } if (!is_val(val_arr) && !is_val(val_arr) && !is_val(val_arr)) { @@ -808,9 +825,7 @@ const func_builtins & value_array_t::get_builtins() const { }}, {"string", [](const func_args & args) -> value { args.ensure_vals(); - auto str = mk_val(); - gather_string_parts_recursive(args.get_pos(0), str); - return str; + return mk_val(args.get_pos(0)->as_string()); }}, {"tojson", tojson}, {"map", [](const func_args & args) -> value { @@ -821,26 +836,26 @@ const func_builtins & value_array_t::get_builtins() const { if (!is_val(args.get_args().at(1))) { throw not_implemented_exception("map: filter-mapping not implemented"); } + value val = args.get_pos(0); value attribute = args.get_kwarg_or_pos("attribute", 1); const bool attr_is_int = is_val(attribute); if (!is_val(attribute) && !attr_is_int) { throw raised_exception("map: attribute must be string or integer"); } const int64_t attr_int = attr_is_int ? attribute->as_int() : 0; - const std::string attr_name = attribute->as_string().str(); value default_val = args.get_kwarg("default", mk_val()); auto out = mk_val(); - auto arr = args.get_pos(0)->as_array(); + auto arr = val->as_array(); for (const auto & item : arr) { value attr_val; if (attr_is_int) { attr_val = is_val(item) ? item->at(attr_int, default_val) : default_val; } else { - attr_val = is_val(item) ? item->at(attr_name, default_val) : default_val; + attr_val = is_val(item) ? item->at(attribute, default_val) : default_val; } out->push_back(attr_val); } - return out; + return is_val(val) ? mk_val(std::move(out->as_array())) : out; }}, {"append", [](const func_args & args) -> value { args.ensure_count(2); @@ -867,6 +882,7 @@ const func_builtins & value_array_t::get_builtins() const { if (!is_val(args.get_pos(0))) { throw raised_exception("sort: first argument must be an array"); } + value val = args.get_pos(0); value val_reverse = args.get_kwarg_or_pos("reverse", 1); value val_case = args.get_kwarg_or_pos("case_sensitive", 2); value attribute = args.get_kwarg_or_pos("attribute", 3); @@ -875,8 +891,7 @@ const func_builtins & value_array_t::get_builtins() const { const bool reverse = val_reverse->as_bool(); // undefined == false const bool attr_is_int = is_val(attribute); const int64_t attr_int = attr_is_int ? attribute->as_int() : 0; - const std::string attr_name = attribute->is_undefined() ? "" : attribute->as_string().str(); - std::vector arr = cast_val(args.get_pos(0))->as_array(); // copy + std::vector arr = val->as_array(); // copy std::sort(arr.begin(), arr.end(),[&](const value & a, const value & b) { value val_a = a; value val_b = b; @@ -884,22 +899,23 @@ const func_builtins & value_array_t::get_builtins() const { if (attr_is_int && is_val(a) && is_val(b)) { val_a = a->at(attr_int); val_b = b->at(attr_int); - } else if (!attr_is_int && !attr_name.empty() && is_val(a) && is_val(b)) { - val_a = a->at(attr_name); - val_b = b->at(attr_name); + } else if (!attr_is_int && is_val(a) && is_val(b)) { + val_a = a->at(attribute); + val_b = b->at(attribute); } else { - throw raised_exception("sort: unsupported object attribute comparison"); + throw raised_exception("sort: unsupported object attribute comparison between " + a->type() + " and " + b->type()); } } return value_compare(val_a, val_b, reverse ? value_compare_op::gt : value_compare_op::lt); }); - return mk_val(arr); + return is_val(val) ? mk_val(std::move(arr)) : mk_val(std::move(arr)); }}, {"reverse", [](const func_args & args) -> value { args.ensure_vals(); - std::vector arr = cast_val(args.get_pos(0))->as_array(); // copy + value val = args.get_pos(0); + std::vector arr = val->as_array(); // copy std::reverse(arr.begin(), arr.end()); - return mk_val(arr); + return is_val(val) ? mk_val(std::move(arr)) : mk_val(std::move(arr)); }}, {"unique", [](const func_args &) -> value { throw not_implemented_exception("Array unique builtin not implemented"); @@ -930,7 +946,7 @@ const func_builtins & value_object_t::get_builtins() const { default_val = args.get_pos(2); } const value obj = args.get_pos(0); - std::string key = args.get_pos(1)->as_string().str(); + const value key = args.get_pos(1); return obj->at(key, default_val); }}, {"keys", [](const func_args & args) -> value { @@ -938,7 +954,7 @@ const func_builtins & value_object_t::get_builtins() const { const auto & obj = args.get_pos(0)->as_ordered_object(); auto result = mk_val(); for (const auto & pair : obj) { - result->push_back(mk_val(pair.first)); + result->push_back(pair.first); } return result; }}, @@ -956,15 +972,16 @@ const func_builtins & value_object_t::get_builtins() const { const auto & obj = args.get_pos(0)->as_ordered_object(); auto result = mk_val(); for (const auto & pair : obj) { - auto item = mk_val(); - item->push_back(mk_val(pair.first)); - item->push_back(pair.second); + auto item = mk_val(pair); result->push_back(std::move(item)); } return result; }}, {"tojson", tojson}, - {"string", tojson}, + {"string", [](const func_args & args) -> value { + args.ensure_vals(); + return mk_val(args.get_pos(0)->as_string()); + }}, {"length", [](const func_args & args) -> value { args.ensure_vals(); const auto & obj = args.get_pos(0)->as_ordered_object(); @@ -985,11 +1002,11 @@ const func_builtins & value_object_t::get_builtins() const { const bool reverse = val_reverse->as_bool(); // undefined == false const bool by_value = is_val(val_by) && val_by->as_string().str() == "value" ? true : false; auto result = mk_val(val_input); // copy - std::sort(result->val_obj.ordered.begin(), result->val_obj.ordered.end(), [&](const auto & a, const auto & b) { + std::sort(result->val_obj.begin(), result->val_obj.end(), [&](const auto & a, const auto & b) { if (by_value) { return value_compare(a.second, b.second, reverse ? value_compare_op::gt : value_compare_op::lt); } else { - return reverse ? a.first > b.first : a.first < b.first; + return value_compare(a.first, b.first, reverse ? value_compare_op::gt : value_compare_op::lt); } }); return result; @@ -1005,7 +1022,12 @@ const func_builtins & value_none_t::get_builtins() const { static const func_builtins builtins = { {"default", default_value}, {"tojson", tojson}, - {"string", [](const func_args &) -> value { return mk_val("None"); }} + {"string", [](const func_args &) -> value { + return mk_val("None"); + }}, + {"safe", [](const func_args &) -> value { + return mk_val("None"); + }}, }; return builtins; } @@ -1014,10 +1036,33 @@ const func_builtins & value_none_t::get_builtins() const { const func_builtins & value_undefined_t::get_builtins() const { static const func_builtins builtins = { {"default", default_value}, - {"tojson", [](const func_args & args) -> value { - args.ensure_vals(); - return mk_val("null"); - }}, + {"capitalize", empty_value_fn}, + {"first", empty_value_fn}, + {"items", empty_value_fn}, + {"join", empty_value_fn}, + {"last", empty_value_fn}, + {"length", empty_value_fn}, + {"list", empty_value_fn}, + {"lower", empty_value_fn}, + {"map", empty_value_fn}, + {"max", empty_value_fn}, + {"min", empty_value_fn}, + {"reject", empty_value_fn}, + {"rejectattr", empty_value_fn}, + {"replace", empty_value_fn}, + {"reverse", empty_value_fn}, + {"safe", empty_value_fn}, + {"select", empty_value_fn}, + {"selectattr", empty_value_fn}, + {"sort", empty_value_fn}, + {"string", empty_value_fn}, + {"strip", empty_value_fn}, + {"sum", empty_value_fn}, + {"title", empty_value_fn}, + {"truncate", empty_value_fn}, + {"unique", empty_value_fn}, + {"upper", empty_value_fn}, + {"wordcount", empty_value_fn}, }; return builtins; } @@ -1134,6 +1179,8 @@ void global_from_json(context & ctx, const nlohmann::ordered_json & json_obj, bo } } +// recursively convert value to JSON string +// TODO: avoid circular references static void value_to_json_internal(std::ostringstream & oss, const value & val, int curr_lvl, int indent, const std::string_view item_sep, const std::string_view key_sep) { auto indent_str = [indent, curr_lvl]() -> std::string { return (indent > 0) ? std::string(curr_lvl * indent, ' ') : ""; @@ -1196,7 +1243,8 @@ static void value_to_json_internal(std::ostringstream & oss, const value & val, size_t i = 0; for (const auto & pair : obj) { oss << indent_str() << (indent > 0 ? std::string(indent, ' ') : ""); - oss << "\"" << pair.first << "\"" << key_sep; + value_to_json_internal(oss, mk_val(pair.first->as_string().str()), curr_lvl + 1, indent, item_sep, key_sep); + oss << key_sep; value_to_json_internal(oss, pair.second, curr_lvl + 1, indent, item_sep, key_sep); if (i < obj.size() - 1) { oss << item_sep; @@ -1219,4 +1267,19 @@ std::string value_to_json(const value & val, int indent, const std::string_view return oss.str(); } +// TODO: avoid circular references +std::string value_to_string_repr(const value & val) { + if (is_val(val)) { + const std::string val_str = val->as_string().str(); + + if (val_str.find('\'') != std::string::npos) { + return value_to_json(val); + } else { + return "'" + val_str + "'"; + } + } else { + return val->as_repr(); + } +} + } // namespace jinja diff --git a/common/jinja/value.h b/common/jinja/value.h index ccb05c6fd4..a2f92d2c69 100644 --- a/common/jinja/value.h +++ b/common/jinja/value.h @@ -1,8 +1,10 @@ #pragma once #include "string.h" +#include "utils.h" #include +#include #include #include #include @@ -93,7 +95,8 @@ void global_from_json(context & ctx, const T_JSON & json_obj, bool mark_input); struct func_args; // function argument values -using func_handler = std::function; +using func_hptr = value(const func_args &); +using func_handler = std::function; using func_builtins = std::map; enum value_compare_op { eq, ge, gt, lt, ne }; @@ -103,28 +106,9 @@ struct value_t { int64_t val_int; double val_flt; string val_str; - bool val_bool; std::vector val_arr; - - struct map { - // once set to true, all keys must be numeric - // caveat: we only allow either all numeric keys or all non-numeric keys - // for now, this only applied to for_statement in case of iterating over object keys/items - bool is_key_numeric = false; - std::map unordered; - std::vector> ordered; - void insert(const std::string & key, const value & val) { - if (unordered.find(key) != unordered.end()) { - // if key exists, remove from ordered list - ordered.erase(std::remove_if(ordered.begin(), ordered.end(), - [&](const std::pair & p) { return p.first == key; }), - ordered.end()); - } - unordered[key] = val; - ordered.push_back({key, val}); - } - } val_obj; + std::vector> val_obj; func_handler val_func; @@ -139,6 +123,7 @@ struct value_t { value_t(const value_t &) = default; virtual ~value_t() = default; + // Note: only for debugging and error reporting purposes virtual std::string type() const { return ""; } virtual int64_t as_int() const { throw std::runtime_error(type() + " is not an int value"); } @@ -146,7 +131,7 @@ struct value_t { virtual string as_string() const { throw std::runtime_error(type() + " is not a string value"); } virtual bool as_bool() const { throw std::runtime_error(type() + " is not a bool value"); } virtual const std::vector & as_array() const { throw std::runtime_error(type() + " is not an array value"); } - virtual const std::vector> & as_ordered_object() const { throw std::runtime_error(type() + " is not an object value"); } + virtual const std::vector> & as_ordered_object() const { throw std::runtime_error(type() + " is not an object value"); } virtual value invoke(const func_args &) const { throw std::runtime_error(type() + " is not a function value"); } virtual bool is_none() const { return false; } virtual bool is_undefined() const { return false; } @@ -154,43 +139,66 @@ struct value_t { throw std::runtime_error("No builtins available for type " + type()); } - virtual bool has_key(const std::string & key) { - return val_obj.unordered.find(key) != val_obj.unordered.end(); - } - virtual value & at(const std::string & key, value & default_val) { - auto it = val_obj.unordered.find(key); - if (it == val_obj.unordered.end()) { - return default_val; - } - return val_obj.unordered.at(key); - } - virtual value & at(const std::string & key) { - auto it = val_obj.unordered.find(key); - if (it == val_obj.unordered.end()) { - throw std::runtime_error("Key '" + key + "' not found in value of type " + type()); - } - return val_obj.unordered.at(key); - } - virtual value & at(int64_t index, value & default_val) { - if (index < 0) { - index += val_arr.size(); - } - if (index < 0 || static_cast(index) >= val_arr.size()) { - return default_val; - } - return val_arr[index]; - } - virtual value & at(int64_t index) { - if (index < 0) { - index += val_arr.size(); - } - if (index < 0 || static_cast(index) >= val_arr.size()) { - throw std::runtime_error("Index " + std::to_string(index) + " out of bounds for array of size " + std::to_string(val_arr.size())); - } - return val_arr[index]; - } + virtual bool has_key(const value &) { throw std::runtime_error(type() + " is not an object value"); } + virtual void insert(const value & /* key */, const value & /* val */) { throw std::runtime_error(type() + " is not an object value"); } + virtual value & at(const value & /* key */, value & /* default_val */) { throw std::runtime_error(type() + " is not an object value"); } + virtual value & at(const value & /* key */) { throw std::runtime_error(type() + " is not an object value"); } + virtual value & at(const std::string & /* key */, value & /* default_val */) { throw std::runtime_error(type() + " is not an object value"); } + virtual value & at(const std::string & /* key */) { throw std::runtime_error(type() + " is not an object value"); } + virtual value & at(int64_t /* idx */, value & /* default_val */) { throw std::runtime_error(type() + " is not an array value"); } + virtual value & at(int64_t /* idx */) { throw std::runtime_error(type() + " is not an array value"); } + virtual bool is_numeric() const { return false; } + virtual bool is_hashable() const { return false; } + virtual bool is_immutable() const { return true; } + virtual hasher unique_hash() const noexcept = 0; + // TODO: C++20 <=> operator + // NOTE: We are treating == as equivalent (for normal comparisons) and != as strict nonequal (for strict (is) comparisons) + virtual bool operator==(const value_t & other) const { return equivalent(other); } + virtual bool operator!=(const value_t & other) const { return nonequal(other); } + + // Note: only for debugging purposes virtual std::string as_repr() const { return as_string().str(); } + +protected: + virtual bool equivalent(const value_t &) const = 0; + virtual bool nonequal(const value_t & other) const { return !equivalent(other); } +}; + +// +// utils +// + +const func_builtins & global_builtins(); + +std::string value_to_json(const value & val, int indent = -1, const std::string_view item_sep = ", ", const std::string_view key_sep = ": "); + +// Note: only used for debugging purposes +std::string value_to_string_repr(const value & val); + +struct not_implemented_exception : public std::runtime_error { + not_implemented_exception(const std::string & msg) : std::runtime_error("NotImplemented: " + msg) {} +}; + +struct value_hasher { + size_t operator()(const value & val) const noexcept { + return val->unique_hash().digest(); + } +}; + +struct value_equivalence { + bool operator()(const value & lhs, const value & rhs) const { + return *lhs == *rhs; + } + bool operator()(const std::pair & lhs, const std::pair & rhs) const { + return *(lhs.first) == *(rhs.first) && *(lhs.second) == *(rhs.second); + } +}; + +struct value_equality { + bool operator()(const value & lhs, const value & rhs) const { + return !(*lhs != *rhs); + } }; // @@ -198,24 +206,49 @@ struct value_t { // struct value_int_t : public value_t { - value_int_t(int64_t v) { val_int = v; } + value_int_t(int64_t v) { + val_int = v; + val_flt = static_cast(v); + if (static_cast(val_flt) != v) { + val_flt = v < 0 ? -INFINITY : INFINITY; + } + } virtual std::string type() const override { return "Integer"; } virtual int64_t as_int() const override { return val_int; } - virtual double as_float() const override { return static_cast(val_int); } + virtual double as_float() const override { return val_flt; } virtual string as_string() const override { return std::to_string(val_int); } virtual bool as_bool() const override { return val_int != 0; } virtual const func_builtins & get_builtins() const override; + virtual bool is_numeric() const override { return true; } + virtual bool is_hashable() const override { return true; } + virtual hasher unique_hash() const noexcept override { + return hasher(typeid(*this)) + .update(&val_int, sizeof(val_int)) + .update(&val_flt, sizeof(val_flt)); + } +protected: + virtual bool equivalent(const value_t & other) const override { + return other.is_numeric() && val_int == other.val_int && val_flt == other.val_flt; + } + virtual bool nonequal(const value_t & other) const override { + return !(typeid(*this) == typeid(other) && val_int == other.val_int); + } }; using value_int = std::shared_ptr; struct value_float_t : public value_t { - value_float_t(double v) { val_flt = v; } + value val; + value_float_t(double v) { + val_flt = v; + val_int = std::isfinite(v) ? static_cast(v) : 0; + val = mk_val(val_int); + } virtual std::string type() const override { return "Float"; } virtual double as_float() const override { return val_flt; } - virtual int64_t as_int() const override { return static_cast(val_flt); } + virtual int64_t as_int() const override { return val_int; } virtual string as_string() const override { std::string out = std::to_string(val_flt); out.erase(out.find_last_not_of('0') + 1, std::string::npos); // remove trailing zeros @@ -226,6 +259,24 @@ struct value_float_t : public value_t { return val_flt != 0.0; } virtual const func_builtins & get_builtins() const override; + virtual bool is_numeric() const override { return true; } + virtual bool is_hashable() const override { return true; } + virtual hasher unique_hash() const noexcept override { + if (static_cast(val_int) == val_flt) { + return val->unique_hash(); + } else { + return hasher(typeid(*this)) + .update(&val_int, sizeof(val_int)) + .update(&val_flt, sizeof(val_flt)); + } + } +protected: + virtual bool equivalent(const value_t & other) const override { + return other.is_numeric() && val_int == other.val_int && val_flt == other.val_flt; + } + virtual bool nonequal(const value_t & other) const override { + return !(typeid(*this) == typeid(other) && val_flt == other.val_flt); + } }; using value_float = std::shared_ptr; @@ -247,19 +298,49 @@ struct value_string_t : public value_t { return val_str.length() > 0; } virtual const func_builtins & get_builtins() const override; + virtual bool is_hashable() const override { return true; } + virtual hasher unique_hash() const noexcept override { + const auto type_hash = typeid(*this).hash_code(); + auto hash = hasher(); + hash.update(&type_hash, sizeof(type_hash)); + val_str.hash_update(hash); + return hash; + } void mark_input() { val_str.mark_input(); } +protected: + virtual bool equivalent(const value_t & other) const override { + return typeid(*this) == typeid(other) && val_str.str() == other.val_str.str(); + } }; using value_string = std::shared_ptr; struct value_bool_t : public value_t { - value_bool_t(bool v) { val_bool = v; } + value val; + value_bool_t(bool v) { + val_int = static_cast(v); + val_flt = static_cast(v); + val = mk_val(val_int); + } virtual std::string type() const override { return "Boolean"; } - virtual bool as_bool() const override { return val_bool; } - virtual string as_string() const override { return std::string(val_bool ? "True" : "False"); } + virtual int64_t as_int() const override { return val_int; } + virtual bool as_bool() const override { return val_int; } + virtual string as_string() const override { return std::string(val_int ? "True" : "False"); } virtual const func_builtins & get_builtins() const override; + virtual bool is_numeric() const override { return true; } + virtual bool is_hashable() const override { return true; } + virtual hasher unique_hash() const noexcept override { + return val->unique_hash(); + } +protected: + virtual bool equivalent(const value_t & other) const override { + return other.is_numeric() && val_int == other.val_int && val_flt == other.val_flt; + } + virtual bool nonequal(const value_t & other) const override { + return !(typeid(*this) == typeid(other) && val_int == other.val_int); + } }; using value_bool = std::shared_ptr; @@ -269,13 +350,34 @@ struct value_array_t : public value_t { value_array_t(value & v) { val_arr = v->val_arr; } + value_array_t(std::vector && arr) { + val_arr = arr; + } value_array_t(const std::vector & arr) { val_arr = arr; } - void reverse() { std::reverse(val_arr.begin(), val_arr.end()); } - void push_back(const value & val) { val_arr.push_back(val); } - void push_back(value && val) { val_arr.push_back(std::move(val)); } + void reverse() { + if (is_immutable()) { + throw std::runtime_error("Attempting to modify immutable type"); + } + std::reverse(val_arr.begin(), val_arr.end()); + } + void push_back(const value & val) { + if (is_immutable()) { + throw std::runtime_error("Attempting to modify immutable type"); + } + val_arr.push_back(val); + } + void push_back(value && val) { + if (is_immutable()) { + throw std::runtime_error("Attempting to modify immutable type"); + } + val_arr.push_back(std::move(val)); + } value pop_at(int64_t index) { + if (is_immutable()) { + throw std::runtime_error("Attempting to modify immutable type"); + } if (index < 0) { index = static_cast(val_arr.size()) + index; } @@ -287,64 +389,225 @@ struct value_array_t : public value_t { return val; } virtual std::string type() const override { return "Array"; } + virtual bool is_immutable() const override { return false; } virtual const std::vector & as_array() const override { return val_arr; } virtual string as_string() const override { + const bool immutable = is_immutable(); std::ostringstream ss; - ss << "["; + ss << (immutable ? "(" : "["); for (size_t i = 0; i < val_arr.size(); i++) { if (i > 0) ss << ", "; - ss << val_arr.at(i)->as_repr(); + value val = val_arr.at(i); + ss << value_to_string_repr(val); } - ss << "]"; + if (immutable && val_arr.size() == 1) { + ss << ","; + } + ss << (immutable ? ")" : "]"); return ss.str(); } virtual bool as_bool() const override { return !val_arr.empty(); } + virtual value & at(int64_t index, value & default_val) override { + if (index < 0) { + index += val_arr.size(); + } + if (index < 0 || static_cast(index) >= val_arr.size()) { + return default_val; + } + return val_arr[index]; + } + virtual value & at(int64_t index) override { + if (index < 0) { + index += val_arr.size(); + } + if (index < 0 || static_cast(index) >= val_arr.size()) { + throw std::runtime_error("Index " + std::to_string(index) + " out of bounds for array of size " + std::to_string(val_arr.size())); + } + return val_arr[index]; + } virtual const func_builtins & get_builtins() const override; + virtual bool is_hashable() const override { + if (std::all_of(val_arr.begin(), val_arr.end(), [&](auto & val) -> bool { + return val->is_immutable() && val->is_hashable(); + })) { + return true; + } + return false; + } + virtual hasher unique_hash() const noexcept override { + auto hash = hasher(typeid(*this)); + for (const auto & val : val_arr) { + // must use digest to prevent problems from "concatenation" property of hasher + // for ex. hash of [ "ab", "c" ] should be different from [ "a", "bc" ] + const size_t val_hash = val->unique_hash().digest(); + hash.update(&val_hash, sizeof(size_t)); + } + return hash; + } +protected: + virtual bool equivalent(const value_t & other) const override { + return typeid(*this) == typeid(other) && is_hashable() && other.is_hashable() && std::equal(val_arr.begin(), val_arr.end(), other.val_arr.begin(), value_equivalence()); + } }; using value_array = std::shared_ptr; +struct value_tuple_t : public value_array_t { + value_tuple_t(value & v) { + val_arr = v->val_arr; + } + value_tuple_t(std::vector && arr) { + val_arr = arr; + } + value_tuple_t(const std::vector & arr) { + val_arr = arr; + } + value_tuple_t(const std::pair & pair) { + val_arr.push_back(pair.first); + val_arr.push_back(pair.second); + } + virtual std::string type() const override { return "Tuple"; } + virtual bool is_immutable() const override { return true; } +}; +using value_tuple = std::shared_ptr; + + struct value_object_t : public value_t { + std::unordered_map unordered; bool has_builtins = true; // context and loop objects do not have builtins value_object_t() = default; value_object_t(value & v) { val_obj = v->val_obj; - } - value_object_t(const std::map & obj) { - for (const auto & pair : obj) { - val_obj.insert(pair.first, pair.second); + for (const auto & pair : val_obj) { + unordered[pair.first] = pair.second; } } - value_object_t(const std::vector> & obj) { + value_object_t(const std::map & obj) { for (const auto & pair : obj) { - val_obj.insert(pair.first, pair.second); + insert(pair.first, pair.second); + } + } + value_object_t(const std::vector> & obj) { + for (const auto & pair : obj) { + insert(pair.first, pair.second); } } void insert(const std::string & key, const value & val) { - val_obj.insert(key, val); + insert(mk_val(key), val); } virtual std::string type() const override { return "Object"; } - virtual const std::vector> & as_ordered_object() const override { return val_obj.ordered; } + virtual bool is_immutable() const override { return false; } + virtual const std::vector> & as_ordered_object() const override { return val_obj; } + virtual string as_string() const override { + std::ostringstream ss; + ss << "{"; + for (size_t i = 0; i < val_obj.size(); i++) { + if (i > 0) ss << ", "; + auto & [key, val] = val_obj.at(i); + ss << value_to_string_repr(key) << ": " << value_to_string_repr(val); + } + ss << "}"; + return ss.str(); + } virtual bool as_bool() const override { - return !val_obj.unordered.empty(); + return !unordered.empty(); + } + virtual bool has_key(const value & key) override { + if (!key->is_immutable() || !key->is_hashable()) { + throw std::runtime_error("Object key of unhashable type: " + key->type()); + } + return unordered.find(key) != unordered.end(); + } + virtual void insert(const value & key, const value & val) override { + bool replaced = false; + if (is_immutable()) { + throw std::runtime_error("Attempting to modify immutable type"); + } + if (has_key(key)) { + // if key exists, replace value in ordered list instead of appending + for (auto & pair : val_obj) { + if (*(pair.first) == *key) { + pair.second = val; + replaced = true; + break; + } + } + } + unordered[key] = val; + if (!replaced) { + val_obj.push_back({key, val}); + } + } + virtual value & at(const value & key, value & default_val) override { + if (!has_key(key)) { + return default_val; + } + return unordered.at(key); + } + virtual value & at(const value & key) override { + if (!has_key(key)) { + throw std::runtime_error("Key '" + key->as_string().str() + "' not found in value of type " + type()); + } + return unordered.at(key); + } + virtual value & at(const std::string & key, value & default_val) override { + value key_val = mk_val(key); + return at(key_val, default_val); + } + virtual value & at(const std::string & key) override { + value key_val = mk_val(key); + return at(key_val); } virtual const func_builtins & get_builtins() const override; + virtual bool is_hashable() const override { + if (std::all_of(val_obj.begin(), val_obj.end(), [&](auto & pair) -> bool { + const auto & val = pair.second; + return val->is_immutable() && val->is_hashable(); + })) { + return true; + } + return false; + } + virtual hasher unique_hash() const noexcept override { + auto hash = hasher(typeid(*this)); + for (const auto & [key, val] : val_obj) { + // must use digest to prevent problems from "concatenation" property of hasher + // for ex. hash of key="ab", value="c" should be different from key="a", value="bc" + const size_t key_hash = key->unique_hash().digest(); + const size_t val_hash = val->unique_hash().digest(); + hash.update(&key_hash, sizeof(key_hash)); + hash.update(&val_hash, sizeof(val_hash)); + } + return hash; + } +protected: + virtual bool equivalent(const value_t & other) const override { + return typeid(*this) == typeid(other) && is_hashable() && other.is_hashable() && std::equal(val_obj.begin(), val_obj.end(), other.val_obj.begin(), value_equivalence()); + } }; using value_object = std::shared_ptr; // -// null and undefined types +// none and undefined types // struct value_none_t : public value_t { virtual std::string type() const override { return "None"; } virtual bool is_none() const override { return true; } virtual bool as_bool() const override { return false; } - virtual string as_string() const override { return string("None"); } + virtual string as_string() const override { return string(type()); } virtual std::string as_repr() const override { return type(); } virtual const func_builtins & get_builtins() const override; + virtual bool is_hashable() const override { return true; } + virtual hasher unique_hash() const noexcept override { + return hasher(typeid(*this)); + } +protected: + virtual bool equivalent(const value_t & other) const override { + return typeid(*this) == typeid(other); + } }; using value_none = std::shared_ptr; @@ -356,6 +619,13 @@ struct value_undefined_t : public value_t { virtual bool as_bool() const override { return false; } virtual std::string as_repr() const override { return type(); } virtual const func_builtins & get_builtins() const override; + virtual hasher unique_hash() const noexcept override { + return hasher(typeid(*this)); + } +protected: + virtual bool equivalent(const value_t & other) const override { + return is_undefined() == other.is_undefined(); + } }; using value_undefined = std::shared_ptr; @@ -436,7 +706,23 @@ struct value_func_t : public value_t { return val_func(new_args); } virtual std::string type() const override { return "Function"; } - virtual std::string as_repr() const override { return type(); } + virtual std::string as_repr() const override { return type() + "<" + name + ">(" + (arg0 ? arg0->as_repr() : "") + ")"; } + virtual bool is_hashable() const override { return false; } + virtual hasher unique_hash() const noexcept override { + // Note: this is unused for now, we don't support function as object keys + // use function pointer as unique identifier + const auto target = val_func.target(); + return hasher(typeid(*this)).update(&target, sizeof(target)); + } +protected: + virtual bool equivalent(const value_t & other) const override { + // Note: this is unused for now, we don't support function as object keys + // compare function pointers + // (val_func == other.val_func does not work as std::function::operator== is only used for nullptr check) + const auto target_this = this->val_func.target(); + const auto target_other = other.val_func.target(); + return typeid(*this) == typeid(other) && target_this == target_other; + } }; using value_func = std::shared_ptr; @@ -447,18 +733,21 @@ struct value_kwarg_t : public value_t { value_kwarg_t(const std::string & k, const value & v) : key(k), val(v) {} virtual std::string type() const override { return "KwArg"; } virtual std::string as_repr() const override { return type(); } + virtual bool is_hashable() const override { return true; } + virtual hasher unique_hash() const noexcept override { + const auto type_hash = typeid(*this).hash_code(); + auto hash = val->unique_hash(); + hash.update(&type_hash, sizeof(type_hash)) + .update(key.data(), key.size()); + return hash; + } +protected: + virtual bool equivalent(const value_t & other) const override { + const value_kwarg_t & other_val = static_cast(other); + return typeid(*this) == typeid(other) && key == other_val.key && val == other_val.val; + } }; using value_kwarg = std::shared_ptr; -// utils - -const func_builtins & global_builtins(); -std::string value_to_json(const value & val, int indent = -1, const std::string_view item_sep = ", ", const std::string_view key_sep = ": "); - -struct not_implemented_exception : public std::runtime_error { - not_implemented_exception(const std::string & msg) : std::runtime_error("NotImplemented: " + msg) {} -}; - - } // namespace jinja diff --git a/common/ngram-cache.cpp b/common/ngram-cache.cpp index d1a4d84c40..dce54b3647 100644 --- a/common/ngram-cache.cpp +++ b/common/ngram-cache.cpp @@ -192,12 +192,12 @@ void common_ngram_cache_draft( break; } - LOG(" - draft candidate: token=%d\n", drafted_token); + LOG_DBG(" - draft candidate: token=%d\n", drafted_token); draft.push_back(drafted_token); } } -void common_ngram_cache_save(common_ngram_cache & ngram_cache, std::string & filename) { +void common_ngram_cache_save(common_ngram_cache & ngram_cache, const std::string & filename) { std::ofstream file_out(filename, std::ios::binary); for (std::pair item : ngram_cache) { const common_ngram ngram = item.first; @@ -217,10 +217,9 @@ void common_ngram_cache_save(common_ngram_cache & ngram_cache, std::string & fil file_out.write(reinterpret_cast(&count), sizeof(int32_t)); } } - } -common_ngram_cache common_ngram_cache_load(std::string & filename) { +common_ngram_cache common_ngram_cache_load(const std::string & filename) { std::ifstream hashmap_file(filename, std::ios::binary); if (!hashmap_file) { throw std::ifstream::failure("Unable to open file " + filename); diff --git a/common/ngram-cache.h b/common/ngram-cache.h index dfe012abe4..6e7cfea966 100644 --- a/common/ngram-cache.h +++ b/common/ngram-cache.h @@ -88,12 +88,12 @@ void common_ngram_cache_draft( // Save an ngram cache to a file. // ngram_cache: the ngram cache to save. // filename: the path under which to save the ngram cache. -void common_ngram_cache_save(common_ngram_cache & ngram_cache, std::string & filename); +void common_ngram_cache_save(common_ngram_cache & ngram_cache, const std::string & filename); // Load an ngram cache saved with common_ngram_cache_save. // filename: the path from which to load the ngram cache. // returns: an ngram cache containing the information saved to filename. -common_ngram_cache common_ngram_cache_load(std::string & filename); +common_ngram_cache common_ngram_cache_load(const std::string & filename); // Merge two ngram caches. // ngram_cache_target: the ngram cache to which to add the information from ngram_cache_add. diff --git a/common/ngram-map.cpp b/common/ngram-map.cpp new file mode 100644 index 0000000000..930e7a3c10 --- /dev/null +++ b/common/ngram-map.cpp @@ -0,0 +1,367 @@ +#include "common.h" +#include "log.h" +#include "ngram-map.h" + +#include +#include +#include +#include + +// n-gram simple +// + +/** + * Perform speculative generation using the model's own token history. + * Searches for a matching pattern in the token history and returns draft tokens. + * + * @param state Current state of this implementation + * @param tokens Token history to search in + * @param sampled Last sampled token + * @return Vector of draft tokens, empty if no matching pattern is found + */ +llama_tokens common_ngram_simple_draft( + common_ngram_simple_state & state, + const llama_tokens & tokens, llama_token sampled) { + + // Simple implementation of self-speculative decoding without a draft model. + // + const size_t cur_len = tokens.size(); + // Only check every check_rate tokens to save compute + // i.e., perform check if (cur_len - idx_last_check) >= check_rate + if (state.idx_last_check + state.config.check_rate > cur_len) { + llama_tokens draft_tokens; + return draft_tokens; + } + + size_t n_draft_min = state.config.size_ngram; // size of n-gram to lookup in token history + size_t n_draft_max = state.config.size_mgram; // the m-gram following the found n-gram is used for draft + + // vector for tokens we want to verify. + // return empty vector if there is no match. + llama_tokens draft_tokens; + + // We need at least n_draft_min + n_draft_max + 1 tokens. + if (cur_len <= static_cast(n_draft_min + n_draft_max + 1)) { + return draft_tokens; + } + + // pattern search + llama_tokens pattern; + pattern.reserve(n_draft_min); + for (size_t j = cur_len - n_draft_min + 1; j < cur_len; ++j) { + pattern.push_back(tokens[j]); + } + pattern.push_back(sampled); // add the last token to the pattern + + // We do a search in the token history. + state.idx_last_check = cur_len; + + size_t match_pos = 0; // we ignore position 0, position 0 == no match + // search backwards, but skip the current match (we are currently there) + for (size_t j = cur_len - n_draft_min - 1; j > 0; --j) { + bool match = true; + for (size_t k = 0; k < pattern.size(); ++k) { + if (tokens[j + k] != pattern[k]) { + match = false; + break; + } + } + if (match) { + match_pos = j; + break; + } + } + if (match_pos == 0) { + return draft_tokens; + } + + const size_t copy_max = std::min( + n_draft_max, + cur_len - (match_pos + n_draft_min) + ); + if (copy_max < n_draft_min) { + return draft_tokens; + } + LOG_DBG("%s: #tokens = %zu: found matching pattern at pos %zu, length %zu, draft length %zu\n", + __func__, cur_len, + match_pos, pattern.size(), copy_max); + + draft_tokens.reserve(copy_max); + for (size_t j = 0; j < copy_max; ++j) { + draft_tokens.push_back(tokens[match_pos + n_draft_min + j]); + } + return draft_tokens; +} + + +// n-gram map +// + +// maximum number of counted values of a ngram map value. +#define COMMON_NGRAM_MAX_VALUE_COUNT 16380 + +static std::string common_tokens_to_str(const llama_tokens & inp, size_t start, size_t length); + +void common_ngram_map_draft(common_ngram_map & map, + const llama_tokens & inp, llama_token sampled, + llama_tokens & draft) { + // reset last key and value. + map.last_draft_created = false; + map.last_draft_key_idx = 0; + map.last_draft_value_idx = 0; + + const size_t cur_len = inp.size(); + const uint16_t n = map.size_key; + const uint16_t m = map.size_value; + if (cur_len < static_cast(2 * n + m)) { + return; + } + + // Only check every check_rate tokens to save compute + // i.e., perform check if (cur_len - idx_last_check) >= check_rate + if (map.idx_last_check + map.check_rate > cur_len) { + return; + } + map.idx_last_check = cur_len; + + // search pattern, the key n-gram + std::vector key_tokens; + key_tokens.reserve(n); + for (size_t j = cur_len - n + 1; j < cur_len; ++j) { + key_tokens.push_back(inp[j]); + } + key_tokens.push_back(sampled); + + // search for the key in the map + size_t match_pos = 0; + for (size_t j = cur_len - n - m - 1; j > 0; --j) { + bool match = true; + for (size_t k = 0; k < n; ++k) { + if (inp[j + k] != key_tokens[k]) { + match = false; + break; + } + } + if (match) { + match_pos = j; + break; + } + } + if (match_pos > 0) { + LOG_INF("%s: cur_len = %zu, n = %d, m = %d, sz_tkns = %zu, sampled = %d, match_pos = %zu\n", __func__, + cur_len, n, m, key_tokens.size(), sampled, match_pos); + } + + if (match_pos == 0) { + return; + } + + // We have a match, now we look for the statistics of the key. + size_t key_offset = map.keys.size(); // offset in the map + // We iterate through the std::vector map->keys. + for (size_t i = 0; i < map.keys.size(); ++i) { + bool match = true; + for (size_t j = 0; j < n; ++j) { + if (inp[map.keys[i].key_idx + j] != key_tokens[j]) { + match = false; + break; + } + } + if (match) { + key_offset = i; + break; + } + } + if (key_offset == map.keys.size()) { + // We create a new key-entry, it will get offset key_offset. + common_ngram_map_key new_key; + new_key.key_idx = match_pos; + new_key.stat_idx = 0; + new_key.key_num = 0; + for (int i = 0; i < COMMON_NGRAM_MAX_VALUES; ++i) { + new_key.values[i].value_num = 0; + new_key.values[i].n_accepted = m; + } + map.keys.push_back(new_key); + } + + // our key n-gram: + common_ngram_map_key & curr_key = map.keys[key_offset]; + + // update number of key hits + curr_key.key_num = (uint16_t) std::min((int) map.keys[key_offset].key_num + 1, + (int) COMMON_NGRAM_MAX_VALUE_COUNT); + + if (map.key_only) { + // simple mode: + // Fill in the draft with the m tokens following the key. + // We work with value values[0] only. + int n_draft_tokens = std::min((int) m, (int) curr_key.values[0].n_accepted); + + for (int i = 0; i < n_draft_tokens; ++i) { + draft.push_back(inp[match_pos + n + i]); + } + + LOG_INF("%s: key_offset = %zu, key_num = %d, draft.size = %zu\n", __func__, + key_offset, curr_key.key_num, draft.size()); + + map.last_draft_created = false; + map.last_draft_key_idx = key_offset; + map.last_draft_value_idx = 0; // value 0 is used for simple mode + return; + } + + if (curr_key.key_num < map.min_hits) { + // not enough hits to consider this a good draft + LOG_DBG("%s: key_offset = %zu, key_num = %d, min_hits = %d, no draft\n", __func__, + key_offset, curr_key.key_num, map.min_hits); + return; + } + + // complex mode: examine the different m-grams after this key n-gram. + // + + // determine all (max COMMON_NGRAM_MAX_VALUES) m-grams after the key n-gram. + for (size_t i = curr_key.stat_idx; i <= match_pos; ++i) { + // begins the key n-gram at index i? + bool match_key = true; + for (size_t k = 0; k < n; ++k) { + if (inp[i + k] != key_tokens[k]) { + match_key = false; + break; + } + } + if (!match_key) { + continue; + } + + // Do we haven a existing value m-gram or a new one after the key at index i? + size_t idx_begin_value_key = i + n; + int idx_value = -1; + for (int v = 0; v < COMMON_NGRAM_MAX_VALUES; ++v) { + size_t idx_begin_value_v = curr_key.values[v].value_idx; + if (idx_begin_value_v == 0) { + // We found an empty value slot => we found a new value m-gram after the key n-gram. + curr_key.values[v].value_idx = idx_begin_value_key; + curr_key.values[v].value_num = 0; + curr_key.values[v].n_accepted = m; + idx_value = v; + break; + } + bool match = true; + for (size_t j = 0; j < m; ++j) { + if (inp[idx_begin_value_key + j] != inp[idx_begin_value_v + j]) { + match = false; + break; + } + } + if (match) { + // We found an existing value m-gram after the key n-gram. + idx_value = v; + break; + } + } + if (idx_value >= 0) { + // We found a value m-gram of the key n-gram. + curr_key.values[idx_value].value_num = (uint16_t) std::min((int) curr_key.values[idx_value].value_num + 1, + (int) COMMON_NGRAM_MAX_VALUE_COUNT); + } + } + // the statistics are updated up to match_pos. + curr_key.stat_idx = match_pos; + + // Do we have a value we could use for the draft? + uint16_t max_occur = 0; + int slot_max = 0; + for (int v = 0; v < COMMON_NGRAM_MAX_VALUES; ++v) { + uint16_t curr_occur = curr_key.values[v].value_num; + if (curr_occur > max_occur) { + max_occur = curr_occur; + slot_max = v; + } + } + // What is sum of the other occurences? + uint32_t sum_occur = 0; + for (int v = 0; v < COMMON_NGRAM_MAX_VALUES; ++v) { + if (v == slot_max) { + continue; + } + uint16_t curr_occur = curr_key.values[v].value_num; + sum_occur += curr_occur; + } + + LOG_INF("%s: key_offset = %zu, max_occur = %d, sum_occur = %d, slot_max = %d [%zu/%d, %zu/%d, %zu/%d, %zu/%d]\n", __func__, + key_offset, + max_occur, sum_occur, slot_max, + curr_key.values[0].value_idx, curr_key.values[0].value_num, + curr_key.values[1].value_idx, curr_key.values[1].value_num, + curr_key.values[2].value_idx, curr_key.values[2].value_num, + curr_key.values[3].value_idx, curr_key.values[3].value_num + ); + // Print the tokens of the four values (if idx != 0), use LOG_INF + for (int v = 0; v < COMMON_NGRAM_MAX_VALUES; ++v) { + if (curr_key.values[v].value_idx != 0) { + LOG_INF("%s: value[%d] = %s\n", __func__, v, common_tokens_to_str(inp, curr_key.values[v].value_idx, m).c_str()); + } + } + + if (sum_occur > 0 && max_occur < 3 * sum_occur) { + // The most frequent value is not much more frequent than the other values. + // We do not use the draft. + return; + } + + // We use the most frequent value values[slot_max] for the draft. + // Fill in the draft with the m tokens following the key. + int n_draft_tokens = std::min((int) m, (int) curr_key.values[slot_max].n_accepted); + + for (int i = 0; i < n_draft_tokens; ++i) { + draft.push_back(inp[match_pos + n + i]); + } + + LOG_INF("%s: key_offset = %zu, slot_max = %d, key_num = %d, draft.size = %zu\n", __func__, + key_offset, slot_max, + curr_key.key_num, draft.size()); + + map.last_draft_created = true; + map.last_draft_key_idx = key_offset; + map.last_draft_value_idx = slot_max; // value used for draft generation. +} + +void common_ngram_map_accept(common_ngram_map & map, uint16_t n_accepted) { + if (!map.last_draft_created) { + return; + } + + // find the key and its chosen value. + const size_t key_idx = map.last_draft_key_idx; + const size_t val_idx = map.last_draft_value_idx; + + // find key corresponding to key_idx. + common_ngram_map_key & curr_key = map.keys[key_idx]; + // find value corresponding to val_idx. + struct common_ngram_map_value & curr_value = curr_key.values[val_idx]; // value used for draft generation. + + // update the value statistics + LOG_INF("common_ngram_map_send_accepted: n_accepted = %d, prev value_num = %d\n", + n_accepted, curr_value.n_accepted); + curr_value.n_accepted = n_accepted; +} + +// Helper functions. +// + +// Print the values of a sublist of `llama_tokens & inp` to a string in the form [v0, v1, v2, ...]. +std::string common_tokens_to_str(const llama_tokens & inp, size_t start, size_t length) { + std::ostringstream oss; + oss << '['; + for (size_t i = 0; i < length; ++i) { + if (i > 0) { + oss << ", "; + } + oss << inp[start + i]; + } + oss << ']'; + return oss.str(); +} + diff --git a/common/ngram-map.h b/common/ngram-map.h new file mode 100644 index 0000000000..bf91883f0c --- /dev/null +++ b/common/ngram-map.h @@ -0,0 +1,105 @@ +#pragma once +// +// common/ngram-map.h: structures used to manage a map from n-grams to a list of m-grams +// +// These structures are used to do a lookup of n-grams followed by m-grams in token history. +// +// There are two algorithms implemented: +// 1. ngram_simple: lookup of n-grams followed by m-grams in token history. +// 2. ngram_map: lookup of n-grams followed by m-grams in token history using a map. +// The map is a vector of key n-grams, and for each key n-gram there is a list of value m-grams. +// + +#include "llama.h" + +#include + +// n-gram simple +// + +// config of n-gram simple. +struct common_ngram_simple_config { + uint16_t size_ngram; // size of n-grams to lookup in self-mode + uint16_t size_mgram; // size of m-grams to draft in self-mode + uint16_t check_rate; // check for speculative decoding without draft model for each check_rate token +}; + +// current state (and config) of n-gram simple. +struct common_ngram_simple_state { + common_ngram_simple_config config; + + size_t idx_last_check = 0; // index of last check in context history (mutable) + + common_ngram_simple_state(const common_ngram_simple_config & config) + : config(config) {} +}; + +// Searches for a n-gram in the history and checks whether a draft sequence should be generated. +// state: the ngram simple state to search in. +// inp: the tokens generated so far. +// sampled: the token that was just sampled. +// draft: vector to store the draft tokens, initially empty. +llama_tokens common_ngram_simple_draft( + common_ngram_simple_state & state, + const llama_tokens & tokens, llama_token sampled); + + +// n-gram map +// + +// maximum number of m-gram values stored for each key n-gram. +#define COMMON_NGRAM_MAX_VALUES 4 + +// statistics of a m-gram after a known n-gram +struct common_ngram_map_value { + size_t value_idx = 0; // index of value m-gram in token-history (0 if unused) + uint16_t value_num = 0; // number of occurences of this value m-gram after the key n-gram (0 in an unused values-slot) + int16_t n_accepted = -1; // number of accepted tokens at last draft (-1 if unused) +}; + +// statistics of a n-gram +struct common_ngram_map_key { + size_t key_idx; // index of key n-gram in token-history + size_t stat_idx; // index of last token of stastistics computation (key_num, values) + + uint16_t key_num; // number of occurences of this key n-gram in token-history + common_ngram_map_value values[COMMON_NGRAM_MAX_VALUES]; // some known values after the key +}; + +// map from n-grams to following m-grams in token-history +struct common_ngram_map { + uint16_t size_key; // size of key n-grams + uint16_t size_value; // size of value m-grams + + bool key_only; // true if only key n-grams are used, no values. + + // first draft: vector only, no map. + std::vector keys; // key n-grams which occur several times in token-history + uint16_t check_rate; // check for speculative decoding without draft model for each check_rate token + uint16_t min_hits; // minimum number of key hits to consider a draft + + common_ngram_map(uint16_t sz_key, uint16_t sz_value, bool only_keys, + uint16_t check_rate, uint16_t min_hits) + : size_key(sz_key), size_value(sz_value), key_only(only_keys), + check_rate(check_rate), min_hits(min_hits) {} + + bool last_draft_created = false; // true if a draft was created at last call. + size_t last_draft_key_idx = 0; // index of last key used for draft generation. + uint16_t last_draft_value_idx = 0; // index of last value used for draft generation. + + size_t idx_last_check = 0; // index of last check in context history +}; + + +// Searches for the n-gram in the history and checks whether a draft sequence should be generated. +// map: the ngram map to search in. +// inp: the tokens generated so far. +// sampled: the token that was just sampled. +// draft: vector to store the draft tokens, initially empty. +void common_ngram_map_draft( + common_ngram_map & map, + const llama_tokens & inp, llama_token sampled, + llama_tokens & draft); + +// Update the statistics of a value after a draft was processed. +void common_ngram_map_accept(common_ngram_map & map, uint16_t n_accepted); diff --git a/common/speculative.cpp b/common/speculative.cpp index 3e83b0964c..3f314b5d57 100644 --- a/common/speculative.cpp +++ b/common/speculative.cpp @@ -1,99 +1,54 @@ #include "speculative.h" +#include "common.h" #include "ggml.h" #include "llama.h" #include "log.h" -#include "common.h" +#include "ngram-cache.h" +#include "ngram-map.h" #include "sampling.h" -#include #include +#include +#include #include #define SPEC_VOCAB_MAX_SIZE_DIFFERENCE 128 #define SPEC_VOCAB_CHECK_START_TOKEN_ID 5 -struct common_speculative { - struct llama_context * ctx_tgt; // only used for retokenizing from ctx_dft - struct llama_context * ctx_dft; - struct common_sampler * smpl; - - llama_batch batch; - llama_tokens prompt_dft; - bool vocab_dft_compatible = true; // whether retokenization is needed - std::map tgt_dft_replacements = {}; +const std::vector common_speculative_types = { + COMMON_SPECULATIVE_TYPE_NONE, + COMMON_SPECULATIVE_TYPE_DRAFT, + COMMON_SPECULATIVE_TYPE_EAGLE3, + COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE, + COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K, + COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V, + COMMON_SPECULATIVE_TYPE_NGRAM_CACHE }; -struct common_speculative * common_speculative_init( - struct llama_context * ctx_tgt, - struct llama_context * ctx_dft) { - auto * result = new common_speculative { - /* .ctx_tgt = */ ctx_tgt, - /* .ctx_dft = */ ctx_dft, - /* .smpl = */ nullptr, - /* .batch = */ llama_batch_init(llama_n_batch(ctx_dft), 0, 1), - /* .prompt_dft = */ {}, - /* .vocab_dft_compatible = */ false, - }; +const std::map common_speculative_type_from_name_map = { + {"none", COMMON_SPECULATIVE_TYPE_NONE}, + {"draft", COMMON_SPECULATIVE_TYPE_DRAFT}, + {"eagle3", COMMON_SPECULATIVE_TYPE_EAGLE3}, + {"ngram_simple", COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE}, + {"ngram_map_k", COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K}, + {"ngram_map_k4v", COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V}, + {"ngram_cache", COMMON_SPECULATIVE_TYPE_NGRAM_CACHE} +}; - // TODO: optimize or pass from outside? -#if 0 - { - common_params_sampling params; - params.no_perf = false; +struct common_speculative_config { + common_speculative_type type; + common_params_speculative params; - params.top_k = 40; - params.top_p = 0.9; + common_speculative_config(common_speculative_type t, + const common_params_speculative & p = common_params_speculative{}) : type(t), params(p) {} +}; - params.samplers = { - COMMON_SAMPLER_TYPE_TOP_K, - COMMON_SAMPLER_TYPE_TOP_P, - COMMON_SAMPLER_TYPE_INFILL, - }; - - result->smpl = common_sampler_init(llama_get_model(ctx_dft), params); - } -#else - { - common_params_sampling params; - params.no_perf = false; - - params.top_k = 10; - - params.samplers = { - COMMON_SAMPLER_TYPE_TOP_K, - }; - - result->smpl = common_sampler_init(llama_get_model(ctx_dft), params); - } -#endif - - result->vocab_dft_compatible = common_speculative_are_compatible(ctx_tgt, ctx_dft); - LOG_DBG("vocab_dft_compatible = %d\n", result->vocab_dft_compatible); - - return result; -} - -void common_speculative_free(struct common_speculative * spec) { - if (spec == nullptr) { - return; - } - - common_sampler_free(spec->smpl); - - llama_batch_free(spec->batch); - - delete spec; -} - -bool common_speculative_are_compatible( - const struct llama_context * ctx_tgt, - const struct llama_context * ctx_dft) { - const struct llama_model * model_tgt = llama_get_model(ctx_tgt); - const struct llama_model * model_dft = llama_get_model(ctx_dft); - - const struct llama_vocab * vocab_tgt = llama_model_get_vocab(model_tgt); - const struct llama_vocab * vocab_dft = llama_model_get_vocab(model_dft); +static bool common_speculative_are_compatible( + const llama_model * model_tgt, + const llama_model * model_dft) { + const llama_vocab * vocab_tgt = llama_model_get_vocab(model_tgt); + const llama_vocab * vocab_dft = llama_model_get_vocab(model_dft); const bool vocab_type_tgt = llama_vocab_type(vocab_tgt); LOG_DBG("%s: vocab_type tgt: %d\n", __func__, vocab_type_tgt); @@ -134,11 +89,12 @@ bool common_speculative_are_compatible( for (int i = SPEC_VOCAB_CHECK_START_TOKEN_ID; i < std::min(n_vocab_tgt, n_vocab_dft); ++i) { const char * token_text_tgt = llama_vocab_get_text(vocab_tgt, i); const char * token_text_dft = llama_vocab_get_text(vocab_dft, i); + if (std::strcmp(token_text_tgt, token_text_dft) != 0) { LOG_DBG("%s: draft model vocab must match target model to use speculation but ", __func__); LOG_DBG("token %d content differs - target '%s', draft '%s'\n", i, - common_token_to_piece(ctx_tgt, i).c_str(), - common_token_to_piece(ctx_dft, i).c_str()); + common_token_to_piece(vocab_tgt, i).c_str(), + common_token_to_piece(vocab_dft, i).c_str()); return false; } } @@ -147,215 +103,779 @@ bool common_speculative_are_compatible( return true; } -void common_speculative_add_replacement_tgt_dft( - struct common_speculative * spec, - const char *source, const char *dest) { - spec->tgt_dft_replacements[source] = dest; -} +// state of an implementation of speculative decoding +// +// each implementation has a unique type and a state that is implementation-specific +// in a subclass of common_speculative_state +struct common_speculative_state { + const enum common_speculative_type type; -static std::string replace_to_dft( - struct common_speculative * spec, - const std::string& input) { - std::string result = input; - for (const auto & pair : spec->tgt_dft_replacements) { - size_t pos = result.find(pair.first); - while (pos != std::string::npos) { - result.replace(pos, pair.first.length(), pair.second); - pos = result.find(pair.first, pos + pair.second.length()); - } - } - return result; -} + size_t drafts_call_count = 0; // number of times this implementation was called. + size_t drafts_generated_count = 0; // number of times a draft or part was generated by this implementation. + size_t drafts_accepted_count = 0; // number of times a draft or part was accepted by the target model. + size_t drafts_generated_tokens = 0; // number of tokens generated by this implementation. + size_t drafts_accepted_tokens = 0; // number of tokens accepted by the target model. -static std::string replace_to_tgt( - struct common_speculative * spec, - const std::string& input) { - std::string result = input; - for (const auto& pair : spec->tgt_dft_replacements) { - size_t pos = result.find(pair.second); - while (pos != std::string::npos) { - result.replace(pos, pair.second.length(), pair.first); - pos = result.find(pair.second, pos + pair.first.length()); - } - } - return result; -} + // TODO: track performance of most recent calls + const bool gen_perf = true; // whether to generate performance stats. + int64_t gen_duration_us = 0; // total time spent in this implementation in microseconds. -llama_tokens common_speculative_gen_draft( - struct common_speculative * spec, - struct common_speculative_params params, - const llama_tokens & prompt_tgt_main_model, // specified in target model vocab - llama_token id_last) { - auto & batch = spec->batch; - auto & ctx_tgt = spec->ctx_tgt; - auto & ctx_dft = spec->ctx_dft; - auto & smpl = spec->smpl; - auto & prompt_dft = spec->prompt_dft; + common_speculative_state(enum common_speculative_type type) : type(type) {} - auto * mem_dft = llama_get_memory(ctx_dft); + virtual ~common_speculative_state() = default; - int reuse_i = 0; - int reuse_n = 0; + virtual void begin(const llama_tokens & prompt) = 0; - const int n_ctx = llama_n_ctx(ctx_dft) - params.n_draft; + virtual void draft( + const common_params_speculative & params, + const llama_tokens & prompt_tgt, + llama_token id_last, + llama_tokens & result) = 0; - llama_tokens prompt_tgt_draft_model; - if (!spec->vocab_dft_compatible) { - std::string text; - text = common_detokenize(ctx_tgt, prompt_tgt_main_model, true); - text = replace_to_dft(spec, text); - LOG_DBG("%s: main->draft detokenized string: '%s'\n", __func__, text.c_str()); - prompt_tgt_draft_model = common_tokenize(ctx_dft, text, false, true); + virtual void accept(uint16_t n_accepted) = 0; +}; - // convert id_last to draft vocab. llama_detokenize is called directly to avoid an allocation - const auto * model_tgt = llama_get_model(ctx_tgt); - const auto * vocab_tgt = llama_model_get_vocab(model_tgt); +struct common_speculative_state_draft : public common_speculative_state { + llama_context * ctx_tgt; // only used for retokenizing from ctx_dft + llama_context * ctx_dft; - int32_t n_chars = llama_detokenize(vocab_tgt, &id_last, 1, nullptr, 0, false, false); - GGML_ASSERT(n_chars < 0 && "failed to detokenize id_last"); - text.resize(-n_chars); - llama_detokenize(vocab_tgt, &id_last, 1, text.data(), text.size(), false, false); - text = replace_to_dft(spec, text); + common_sampler * smpl; - LOG_DBG("main->draft detokenized id_last(%d): '%s'\n", id_last, text.c_str()); - id_last = common_tokenize(ctx_dft, text, false, true)[0]; - } - // prompt_tgt's tokens will always be compatible with ctx_dft - const llama_tokens &prompt_tgt = - spec->vocab_dft_compatible ? prompt_tgt_main_model : prompt_tgt_draft_model; + llama_batch batch; + llama_tokens prompt_dft; - const int i_start = std::max(0, (int) prompt_tgt.size() - n_ctx); + bool vocab_cmpt = true; // whether retokenization is needed + std::unordered_map vocab_map; - // reuse as much as possible from the old draft context - // ideally, the draft context should be as big as the target context and we will always reuse the entire prompt - for (int i = 0; i < (int) prompt_dft.size(); ++i) { - int cur = 0; - while (i_start + cur < (int) prompt_tgt.size() && - i + cur < (int) prompt_dft.size() && - prompt_tgt[i_start + cur] == prompt_dft[i + cur]) { - cur++; + common_speculative_state_draft( + enum common_speculative_type type, + llama_context * ctx_tgt, + llama_context * ctx_dft, + const std::vector> & replacements) + : common_speculative_state(type) + , ctx_tgt(ctx_tgt) + , ctx_dft(ctx_dft) + { + batch = llama_batch_init(llama_n_batch(ctx_dft), 0, 1); + smpl = nullptr; + + // TODO: optimize or pass from outside? + // { + // common_params_sampling params; + // params.no_perf = false; + // + // params.top_k = 40; + // params.top_p = 0.9; + // + // params.samplers = { + // COMMON_SAMPLER_TYPE_TOP_K, + // COMMON_SAMPLER_TYPE_TOP_P, + // COMMON_SAMPLER_TYPE_INFILL, + // }; + // + // result->smpl = common_sampler_init(llama_get_model(ctx_dft), params); + // } + { + common_params_sampling params; + params.no_perf = false; + params.top_k = 10; + params.samplers = { + COMMON_SAMPLER_TYPE_TOP_K, + }; + + smpl = common_sampler_init(llama_get_model(ctx_dft), params); } - if ((cur >= params.n_reuse || n_ctx >= (int) prompt_tgt.size()) && cur > reuse_n) { - reuse_i = i; - reuse_n = cur; + vocab_cmpt = common_speculative_are_compatible(llama_get_model(ctx_tgt), llama_get_model(ctx_dft)); + LOG_DBG("vocab_cmpt = %d\n", vocab_cmpt); + + if (!vocab_cmpt) { + LOG_WRN("the target and draft vocabs are not compatible - tokens will be translated between the two\n"); + + for (const auto & pair : replacements) { + vocab_map[pair.first] = pair.second; + } } } - LOG_DBG("%s: reuse_i = %d, reuse_n = %d, prompt = %d\n", __func__, reuse_i, reuse_n, (int) prompt_dft.size()); + ~common_speculative_state_draft() override { + llama_perf_context_print(ctx_dft); - llama_tokens result; - result.reserve(params.n_draft); + llama_free(ctx_dft); - if (reuse_n == 0) { - llama_memory_clear(mem_dft, false); - prompt_dft.clear(); - } else { - // this happens when a previous draft has been discarded (for example, due to being too small), but the - // target model agreed with it. in this case, we simply pass back the previous results to save compute - if (reuse_i + reuse_n < (int) prompt_dft.size() && prompt_dft[reuse_i + reuse_n] == id_last) { - for (int i = reuse_i + reuse_n + 1; i < (int) prompt_dft.size(); ++i) { - result.push_back(prompt_dft[i]); + common_sampler_free(smpl); - if (params.n_draft <= (int) result.size()) { - break; - } + llama_batch_free(batch); + } + + void begin(const llama_tokens & prompt) override { + GGML_UNUSED(prompt); + } + + void draft( + const common_params_speculative & params, + const llama_tokens & prompt_tgt, + llama_token id_last, + llama_tokens & result) override { + auto * spec = this; + + auto & batch = spec->batch; + auto & ctx_tgt = spec->ctx_tgt; + auto & ctx_dft = spec->ctx_dft; + auto & smpl = spec->smpl; + auto & prompt_dft = spec->prompt_dft; + + auto * mem_dft = llama_get_memory(ctx_dft); + + int reuse_i = 0; + int reuse_n = 0; + + const int n_ctx = llama_n_ctx(ctx_dft) - params.n_max; + + llama_tokens prompt_cnv; + if (!spec->vocab_cmpt) { + std::string text; + + text = common_detokenize(ctx_tgt, prompt_tgt, true); + text = replace_to_dft(text); + + LOG_DBG("%s: main->draft detokenized string: '%s'\n", __func__, text.c_str()); + + prompt_cnv = common_tokenize(ctx_dft, text, false, true); + + // convert id_last to draft vocab. llama_detokenize is called directly to avoid an allocation + const auto * model_tgt = llama_get_model(ctx_tgt); + const auto * vocab_tgt = llama_model_get_vocab(model_tgt); + + int32_t n_chars = llama_detokenize(vocab_tgt, &id_last, 1, nullptr, 0, false, false); + GGML_ASSERT(n_chars < 0 && "failed to detokenize id_last"); + + text.resize(-n_chars); + llama_detokenize(vocab_tgt, &id_last, 1, text.data(), text.size(), false, false); + text = replace_to_dft(text); + + LOG_DBG("main->draft detokenized id_last(%d): '%s'\n", id_last, text.c_str()); + id_last = common_tokenize(ctx_dft, text, false, true)[0]; + } + + const llama_tokens & prompt_cur = spec->vocab_cmpt ? prompt_tgt : prompt_cnv; + + const int i_start = std::max(0, (int) prompt_cur.size() - n_ctx); + + // reuse as much as possible from the old draft context + // ideally, the draft context should be as big as the target context and we will always reuse the entire prompt + for (int i = 0; i < (int) prompt_dft.size(); ++i) { + int cur = 0; + while (i_start + cur < (int) prompt_cur.size() && + i + cur < (int) prompt_dft.size() && + prompt_cur[i_start + cur] == prompt_dft[i + cur]) { + cur++; } - return result; + if ((cur >= 256 || n_ctx >= (int) prompt_cur.size()) && cur > reuse_n) { + reuse_i = i; + reuse_n = cur; + } } - if (reuse_i > 0) { - llama_memory_seq_rm (mem_dft, 0, 0, reuse_i); - llama_memory_seq_add(mem_dft, 0, reuse_i, -1, -reuse_i); + LOG_DBG("%s: reuse_i = %d, reuse_n = %d, prompt = %d\n", __func__, reuse_i, reuse_n, (int) prompt_dft.size()); - prompt_dft.erase(prompt_dft.begin(), prompt_dft.begin() + reuse_i); + result.clear(); + result.reserve(params.n_max); + + if (reuse_n == 0) { + llama_memory_clear(mem_dft, false); + prompt_dft.clear(); + } else { + // this happens when a previous draft has been discarded (for example, due to being too small), but the + // target model agreed with it. in this case, we simply pass back the previous results to save compute + if (reuse_i + reuse_n < (int) prompt_dft.size() && prompt_dft[reuse_i + reuse_n] == id_last) { + for (int i = reuse_i + reuse_n + 1; i < (int) prompt_dft.size(); ++i) { + result.push_back(prompt_dft[i]); + + if (params.n_max <= (int) result.size()) { + break; + } + } + + return; + } + + if (reuse_i > 0) { + llama_memory_seq_rm (mem_dft, 0, 0, reuse_i); + llama_memory_seq_add(mem_dft, 0, reuse_i, -1, -reuse_i); + + prompt_dft.erase(prompt_dft.begin(), prompt_dft.begin() + reuse_i); + } + + if (reuse_n < (int) prompt_dft.size()) { + llama_memory_seq_rm (mem_dft, 0, reuse_n, -1); + prompt_dft.erase(prompt_dft.begin() + reuse_n, prompt_dft.end()); + } } - if (reuse_n < (int) prompt_dft.size()) { - llama_memory_seq_rm (mem_dft, 0, reuse_n, -1); - prompt_dft.erase(prompt_dft.begin() + reuse_n, prompt_dft.end()); - } - } - - // prepare a batch to evaluate any new tokens in the prompt - common_batch_clear(batch); - - for (size_t i = i_start + reuse_n; i < prompt_tgt.size(); ++i) { - //LOG_DBG("i = %d, i_start = %d, reuse_n = %d, i - i_start = %d, id = %6d\n", i, i_start, reuse_n, i - i_start, prompt_tgt[i]); - common_batch_add(batch, prompt_tgt[i], i - i_start, { 0 }, false); - - prompt_dft.push_back(prompt_tgt[i]); - } - - // we should rarely end-up here during normal decoding - if (batch.n_tokens > 0) { - //LOG_DBG("%s: draft prompt batch: %s\n", __func__, string_from(ctx, batch).c_str()); - - llama_decode(ctx_dft, batch); - } - - const llama_pos n_past = prompt_dft.size(); - - LOG_DBG("%s: n_past = %d\n", __func__, n_past); - - common_batch_clear(batch); - common_batch_add (batch, id_last, n_past, { 0 }, true); - - prompt_dft.push_back(id_last); - - LOG_DBG("%s: draft prompt: %s\n", __func__, string_from(ctx_dft, prompt_dft).c_str()); - - llama_decode(ctx_dft, batch); - - common_sampler_reset(smpl); - - // sample n_draft tokens from the draft model - for (int i = 0; i < params.n_draft; ++i) { + // prepare a batch to evaluate any new tokens in the prompt common_batch_clear(batch); - common_sampler_sample(smpl, ctx_dft, 0, true); + for (size_t i = i_start + reuse_n; i < prompt_cur.size(); ++i) { + //LOG_DBG("i = %d, i_start = %d, reuse_n = %d, i - i_start = %d, id = %6d\n", i, i_start, reuse_n, i - i_start, prompt_cur[i]); + common_batch_add(batch, prompt_cur[i], i - i_start, { 0 }, false); - const auto * cur_p = common_sampler_get_candidates(smpl, true); - - for (int k = 0; k < std::min(3, (int) cur_p->size); ++k) { - LOG_DBG(" - draft candidate %3d, pos %3d: %6d (%8.3f) '%s'\n", - k, i, cur_p->data[k].id, cur_p->data[k].p, common_token_to_piece(ctx_dft, cur_p->data[k].id).c_str()); + prompt_dft.push_back(prompt_cur[i]); } - // add drafted token for each sequence - const llama_token id = cur_p->data[0].id; + // we should rarely end-up here during normal decoding + if (batch.n_tokens > 0) { + //LOG_DBG("%s: draft prompt batch: %s\n", __func__, string_from(ctx, batch).c_str()); - common_sampler_accept(smpl, id, true); - - result.push_back(id); - - if (params.n_draft <= (int) result.size()) { - break; + llama_decode(ctx_dft, batch); } - // only collect very high-confidence draft tokens - if (cur_p->data[0].p < params.p_min) { - break; - } + const llama_pos n_past = prompt_dft.size(); - common_batch_add(batch, id, n_past + i + 1, { 0 }, true); + LOG_DBG("%s: n_past = %d\n", __func__, n_past); + + common_batch_clear(batch); + common_batch_add (batch, id_last, n_past, { 0 }, true); + + prompt_dft.push_back(id_last); + + LOG_DBG("%s: draft prompt: %s\n", __func__, string_from(ctx_dft, prompt_dft).c_str()); - // evaluate the drafted tokens on the draft model llama_decode(ctx_dft, batch); - prompt_dft.push_back(id); + common_sampler_reset(smpl); + + // sample n_draft tokens from the draft model + for (int i = 0; i < params.n_max; ++i) { + common_batch_clear(batch); + + common_sampler_sample(smpl, ctx_dft, 0, true); + + const auto * cur_p = common_sampler_get_candidates(smpl, true); + + for (int k = 0; k < std::min(3, (int) cur_p->size); ++k) { + LOG_DBG(" - draft candidate %3d, pos %3d: %6d (%8.3f) '%s'\n", + k, i, cur_p->data[k].id, cur_p->data[k].p, common_token_to_piece(ctx_dft, cur_p->data[k].id).c_str()); + } + + // add drafted token for each sequence + const llama_token id = cur_p->data[0].id; + + common_sampler_accept(smpl, id, true); + + result.push_back(id); + + if (params.n_max <= (int) result.size()) { + break; + } + + // only collect very high-confidence draft tokens + if (cur_p->data[0].p < params.p_min) { + break; + } + + common_batch_add(batch, id, n_past + i + 1, { 0 }, true); + + // evaluate the drafted tokens on the draft model + llama_decode(ctx_dft, batch); + + prompt_dft.push_back(id); + } + + if (!spec->vocab_cmpt) { + std::string detokenized = common_detokenize(ctx_dft, result, true); + detokenized = replace_to_tgt(detokenized); + LOG_DBG("draft->main detokenized string: '%s'\n", detokenized.c_str()); + result = common_tokenize(ctx_tgt, detokenized, false, true); + if (result.size() > (size_t)params.n_max) { + result.resize(params.n_max); + } + } } - if (!spec->vocab_dft_compatible) { - std::string detokenized = common_detokenize(ctx_dft, result, true); - detokenized = replace_to_tgt(spec, detokenized); - LOG_DBG("draft->main detokenized string: '%s'\n", detokenized.c_str()); - result = common_tokenize(ctx_tgt, detokenized, false, true); - if (result.size() > (size_t)params.n_draft) { - result.resize(params.n_draft); + void accept(uint16_t n_accepted) override { + // noop + GGML_UNUSED(n_accepted); + } + + std::string replace_to_dft(const std::string & input) const { + std::string result = input; + + for (const auto & pair : this->vocab_map) { + size_t pos = result.find(pair.first); + while (pos != std::string::npos) { + result.replace(pos, pair.first.length(), pair.second); + pos = result.find(pair.first, pos + pair.second.length()); + } } + + return result; + } + + std::string replace_to_tgt(const std::string & input) const { + std::string result = input; + + for (const auto & pair : this->vocab_map) { + size_t pos = result.find(pair.second); + while (pos != std::string::npos) { + result.replace(pos, pair.second.length(), pair.first); + pos = result.find(pair.second, pos + pair.first.length()); + } + } + + return result; + } +}; + +struct common_speculative_state_eagle3 : public common_speculative_state { + common_speculative_state_eagle3(enum common_speculative_type type) : common_speculative_state(type) {} + + void begin(const llama_tokens & prompt) override { + GGML_UNUSED(prompt); + } + + void draft( + const common_params_speculative & params, + const llama_tokens & prompt_tgt, + llama_token id_last, + llama_tokens & draft_tokens) override { + // TODO: implement + GGML_UNUSED(params); + GGML_UNUSED(prompt_tgt); + GGML_UNUSED(id_last); + GGML_UNUSED(draft_tokens); + } + + void accept(uint16_t n_accepted) override { + // noop + GGML_UNUSED(n_accepted); + } +}; + +// state of self-speculation (simple implementation, not ngram-map) +struct common_speculative_state_ngram_simple : public common_speculative_state { + common_ngram_simple_state state; + + common_speculative_state_ngram_simple( + enum common_speculative_type type, + common_ngram_simple_state state) + : common_speculative_state(type), state(state) {} + + void begin(const llama_tokens & prompt) override { + GGML_UNUSED(prompt); + } + + void draft( + const common_params_speculative & params, + const llama_tokens & prompt_tgt, + llama_token id_last, + llama_tokens & result) override { + result = common_ngram_simple_draft(state, prompt_tgt, id_last); + GGML_UNUSED(params); + } + + void accept(uint16_t n_accepted) override { + // noop + GGML_UNUSED(n_accepted); + } +}; + +struct common_speculative_state_ngram_map_k : public common_speculative_state { + // draft ngram map for speculative decoding without draft model + common_ngram_map map; + + common_speculative_state_ngram_map_k( + enum common_speculative_type type, + common_ngram_map map) + : common_speculative_state(type), map(std::move(map)) {} + + void begin(const llama_tokens & prompt) override { + GGML_UNUSED(prompt); + } + + void draft( + const common_params_speculative & params, + const llama_tokens & prompt_tgt, + llama_token id_last, + llama_tokens & result) override { + common_ngram_map_draft(map, prompt_tgt, id_last, result); + GGML_UNUSED(params); + } + + void accept(uint16_t n_accepted) override { + common_ngram_map_accept(map, n_accepted); + } +}; + +struct common_speculative_state_ngram_cache : public common_speculative_state { + uint16_t n_draft; + bool save_dynamic; + bool save_static; + + common_ngram_cache ngram_cache_context; + common_ngram_cache ngram_cache_dynamic; + common_ngram_cache ngram_cache_static; + + size_t cache_size = 0; // number of tokens in n-gram cache + + common_speculative_state_ngram_cache( + const enum common_speculative_type type, + const std::string & path_static, + const std::string & path_dynamic, + uint16_t n_draft, + bool save_dynamic, + bool save_static) + : common_speculative_state(type) + , n_draft(n_draft) + , save_dynamic(save_dynamic) + , save_static(save_static) + { + if (!path_static.empty()) { + try { + ngram_cache_static = common_ngram_cache_load(path_static); + } catch (...) { + LOG_ERR("failed to open static lookup cache: %s", path_static.c_str()); + GGML_ABORT("Couldn't read static lookup cache"); + } + } + + if (!path_dynamic.empty()) { + try { + ngram_cache_dynamic = common_ngram_cache_load(path_dynamic); + } catch (...) { + LOG_ERR("failed to open dynamic lookup cache: %s", path_dynamic.c_str()); + GGML_ABORT("Couldn't read dynamic lookup cache"); + } + } + } + + void begin(const llama_tokens & prompt) override { + GGML_UNUSED(prompt); + } + + void draft( + const common_params_speculative & params, + const llama_tokens & prompt_tgt, + llama_token id_last, + llama_tokens & result) override { + GGML_UNUSED(params); + + if (cache_size < prompt_tgt.size() + 1) { + llama_tokens tokens_new; + tokens_new.reserve(prompt_tgt.size() + 1 - cache_size); + for (size_t j = cache_size; j < prompt_tgt.size(); ++j) { + tokens_new.push_back(prompt_tgt[j]); + } + tokens_new.push_back(id_last); // add the last token + + // Update context ngram cache with new prompt_tgt: + common_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, + tokens_new, tokens_new.size(), false); + cache_size = prompt_tgt.size() + 1; + } + + llama_tokens inp; + inp.reserve(prompt_tgt.size() + 1); + for (size_t j = 0; j < prompt_tgt.size(); ++j) { + inp.push_back(prompt_tgt[j]); + } + inp.push_back(id_last); + + result.push_back(id_last); + + common_ngram_cache_draft(inp, result, n_draft, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, + ngram_cache_context, + ngram_cache_dynamic, + ngram_cache_static); + + if (result.size() > 0) { + // delete first token in result (which is the id_last token) + result.erase(result.begin()); + } + } + + void accept(uint16_t n_accepted) override { + // TODO: noop + GGML_UNUSED(n_accepted); + } +}; + +struct common_speculative { + std::vector> impls; // list of implementations to use and their states + common_speculative_state * curr_impl = nullptr; // current implementation in use (for stats) +}; + +static common_ngram_map get_common_ngram_map(const common_speculative_config & config) { + uint16_t size_key = config.params.ngram_size_n; + uint16_t size_value = config.params.ngram_size_m; + bool key_only = (config.type == COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K); + uint16_t check_rate = config.params.ngram_check_rate; + uint16_t min_hits = config.params.ngram_min_hits; + + return common_ngram_map(size_key, size_value, key_only, check_rate, min_hits); +} + +static common_speculative_state_ngram_cache create_state_ngram_cache( + const std::string & path_static, const std::string & path_dynamic, + const common_speculative_config & config) { + uint16_t n_draft = 8; // TODO get from config? + + // TODO bool param in common/common.h to set save_static/save_dynamic? + bool save_static = false; + bool save_dynamic = false; + + common_speculative_state_ngram_cache state(config.type, path_static, path_dynamic, n_draft, save_static, save_dynamic); + + return state; +} + +std::string common_speculative_type_name_str() { + std::string result; + for (size_t i = 0; i < common_speculative_types.size(); i++) { + if (i > 0) { + result += ", "; + } + result += common_speculative_type_to_str(common_speculative_types[i]); } return result; } + +std::string common_speculative_type_to_str(enum common_speculative_type type) { + switch (type) { + case COMMON_SPECULATIVE_TYPE_NONE: return "none"; + case COMMON_SPECULATIVE_TYPE_DRAFT: return "draft"; + case COMMON_SPECULATIVE_TYPE_EAGLE3: return "eagle3"; + case COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE: return "ngram_simple"; + case COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K: return "ngram_map_k"; + case COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V: return "ngram_map_k4v"; + case COMMON_SPECULATIVE_TYPE_NGRAM_CACHE: return "ngram_cache"; + default: return "unknown"; + } +} + +enum common_speculative_type common_speculative_type_from_name(const std::string & name) { + const auto it = common_speculative_type_from_name_map.find(name); + if (it == common_speculative_type_from_name_map.end()) { + return COMMON_SPECULATIVE_TYPE_COUNT; + } + return it->second; +} + +// initialization of the speculative decoding system +// +common_speculative * common_speculative_init( + const common_params_speculative & params, + llama_context * ctx_tgt) { + llama_context * ctx_dft = nullptr; + if (params.model_dft) { + ctx_dft = llama_init_from_model(params.model_dft, params.cparams_dft); + if (ctx_dft == nullptr) { + LOG_ERR("%s", "failed to create draft context\n"); + return nullptr; + } + } + + // Compute the implementations to use based on the config and their order of preference + std::vector configs = {}; // list of speculative configs to try + { + bool has_draft = !params.mparams_dft.path.empty(); + bool has_draft_eagle3 = false; // TODO PR-18039: if params.speculative.eagle3 + + bool has_ngram_cache = (params.type == COMMON_SPECULATIVE_TYPE_NGRAM_CACHE); + bool has_ngram_simple = (params.type == COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE); + bool has_ngram_map_k = (params.type == COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K); + bool has_ngram_map_k4v = (params.type == COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V); + + // In a more complex implementation we could use the same implementation but with different parameters. + // This was initially used in PR-18471 but removed to simplify the code. + if (has_ngram_simple) { + // This implementation can guess a lot of tokens without any draft model. + configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE, params)); + } + if (has_ngram_map_k) { + configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K, params)); + } + if (has_ngram_map_k4v) { + // This implementation can guess tokens with high acceptance rate but is more expensive. + configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V, params)); + } + if (has_ngram_cache) { + configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_NGRAM_CACHE, params)); + } + if (has_draft) { + configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_DRAFT, params)); + } + if (has_draft_eagle3) { + configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_EAGLE3, params)); + } + } + + std::vector> impls = {}; + + for (const common_speculative_config & config : configs) { + LOG_DBG("%s: adding implementation %s\n", __func__, common_speculative_type_to_str(config.type).c_str()); + switch (config.type) { + case COMMON_SPECULATIVE_TYPE_NONE: + break; + case COMMON_SPECULATIVE_TYPE_DRAFT: { + impls.push_back(std::make_unique(config.type, + /* .ctx_tgt = */ ctx_tgt, + /* .ctx_dft = */ ctx_dft, + /* .replacements = */ params.replacements + )); + break; + } + case COMMON_SPECULATIVE_TYPE_EAGLE3: { + impls.push_back(std::make_unique(config.type)); + break; + } + case COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE: { + common_ngram_map ngram_map = get_common_ngram_map(config); + + uint16_t ngram_size_key = ngram_map.size_key; + uint16_t mgram_size_value = ngram_map.size_value; + uint16_t check_rate = ngram_map.check_rate; + + auto config_simple = common_ngram_simple_config{ + /* .size_ngram = */ ngram_size_key, + /* .size_mgram = */ mgram_size_value, + /* .check_rate = */ check_rate + }; + auto state = std::make_unique( + /* .type = */ config.type, + /* .state = */ common_ngram_simple_state(config_simple) + ); + impls.push_back(std::move(state)); + break; + } + case COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K: + case COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V: { + impls.push_back(std::make_unique( + (config.type), + get_common_ngram_map(config) + )); + break; + } + case COMMON_SPECULATIVE_TYPE_NGRAM_CACHE: { + auto state = create_state_ngram_cache( + params.lookup_cache_static, params.lookup_cache_dynamic, config); + impls.push_back(std::make_unique(state)); + break; + } + default: + break; + } + } + + if (impls.empty()) { + LOG_WRN("%s", "no implementations specified for speculative decoding\n"); + return nullptr; + } + + auto * result = new common_speculative { + /* .impls = */ std::move(impls) + }; + + return result; +} + +void common_speculative_free(common_speculative * spec) { + if (spec == nullptr) { + return; + } + + delete spec; +} + +void common_speculative_begin(common_speculative * spec, const llama_tokens & prompt) { + if (spec == nullptr) { + return; + } + + for (auto & impl : spec->impls) { + impl->begin(prompt); + } +} + +llama_tokens common_speculative_draft( + common_speculative * spec, + const common_params_speculative & params, + const llama_tokens & prompt_tgt, // specified in target model vocab + llama_token id_last) { + llama_tokens result; + + spec->curr_impl = nullptr; // reset current implementation + + for (auto & impl : spec->impls) { + { + const int64_t t_start_us = impl->gen_perf ? ggml_time_us() : 0; + + impl->draft(params, prompt_tgt, id_last, result); + + const int64_t t_now_us = impl->gen_perf ? ggml_time_us() : 0; + + impl->drafts_call_count++; + impl->gen_duration_us += t_now_us - t_start_us; // accumulate duration for this implementation + } + + if (!result.empty()) { + LOG_DBG("%s: called impl %s, hist size = %zu, call_count = %zu, gen = %zu\n", __func__, + common_speculative_type_to_str(impl.get()->type).c_str(), + prompt_tgt.size(), + impl.get()->drafts_call_count, result.size()); + + spec->curr_impl = impl.get(); // set current implementation for stats + impl->drafts_generated_count++; + impl->drafts_generated_tokens += result.size(); + + break; // We have a draft, so break out of the loop and return it. + } + } + + return result; +} + +void common_speculative_accept(common_speculative * spec, uint16_t n_accepted) { + if (n_accepted == 0) { + return; + } + + common_speculative_state * impl = spec->curr_impl; + + GGML_ASSERT(impl); + + if (n_accepted > 0) { + impl->drafts_accepted_count++; + impl->drafts_accepted_tokens += n_accepted; + } + + impl->accept(n_accepted); +} + +void common_speculative_print_stats(const common_speculative * spec) { + if (spec == nullptr) { + return; + } + + for (const auto & impl : spec->impls) { + std::string str_perf; + if (impl->gen_perf) { + std::ostringstream oss; + oss << std::fixed << std::setprecision(3) << impl->gen_duration_us / 1000.0; + str_perf = ", dur = " + oss.str() + " ms"; + } else { + str_perf = ""; + } + + LOG_INF("statistics %s: #calls = %zu, #gen drafts = %zu, #acc drafts = %zu, #gen tokens = %zu, #acc tokens = %zu%s\n", + common_speculative_type_to_str(impl->type).c_str(), + impl->drafts_call_count, + impl->drafts_generated_count, + impl->drafts_accepted_count, + impl->drafts_generated_tokens, + impl->drafts_accepted_tokens, + str_perf.c_str()); + } +} diff --git a/common/speculative.h b/common/speculative.h index e69d7aaa1e..9e1888e4be 100644 --- a/common/speculative.h +++ b/common/speculative.h @@ -5,31 +5,33 @@ struct common_speculative; -struct common_speculative_params { - int n_draft = 16; // max drafted tokens - int n_reuse = 256; +// comma separated list of all types +std::string common_speculative_type_name_str(); - float p_min = 0.75f; // min probability required to accept a token in the draft -}; +// convert string to type +enum common_speculative_type common_speculative_type_from_name(const std::string & name); -struct common_speculative * common_speculative_init( - struct llama_context * ctx_tgt, - struct llama_context * ctx_dft -); +// convert type to string +std::string common_speculative_type_to_str(enum common_speculative_type type); -void common_speculative_free(struct common_speculative * spec); +common_speculative * common_speculative_init( + const common_params_speculative & params, + llama_context * ctx_tgt); -bool common_speculative_are_compatible( - const struct llama_context * ctx_tgt, - const struct llama_context * ctx_dft); +void common_speculative_free(common_speculative * spec); -void common_speculative_add_replacement_tgt_dft( - struct common_speculative * spec, - const char *source, const char *dest); +// optionally call once at the beginning of a new generation +void common_speculative_begin(common_speculative * spec, const llama_tokens & prompt); // sample up to n_draft tokens and add them to the batch using the draft model -llama_tokens common_speculative_gen_draft( - struct common_speculative * spec, - struct common_speculative_params params, - const llama_tokens & prompt, - llama_token id_last); +llama_tokens common_speculative_draft( + common_speculative * spec, + const common_params_speculative & params, + const llama_tokens & prompt, + llama_token id_last); + +// informs the speculative decoder that n_accepted tokens were accepted by the target model +void common_speculative_accept(common_speculative * spec, uint16_t n_accepted); + +// print statistics about the speculative decoding +void common_speculative_print_stats(const common_speculative * spec); diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py index 8cc4963fb2..a391717e32 100755 --- a/convert_hf_to_gguf.py +++ b/convert_hf_to_gguf.py @@ -2736,7 +2736,7 @@ class AfmoeModel(LlamaModel): data_torch = torch.stack(datas, dim=0) merged_name = f"model.layers.{bid}.mlp.experts.{w_name}.weight" - yield from super().modify_tensors(data_torch, merged_name, bid) + yield from ModelBase.modify_tensors(self, data_torch, merged_name, bid) return else: @@ -2745,7 +2745,7 @@ class AfmoeModel(LlamaModel): if name.endswith(".expert_bias"): name = name.replace(".expert_bias", ".expert_bias.bias") - yield from super().modify_tensors(data_torch, name, bid) + yield from ModelBase.modify_tensors(self, data_torch, name, bid) @ModelBase.register( @@ -3799,7 +3799,7 @@ class Ernie4_5MoeModel(Ernie4_5Model): merged_name = f"model.layers.{bid}.mlp.experts.{w_name}.weight" yield from super().modify_tensors(data_torch, merged_name, bid) else: - yield from super().modify_tensors(data_torch, name, bid) + yield from ModelBase.modify_tensors(self, data_torch, name, bid) def prepare_tensors(self): super().prepare_tensors() @@ -6145,7 +6145,8 @@ class Gemma3nVisionAudioModel(ConformerAudioModel): if name.startswith("model.vision_tower.timm_model.blocks."): # Double-indexed block tensors through custom logic - new_name = self.custom_map(name) + yield (self.custom_map(name), data_torch) + return else: # Route non-repeating (conv_stem, msfa, embedding, etc.) and un-catched through tensor_mapping.py new_name = self.map_tensor_name(name) @@ -6153,7 +6154,7 @@ class Gemma3nVisionAudioModel(ConformerAudioModel): if new_name.endswith("conv_stem.conv.bias") or new_name.endswith("layer_scale.gamma"): data_torch = data_torch.unsqueeze(0).unsqueeze(-1).unsqueeze(-1) # [1, C, 1, 1] - yield from super().modify_tensors(data_torch, new_name, bid) + yield from ModelBase.modify_tensors(self, data_torch, new_name, bid) @ModelBase.register("Gemma3nForCausalLM", "Gemma3nForConditionalGeneration") @@ -6253,7 +6254,7 @@ class Gemma3NModel(Gemma3Model): # Continue with normal processing name = name.replace("language_model.", "") - yield from super().modify_tensors(data_torch, name, bid) + yield from ModelBase.modify_tensors(self, data_torch, name, bid) return if "altup_unembed_projections" in name: @@ -6270,7 +6271,7 @@ class Gemma3NModel(Gemma3Model): raise ValueError(f"Unknown name: {name}") out = self._stack_matrices(self._altup_unembd) if out is not None: - yield from super().modify_tensors(out, "model.altup_unembed_projections.weight", bid) + yield from ModelBase.modify_tensors(self, out, "model.altup_unembed_projections.weight", bid) return else: return @@ -6287,7 +6288,7 @@ class Gemma3NModel(Gemma3Model): raise ValueError(f"Unknown name: {name}") out = self._stack_matrices(self._altup_proj) if out is not None: - yield from super().modify_tensors(out, "model.altup_projections.weight", bid) + yield from ModelBase.modify_tensors(self, out, "model.altup_projections.weight", bid) return else: return @@ -8803,8 +8804,8 @@ class GraniteMoeModel(GraniteModel): ffn_dim = self.hparams["intermediate_size"] assert data_torch.shape[-2] == 2 * ffn_dim, "Merged FFN tensor size must be 2 * intermediate_size" gate, up = data_torch.split(ffn_dim, dim=-2) - yield from super().modify_tensors(gate, self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE_EXP, bid), bid) - yield from super().modify_tensors(up, self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP_EXP, bid), bid) + yield from ModelBase.modify_tensors(self, gate, self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE_EXP, bid), bid) + yield from ModelBase.modify_tensors(self, up, self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP_EXP, bid), bid) has_experts = bool(self.hparams.get('num_local_experts')) @@ -8813,15 +8814,15 @@ class GraniteMoeModel(GraniteModel): assert data_torch.shape[-2] == 2 * ffn_dim, "Merged FFN tensor size must be 2 * shared_intermediate_size" gate, up = data_torch.split(ffn_dim, dim=-2) if has_experts: - yield from super().modify_tensors(gate,self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE_SHEXP, bid), bid) - yield from super().modify_tensors(up, self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP_SHEXP, bid), bid) + yield from ModelBase.modify_tensors(self, gate,self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE_SHEXP, bid), bid) + yield from ModelBase.modify_tensors(self, up, self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP_SHEXP, bid), bid) return - yield from super().modify_tensors(gate, self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE, bid), bid) - yield from super().modify_tensors(up, self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP, bid), bid) + yield from ModelBase.modify_tensors(self, gate, self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE, bid), bid) + yield from ModelBase.modify_tensors(self, up, self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP, bid), bid) return if not has_experts and name.endswith("shared_mlp.output_linear.weight"): - yield from super().modify_tensors(data_torch, self.format_tensor_name(gguf.MODEL_TENSOR.FFN_DOWN, bid), bid) + yield from ModelBase.modify_tensors(self, data_torch, self.format_tensor_name(gguf.MODEL_TENSOR.FFN_DOWN, bid), bid) return yield from super().modify_tensors(data_torch, name, bid) @@ -8911,14 +8912,17 @@ class GraniteHybridModel(Mamba2Model, GraniteMoeModel): name.endswith("block_sparse_moe.input_linear.weight") or "shared_mlp" in name ): - return GraniteMoeModel.modify_tensors(self, data_torch, name, bid) + yield from GraniteMoeModel.modify_tensors(self, data_torch, name, bid) + return # Determine whether this is a mamba layer or an attention layer if bid in self._ssm_layers: - return Mamba2Model.modify_tensors(self, data_torch, name, bid) + yield from Mamba2Model.modify_tensors(self, data_torch, name, bid) + return elif bid in self._attn_layers: - return GraniteMoeModel.modify_tensors(self, data_torch, name, bid) - yield from super().modify_tensors(data_torch, name, bid) + yield from GraniteMoeModel.modify_tensors(self, data_torch, name, bid) + return + yield from ModelBase.modify_tensors(self, data_torch, name, bid) def set_gguf_parameters(self): """This method merges params from both parents and some that are @@ -9050,33 +9054,33 @@ class NemotronHModel(GraniteHybridModel): if self.is_moe and bid is not None: if name.endswith("mixer.gate.e_score_correction_bias"): new_name = name.replace("e_score_correction_bias", "e_score_correction.bias") - yield from super().modify_tensors(data_torch, new_name, bid) + yield from ModelBase.modify_tensors(self, data_torch, new_name, bid) return if name.endswith("mixer.dt_bias"): new_name = name.replace("dt_bias", "dt.bias") - yield from super().modify_tensors(data_torch, new_name, bid) + yield from ModelBase.modify_tensors(self, data_torch, new_name, bid) return if name.endswith("mixer.conv1d.weight"): squeezed_data = data_torch.squeeze() - yield from super().modify_tensors(squeezed_data, name, bid) + yield from ModelBase.modify_tensors(self, squeezed_data, name, bid) return if name.endswith("mixer.A_log"): transformed_data = -torch.exp(data_torch) reshaped_data = transformed_data.squeeze().reshape(-1, 1) - yield from super().modify_tensors(reshaped_data, name, bid) + yield from ModelBase.modify_tensors(self, reshaped_data, name, bid) return if name.endswith("mixer.D"): reshaped_data = data_torch.squeeze().reshape(-1, 1) - yield from super().modify_tensors(reshaped_data, name, bid) + yield from ModelBase.modify_tensors(self, reshaped_data, name, bid) return if name.endswith("mixer.norm.weight"): reshaped_data = data_torch.reshape(self.n_group, -1) - yield from super().modify_tensors(reshaped_data, name, bid) + yield from ModelBase.modify_tensors(self, reshaped_data, name, bid) return if name.find("mixer.experts") != -1: @@ -9101,7 +9105,7 @@ class NemotronHModel(GraniteHybridModel): data_torch = torch.stack(datas, dim=0) merged_name = f"model.layers.{bid}.mlp.experts.{w_name}.weight" - yield from super().modify_tensors(data_torch, merged_name, bid) + yield from ModelBase.modify_tensors(self, data_torch, merged_name, bid) return else: return @@ -10731,7 +10735,7 @@ class CogVLMModel(LlamaModel): if name.startswith("model.vision."): return - yield from super().modify_tensors(data_torch, name, bid) + yield from ModelBase.modify_tensors(self, data_torch, name, bid) @ModelBase.register("JanusForConditionalGeneration") diff --git a/docs/build.md b/docs/build.md index fce9361b2d..3a43f2a45a 100644 --- a/docs/build.md +++ b/docs/build.md @@ -144,7 +144,7 @@ We also have a [guide](./backend/CUDA-FEDORA.md) for setting up CUDA toolkit in - ***Necessary*** for users of [Atomic Desktops for Fedora](https://fedoraproject.org/atomic-desktops/); such as: [Silverblue](https://fedoraproject.org/atomic-desktops/silverblue/) and [Kinoite](https://fedoraproject.org/atomic-desktops/kinoite/). - (there are no supported CUDA packages for these systems) - ***Necessary*** for users that have a host that is not a: [Supported Nvidia CUDA Release Platform](https://developer.nvidia.com/cuda-downloads). - - (for example, you may have [Fedora 42 Beta](https://fedoramagazine.org/announcing-fedora-linux-42-beta/) as your your host operating system) + - (for example, you may have [Fedora 42 Beta](https://fedoramagazine.org/announcing-fedora-linux-42-beta/) as your host operating system) - ***Convenient*** For those running [Fedora Workstation](https://fedoraproject.org/workstation/) or [Fedora KDE Plasma Desktop](https://fedoraproject.org/spins/kde), and want to keep their host system clean. - *Optionally* toolbox packages are available: [Arch Linux](https://archlinux.org/), [Red Hat Enterprise Linux >= 8.5](https://www.redhat.com/en/technologies/linux-platforms/enterprise-linux), or [Ubuntu](https://ubuntu.com/download) @@ -248,6 +248,14 @@ You may set the [cuda environmental variables](https://docs.nvidia.com/cuda/cuda CUDA_VISIBLE_DEVICES="-0" ./build/bin/llama-server --model /srv/models/llama.gguf ``` +#### CUDA_SCALE_LAUNCH_QUEUES + +The environment variable [`CUDA_SCALE_LAUNCH_QUEUES`](https://docs.nvidia.com/cuda/cuda-programming-guide/05-appendices/environment-variables.html#cuda-scale-launch-queues) controls the size of CUDA's command buffer, which determines how many GPU operations can be queued before the CPU must wait for the GPU to catch up. A larger buffer reduces CPU-side stalls and allows more work to be queued on a GPU. + +**Default behavior:** llama.cpp automatically sets `CUDA_SCALE_LAUNCH_QUEUES=4x`, which increases the CUDA command buffer to 4 times its default size. This optimization is particularly beneficial for **Multi-GPU setups with pipeline parallelism**, where it significantly improves prompt processing throughput by allowing more operations to be enqueued across GPUs. + +See PR [#19042](https://github.com/ggml-org/llama.cpp/pull/19042) for performance benchmarks and technical details. + ### Unified Memory The environment variable `GGML_CUDA_ENABLE_UNIFIED_MEMORY=1` can be used to enable unified memory in Linux. This allows swapping to system RAM instead of crashing when the GPU VRAM is exhausted. In Windows this setting is available in the NVIDIA control panel as `System Memory Fallback`. @@ -487,6 +495,37 @@ Finally, after finishing your build, you should be able to do something like thi # ggml_vulkan: Using Intel(R) Graphics (ADL GT2) | uma: 1 | fp16: 1 | warp size: 32 ``` +### For Mac users: + +Generally, follow LunarG's [Getting Started with the MacOS Vulkan SDK](https://vulkan.lunarg.com/doc/sdk/latest/mac/getting_started.html) guide for installation and setup of the Vulkan SDK. There are two options of Vulkan drivers on macOS, both of which implement translation layers to map Vulkan to Metal. They can be hot-swapped by setting the `VK_ICD_FILENAMES` environment variable to point to the respective ICD JSON file. + +Check the box for "KosmicKrisp" during the LunarG Vulkan SDK installation. + +Set environment variable for the LunarG Vulkan SDK after installation (and optionally add to your shell profile for persistence): +```bash +source /path/to/vulkan-sdk/setup-env.sh +``` + +#### Using MoltenVK + +MoltenVK is the default Vulkan driver installed with the LunarG Vulkan SDK on macOS, so you can use the above environment variable settings as is. + +#### Using KosmicKrisp + +Override the environment variable for KosmicKrisp: +```bash +export VK_ICD_FILENAMES=$VULKAN_SDK/share/vulkan/icd.d/libkosmickrisp_icd.json +export VK_DRIVER_FILES=$VULKAN_SDK/share/vulkan/icd.d/libkosmickrisp_icd.json +``` + +#### Build + +This is the only step different from [above](#common-steps) instructions. +```bash +cmake -B build -DGGML_VULKAN=1 -DGGML_METAL=OFF +cmake --build build --config Release +``` + ## CANN This provides NPU acceleration using the AI cores of your Ascend NPU. And [CANN](https://www.hiascend.com/en/software/cann) is a hierarchical APIs to help you to quickly build AI applications and service based on Ascend NPU. diff --git a/docs/ops/SYCL.csv b/docs/ops/SYCL.csv index 91b442bde8..091a5caed7 100644 --- a/docs/ops/SYCL.csv +++ b/docs/ops/SYCL.csv @@ -878,6 +878,54 @@ "SYCL0","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=3,s0=2,s1=2,p0=0,p1=1","support","1","yes","SYCL" "SYCL0","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=3,s0=2,s1=2,p0=1,p1=0","support","1","yes","SYCL" "SYCL0","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=3,s0=2,s1=2,p0=1,p1=1","support","1","yes","SYCL" +"SYCL0","POOL_1D","pool_type=avg,type_input=f32,ne_input=[10,3,2,1],k0=1,s0=1,p0=0","support","0","no","SYCL" +"SYCL0","POOL_1D","pool_type=avg,type_input=f32,ne_input=[11,1,3,2],k0=1,s0=1,p0=0","support","0","no","SYCL" +"SYCL0","POOL_1D","pool_type=avg,type_input=f32,ne_input=[128,2,1,3],k0=1,s0=1,p0=0","support","0","no","SYCL" +"SYCL0","POOL_1D","pool_type=avg,type_input=f32,ne_input=[10,3,2,1],k0=1,s0=1,p0=1","support","0","no","SYCL" +"SYCL0","POOL_1D","pool_type=avg,type_input=f32,ne_input=[11,1,3,2],k0=1,s0=1,p0=1","support","0","no","SYCL" +"SYCL0","POOL_1D","pool_type=avg,type_input=f32,ne_input=[128,2,1,3],k0=1,s0=1,p0=1","support","0","no","SYCL" +"SYCL0","POOL_1D","pool_type=avg,type_input=f32,ne_input=[10,3,2,1],k0=1,s0=2,p0=0","support","0","no","SYCL" +"SYCL0","POOL_1D","pool_type=avg,type_input=f32,ne_input=[11,1,3,2],k0=1,s0=2,p0=0","support","0","no","SYCL" +"SYCL0","POOL_1D","pool_type=avg,type_input=f32,ne_input=[128,2,1,3],k0=1,s0=2,p0=0","support","0","no","SYCL" +"SYCL0","POOL_1D","pool_type=avg,type_input=f32,ne_input=[10,3,2,1],k0=1,s0=2,p0=1","support","0","no","SYCL" +"SYCL0","POOL_1D","pool_type=avg,type_input=f32,ne_input=[11,1,3,2],k0=1,s0=2,p0=1","support","0","no","SYCL" +"SYCL0","POOL_1D","pool_type=avg,type_input=f32,ne_input=[128,2,1,3],k0=1,s0=2,p0=1","support","0","no","SYCL" +"SYCL0","POOL_1D","pool_type=avg,type_input=f32,ne_input=[10,3,2,1],k0=3,s0=1,p0=0","support","0","no","SYCL" +"SYCL0","POOL_1D","pool_type=avg,type_input=f32,ne_input=[11,1,3,2],k0=3,s0=1,p0=0","support","0","no","SYCL" +"SYCL0","POOL_1D","pool_type=avg,type_input=f32,ne_input=[128,2,1,3],k0=3,s0=1,p0=0","support","0","no","SYCL" +"SYCL0","POOL_1D","pool_type=avg,type_input=f32,ne_input=[10,3,2,1],k0=3,s0=1,p0=1","support","0","no","SYCL" +"SYCL0","POOL_1D","pool_type=avg,type_input=f32,ne_input=[11,1,3,2],k0=3,s0=1,p0=1","support","0","no","SYCL" +"SYCL0","POOL_1D","pool_type=avg,type_input=f32,ne_input=[128,2,1,3],k0=3,s0=1,p0=1","support","0","no","SYCL" +"SYCL0","POOL_1D","pool_type=avg,type_input=f32,ne_input=[10,3,2,1],k0=3,s0=2,p0=0","support","0","no","SYCL" +"SYCL0","POOL_1D","pool_type=avg,type_input=f32,ne_input=[11,1,3,2],k0=3,s0=2,p0=0","support","0","no","SYCL" +"SYCL0","POOL_1D","pool_type=avg,type_input=f32,ne_input=[128,2,1,3],k0=3,s0=2,p0=0","support","0","no","SYCL" +"SYCL0","POOL_1D","pool_type=avg,type_input=f32,ne_input=[10,3,2,1],k0=3,s0=2,p0=1","support","0","no","SYCL" +"SYCL0","POOL_1D","pool_type=avg,type_input=f32,ne_input=[11,1,3,2],k0=3,s0=2,p0=1","support","0","no","SYCL" +"SYCL0","POOL_1D","pool_type=avg,type_input=f32,ne_input=[128,2,1,3],k0=3,s0=2,p0=1","support","0","no","SYCL" +"SYCL0","POOL_1D","pool_type=max,type_input=f32,ne_input=[10,3,2,1],k0=1,s0=1,p0=0","support","0","no","SYCL" +"SYCL0","POOL_1D","pool_type=max,type_input=f32,ne_input=[11,1,3,2],k0=1,s0=1,p0=0","support","0","no","SYCL" +"SYCL0","POOL_1D","pool_type=max,type_input=f32,ne_input=[128,2,1,3],k0=1,s0=1,p0=0","support","0","no","SYCL" +"SYCL0","POOL_1D","pool_type=max,type_input=f32,ne_input=[10,3,2,1],k0=1,s0=1,p0=1","support","0","no","SYCL" +"SYCL0","POOL_1D","pool_type=max,type_input=f32,ne_input=[11,1,3,2],k0=1,s0=1,p0=1","support","0","no","SYCL" +"SYCL0","POOL_1D","pool_type=max,type_input=f32,ne_input=[128,2,1,3],k0=1,s0=1,p0=1","support","0","no","SYCL" +"SYCL0","POOL_1D","pool_type=max,type_input=f32,ne_input=[10,3,2,1],k0=1,s0=2,p0=0","support","0","no","SYCL" +"SYCL0","POOL_1D","pool_type=max,type_input=f32,ne_input=[11,1,3,2],k0=1,s0=2,p0=0","support","0","no","SYCL" +"SYCL0","POOL_1D","pool_type=max,type_input=f32,ne_input=[128,2,1,3],k0=1,s0=2,p0=0","support","0","no","SYCL" +"SYCL0","POOL_1D","pool_type=max,type_input=f32,ne_input=[10,3,2,1],k0=1,s0=2,p0=1","support","0","no","SYCL" +"SYCL0","POOL_1D","pool_type=max,type_input=f32,ne_input=[11,1,3,2],k0=1,s0=2,p0=1","support","0","no","SYCL" +"SYCL0","POOL_1D","pool_type=max,type_input=f32,ne_input=[128,2,1,3],k0=1,s0=2,p0=1","support","0","no","SYCL" +"SYCL0","POOL_1D","pool_type=max,type_input=f32,ne_input=[10,3,2,1],k0=3,s0=1,p0=0","support","0","no","SYCL" +"SYCL0","POOL_1D","pool_type=max,type_input=f32,ne_input=[11,1,3,2],k0=3,s0=1,p0=0","support","0","no","SYCL" +"SYCL0","POOL_1D","pool_type=max,type_input=f32,ne_input=[128,2,1,3],k0=3,s0=1,p0=0","support","0","no","SYCL" +"SYCL0","POOL_1D","pool_type=max,type_input=f32,ne_input=[10,3,2,1],k0=3,s0=1,p0=1","support","0","no","SYCL" +"SYCL0","POOL_1D","pool_type=max,type_input=f32,ne_input=[11,1,3,2],k0=3,s0=1,p0=1","support","0","no","SYCL" +"SYCL0","POOL_1D","pool_type=max,type_input=f32,ne_input=[128,2,1,3],k0=3,s0=1,p0=1","support","0","no","SYCL" +"SYCL0","POOL_1D","pool_type=max,type_input=f32,ne_input=[10,3,2,1],k0=3,s0=2,p0=0","support","0","no","SYCL" +"SYCL0","POOL_1D","pool_type=max,type_input=f32,ne_input=[11,1,3,2],k0=3,s0=2,p0=0","support","0","no","SYCL" +"SYCL0","POOL_1D","pool_type=max,type_input=f32,ne_input=[128,2,1,3],k0=3,s0=2,p0=0","support","0","no","SYCL" +"SYCL0","POOL_1D","pool_type=max,type_input=f32,ne_input=[10,3,2,1],k0=3,s0=2,p0=1","support","0","no","SYCL" +"SYCL0","POOL_1D","pool_type=max,type_input=f32,ne_input=[11,1,3,2],k0=3,s0=2,p0=1","support","0","no","SYCL" +"SYCL0","POOL_1D","pool_type=max,type_input=f32,ne_input=[128,2,1,3],k0=3,s0=2,p0=1","support","0","no","SYCL" "SYCL0","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[3000,128,1,1],ne_kernel=[3,128,1280,1],s0=1,s1=0,p0=1,p1=0,d0=1,d1=0,is_2D=0","support","1","yes","SYCL" "SYCL0","IM2COL","type_input=f32,type_kernel=f16,dst_type=f32,ne_input=[3000,128,1,1],ne_kernel=[3,128,1280,1],s0=1,s1=0,p0=1,p1=0,d0=1,d1=0,is_2D=0","support","1","yes","SYCL" "SYCL0","IM2COL","type_input=f32,type_kernel=f16,dst_type=f16,ne_input=[3000,128,1,1],ne_kernel=[3,128,1280,1],s0=1,s1=0,p0=1,p1=0,d0=1,d1=0,is_2D=0","support","1","yes","SYCL" @@ -965,6 +1013,7 @@ "SYCL0","IM2COL","type_input=f32,type_kernel=f16,dst_type=f16,ne_input=[12,12,1,2560],ne_kernel=[3,3,1,2560],s0=1,s1=1,p0=1,p1=1,d0=1,d1=1,is_2D=1","support","1","yes","SYCL" "SYCL0","IM2COL","type_input=f32,type_kernel=f16,dst_type=f16,ne_input=[12,12,2,2560],ne_kernel=[3,3,2,2560],s0=1,s1=1,p0=1,p1=1,d0=1,d1=1,is_2D=1","support","1","yes","SYCL" "SYCL0","IM2COL","type_input=f32,type_kernel=f16,dst_type=f16,ne_input=[5,5,1,32],ne_kernel=[3,4,1,32],s0=1,s1=1,p0=0,p1=0,d0=1,d1=1,is_2D=1","support","1","yes","SYCL" +"SYCL0","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[2,2,1536,729],ne_kernel=[2,2,1536,4096],s0=1,s1=1,p0=0,p1=0,d0=1,d1=1,is_2D=1","support","1","yes","SYCL" "SYCL0","IM2COL_3D","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[10,10,10,9],ne_kernel=[3,3,3,1],IC=3,s0=1,s1=1,s2=1,p0=1,p1=1,p2=1,d0=1,d1=1,d2=1,v=0","support","0","no","SYCL" "SYCL0","IM2COL_3D","type_input=f32,type_kernel=f16,dst_type=f32,ne_input=[10,10,10,9],ne_kernel=[3,3,3,1],IC=3,s0=1,s1=1,s2=1,p0=1,p1=1,p2=1,d0=1,d1=1,d2=1,v=0","support","0","no","SYCL" "SYCL0","IM2COL_3D","type_input=f32,type_kernel=f16,dst_type=f16,ne_input=[10,10,10,9],ne_kernel=[3,3,3,1],IC=3,s0=1,s1=1,s2=1,p0=1,p1=1,p2=1,d0=1,d1=1,d2=1,v=0","support","0","no","SYCL" @@ -5696,35 +5745,58 @@ "SYCL0","RMS_NORM","type=f32,ne=[64,5,4,3],v=1,eps=0.000000,inplace=0","support","1","yes","SYCL" "SYCL0","RMS_NORM_BACK","type=f32,ne=[64,5,4,3],eps=0.000000","support","1","yes","SYCL" "SYCL0","L2_NORM","type=f32,ne=[64,5,4,3]","support","1","yes","SYCL" +"SYCL0","NORM","type=f32,ne=[1025,5,4,3],v=0,eps=0.000000","support","1","yes","SYCL" +"SYCL0","RMS_NORM","type=f32,ne=[1025,5,4,3],v=0,eps=0.000000,inplace=0","support","1","yes","SYCL" +"SYCL0","NORM","type=f32,ne=[1025,5,4,3],v=1,eps=0.000000","support","1","yes","SYCL" +"SYCL0","RMS_NORM","type=f32,ne=[1025,5,4,3],v=1,eps=0.000000,inplace=0","support","1","yes","SYCL" +"SYCL0","RMS_NORM_BACK","type=f32,ne=[1025,5,4,3],eps=0.000000","support","1","yes","SYCL" +"SYCL0","L2_NORM","type=f32,ne=[1025,5,4,3]","support","1","yes","SYCL" "SYCL0","NORM","type=f32,ne=[64,5,4,3],v=0,eps=0.000001","support","1","yes","SYCL" "SYCL0","RMS_NORM","type=f32,ne=[64,5,4,3],v=0,eps=0.000001,inplace=0","support","1","yes","SYCL" "SYCL0","NORM","type=f32,ne=[64,5,4,3],v=1,eps=0.000001","support","1","yes","SYCL" "SYCL0","RMS_NORM","type=f32,ne=[64,5,4,3],v=1,eps=0.000001,inplace=0","support","1","yes","SYCL" "SYCL0","RMS_NORM_BACK","type=f32,ne=[64,5,4,3],eps=0.000001","support","1","yes","SYCL" "SYCL0","L2_NORM","type=f32,ne=[64,5,4,3]","support","1","yes","SYCL" +"SYCL0","NORM","type=f32,ne=[1025,5,4,3],v=0,eps=0.000001","support","1","yes","SYCL" +"SYCL0","RMS_NORM","type=f32,ne=[1025,5,4,3],v=0,eps=0.000001,inplace=0","support","1","yes","SYCL" +"SYCL0","NORM","type=f32,ne=[1025,5,4,3],v=1,eps=0.000001","support","1","yes","SYCL" +"SYCL0","RMS_NORM","type=f32,ne=[1025,5,4,3],v=1,eps=0.000001,inplace=0","support","1","yes","SYCL" +"SYCL0","RMS_NORM_BACK","type=f32,ne=[1025,5,4,3],eps=0.000001","support","1","yes","SYCL" +"SYCL0","L2_NORM","type=f32,ne=[1025,5,4,3]","support","1","yes","SYCL" "SYCL0","NORM","type=f32,ne=[64,5,4,3],v=0,eps=0.000100","support","1","yes","SYCL" "SYCL0","RMS_NORM","type=f32,ne=[64,5,4,3],v=0,eps=0.000100,inplace=0","support","1","yes","SYCL" "SYCL0","NORM","type=f32,ne=[64,5,4,3],v=1,eps=0.000100","support","1","yes","SYCL" "SYCL0","RMS_NORM","type=f32,ne=[64,5,4,3],v=1,eps=0.000100,inplace=0","support","1","yes","SYCL" "SYCL0","RMS_NORM_BACK","type=f32,ne=[64,5,4,3],eps=0.000100","support","1","yes","SYCL" "SYCL0","L2_NORM","type=f32,ne=[64,5,4,3]","support","1","yes","SYCL" +"SYCL0","NORM","type=f32,ne=[1025,5,4,3],v=0,eps=0.000100","support","1","yes","SYCL" +"SYCL0","RMS_NORM","type=f32,ne=[1025,5,4,3],v=0,eps=0.000100,inplace=0","support","1","yes","SYCL" +"SYCL0","NORM","type=f32,ne=[1025,5,4,3],v=1,eps=0.000100","support","1","yes","SYCL" +"SYCL0","RMS_NORM","type=f32,ne=[1025,5,4,3],v=1,eps=0.000100,inplace=0","support","1","yes","SYCL" +"SYCL0","RMS_NORM_BACK","type=f32,ne=[1025,5,4,3],eps=0.000100","support","1","yes","SYCL" +"SYCL0","L2_NORM","type=f32,ne=[1025,5,4,3]","support","1","yes","SYCL" "SYCL0","NORM","type=f32,ne=[64,5,4,3],v=0,eps=0.100000","support","1","yes","SYCL" "SYCL0","RMS_NORM","type=f32,ne=[64,5,4,3],v=0,eps=0.100000,inplace=0","support","1","yes","SYCL" "SYCL0","NORM","type=f32,ne=[64,5,4,3],v=1,eps=0.100000","support","1","yes","SYCL" "SYCL0","RMS_NORM","type=f32,ne=[64,5,4,3],v=1,eps=0.100000,inplace=0","support","1","yes","SYCL" "SYCL0","RMS_NORM_BACK","type=f32,ne=[64,5,4,3],eps=0.100000","support","1","yes","SYCL" "SYCL0","L2_NORM","type=f32,ne=[64,5,4,3]","support","1","yes","SYCL" +"SYCL0","NORM","type=f32,ne=[1025,5,4,3],v=0,eps=0.100000","support","1","yes","SYCL" +"SYCL0","RMS_NORM","type=f32,ne=[1025,5,4,3],v=0,eps=0.100000,inplace=0","support","1","yes","SYCL" +"SYCL0","NORM","type=f32,ne=[1025,5,4,3],v=1,eps=0.100000","support","1","yes","SYCL" +"SYCL0","RMS_NORM","type=f32,ne=[1025,5,4,3],v=1,eps=0.100000,inplace=0","support","1","yes","SYCL" +"SYCL0","RMS_NORM_BACK","type=f32,ne=[1025,5,4,3],eps=0.100000","support","1","yes","SYCL" +"SYCL0","L2_NORM","type=f32,ne=[1025,5,4,3]","support","1","yes","SYCL" "SYCL0","RMS_NORM","type=f32,ne=[64,5,4,3],v=0,eps=0.000001,inplace=1","support","1","yes","SYCL" -"SYCL0","L2_NORM","type=f32,ne=[64,5,4,3]","support","1","yes","SYCL" -"SYCL0","SSM_CONV","type=f32,ne_a=[4,1024,1,1],ne_b=[3,1024,1,1]","support","1","yes","SYCL" -"SYCL0","SSM_CONV","type=f32,ne_a=[8,1024,1,1],ne_b=[3,1024,1,1]","support","1","yes","SYCL" -"SYCL0","SSM_CONV","type=f32,ne_a=[4,1024,4,1],ne_b=[3,1024,1,1]","support","1","yes","SYCL" -"SYCL0","SSM_CONV","type=f32,ne_a=[4,1536,1,1],ne_b=[3,1536,1,1]","support","1","yes","SYCL" -"SYCL0","SSM_CONV","type=f32,ne_a=[8,1536,1,1],ne_b=[3,1536,1,1]","support","1","yes","SYCL" -"SYCL0","SSM_CONV","type=f32,ne_a=[4,1536,4,1],ne_b=[3,1536,1,1]","support","1","yes","SYCL" -"SYCL0","SSM_CONV","type=f32,ne_a=[4,2048,1,1],ne_b=[3,2048,1,1]","support","1","yes","SYCL" -"SYCL0","SSM_CONV","type=f32,ne_a=[8,2048,1,1],ne_b=[3,2048,1,1]","support","1","yes","SYCL" -"SYCL0","SSM_CONV","type=f32,ne_a=[4,2048,4,1],ne_b=[3,2048,1,1]","support","1","yes","SYCL" +"SYCL0","SSM_CONV","type=f32,ne_a=[3,1024,1,1],ne_b=[3,1024,1,1]","support","1","yes","SYCL" +"SYCL0","SSM_CONV","type=f32,ne_a=[6,1024,1,1],ne_b=[3,1024,1,1]","support","1","yes","SYCL" +"SYCL0","SSM_CONV","type=f32,ne_a=[3,1024,4,1],ne_b=[3,1024,1,1]","support","1","yes","SYCL" +"SYCL0","SSM_CONV","type=f32,ne_a=[3,1536,1,1],ne_b=[3,1536,1,1]","support","1","yes","SYCL" +"SYCL0","SSM_CONV","type=f32,ne_a=[6,1536,1,1],ne_b=[3,1536,1,1]","support","1","yes","SYCL" +"SYCL0","SSM_CONV","type=f32,ne_a=[3,1536,4,1],ne_b=[3,1536,1,1]","support","1","yes","SYCL" +"SYCL0","SSM_CONV","type=f32,ne_a=[3,2048,1,1],ne_b=[3,2048,1,1]","support","1","yes","SYCL" +"SYCL0","SSM_CONV","type=f32,ne_a=[6,2048,1,1],ne_b=[3,2048,1,1]","support","1","yes","SYCL" +"SYCL0","SSM_CONV","type=f32,ne_a=[3,2048,4,1],ne_b=[3,2048,1,1]","support","1","yes","SYCL" "SYCL0","SSM_CONV","type=f32,ne_a=[4,1024,1,1],ne_b=[4,1024,1,1]","support","1","yes","SYCL" "SYCL0","SSM_CONV","type=f32,ne_a=[8,1024,1,1],ne_b=[4,1024,1,1]","support","1","yes","SYCL" "SYCL0","SSM_CONV","type=f32,ne_a=[4,1024,4,1],ne_b=[4,1024,1,1]","support","1","yes","SYCL" @@ -5734,6 +5806,15 @@ "SYCL0","SSM_CONV","type=f32,ne_a=[4,2048,1,1],ne_b=[4,2048,1,1]","support","1","yes","SYCL" "SYCL0","SSM_CONV","type=f32,ne_a=[8,2048,1,1],ne_b=[4,2048,1,1]","support","1","yes","SYCL" "SYCL0","SSM_CONV","type=f32,ne_a=[4,2048,4,1],ne_b=[4,2048,1,1]","support","1","yes","SYCL" +"SYCL0","SSM_CONV","type=f32,ne_a=[9,1024,1,1],ne_b=[9,1024,1,1]","support","1","yes","SYCL" +"SYCL0","SSM_CONV","type=f32,ne_a=[18,1024,1,1],ne_b=[9,1024,1,1]","support","1","yes","SYCL" +"SYCL0","SSM_CONV","type=f32,ne_a=[9,1024,4,1],ne_b=[9,1024,1,1]","support","1","yes","SYCL" +"SYCL0","SSM_CONV","type=f32,ne_a=[9,1536,1,1],ne_b=[9,1536,1,1]","support","1","yes","SYCL" +"SYCL0","SSM_CONV","type=f32,ne_a=[18,1536,1,1],ne_b=[9,1536,1,1]","support","1","yes","SYCL" +"SYCL0","SSM_CONV","type=f32,ne_a=[9,1536,4,1],ne_b=[9,1536,1,1]","support","1","yes","SYCL" +"SYCL0","SSM_CONV","type=f32,ne_a=[9,2048,1,1],ne_b=[9,2048,1,1]","support","1","yes","SYCL" +"SYCL0","SSM_CONV","type=f32,ne_a=[18,2048,1,1],ne_b=[9,2048,1,1]","support","1","yes","SYCL" +"SYCL0","SSM_CONV","type=f32,ne_a=[9,2048,4,1],ne_b=[9,2048,1,1]","support","1","yes","SYCL" "SYCL0","SSM_SCAN","type=f32,d_state=16,head_dim=1,n_head=1024,n_group=1,n_seq_tokens=32,n_seqs=4","support","0","no","SYCL" "SYCL0","SSM_SCAN","type=f32,d_state=128,head_dim=64,n_head=16,n_group=2,n_seq_tokens=32,n_seqs=4","support","0","no","SYCL" "SYCL0","SSM_SCAN","type=f32,d_state=256,head_dim=64,n_head=8,n_group=2,n_seq_tokens=32,n_seqs=4","support","0","no","SYCL" @@ -6593,6 +6674,30 @@ "SYCL0","MUL_MAT","type_a=f16,type_b=f32,m=1056,n=1,k=67,bs=[1,1],nr=[4,1],per=[0,2,1,3],k_v=0,o=1","support","1","yes","SYCL" "SYCL0","MUL_MAT","type_a=f32,type_b=f32,m=64,n=77,k=77,bs=[12,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","1","yes","SYCL" "SYCL0","MUL_MAT","type_a=q4_0,type_b=f32,m=576,n=512,k=576,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","1","yes","SYCL" +"SYCL0","MUL_MAT","type_a=q4_0,type_b=f32,m=1,n=2048,k=8192,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","1","yes","SYCL" +"SYCL0","MUL_MAT","type_a=f32,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","1","yes","SYCL" +"SYCL0","MUL_MAT","type_a=f16,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","1","yes","SYCL" +"SYCL0","MUL_MAT","type_a=bf16,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","0","no","SYCL" +"SYCL0","MUL_MAT","type_a=q4_0,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","1","yes","SYCL" +"SYCL0","MUL_MAT","type_a=q4_1,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","1","yes","SYCL" +"SYCL0","MUL_MAT","type_a=q5_0,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","1","yes","SYCL" +"SYCL0","MUL_MAT","type_a=q5_1,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","1","yes","SYCL" +"SYCL0","MUL_MAT","type_a=q8_0,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","1","yes","SYCL" +"SYCL0","MUL_MAT","type_a=mxfp4,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","1","yes","SYCL" +"SYCL0","MUL_MAT","type_a=q2_K,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","1","yes","SYCL" +"SYCL0","MUL_MAT","type_a=q3_K,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","1","yes","SYCL" +"SYCL0","MUL_MAT","type_a=q4_K,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","1","yes","SYCL" +"SYCL0","MUL_MAT","type_a=q5_K,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","1","yes","SYCL" +"SYCL0","MUL_MAT","type_a=q6_K,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","1","yes","SYCL" +"SYCL0","MUL_MAT","type_a=iq2_xxs,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","1","yes","SYCL" +"SYCL0","MUL_MAT","type_a=iq2_xs,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","1","yes","SYCL" +"SYCL0","MUL_MAT","type_a=iq2_s,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","1","yes","SYCL" +"SYCL0","MUL_MAT","type_a=iq3_xxs,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","1","yes","SYCL" +"SYCL0","MUL_MAT","type_a=iq1_s,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","1","yes","SYCL" +"SYCL0","MUL_MAT","type_a=iq1_m,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","1","yes","SYCL" +"SYCL0","MUL_MAT","type_a=iq4_nl,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","1","yes","SYCL" +"SYCL0","MUL_MAT","type_a=iq3_s,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","1","yes","SYCL" +"SYCL0","MUL_MAT","type_a=iq4_xs,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","1","yes","SYCL" "SYCL0","MUL_MAT","type_a=f16,type_b=f32,m=1056,n=1,k=128,bs=[1,1],nr=[1,1],per=[0,2,1,3],k_v=0,o=1","support","1","yes","SYCL" "SYCL0","MUL_MAT","type_a=f16,type_b=f32,m=128,n=1,k=1056,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=2112,o=1","support","1","yes","SYCL" "SYCL0","MUL_MAT","type_a=bf16,type_b=f32,m=1056,n=1,k=128,bs=[1,1],nr=[1,1],per=[0,2,1,3],k_v=0,o=1","support","0","no","SYCL" @@ -8917,6 +9022,11 @@ "SYCL0","SOFT_MAX","type=f32,ne=[32,2,32,1],mask=1,sinks=0,m_prec=f16,nr23=[1,1],scale=0.100000,max_bias=0.000000,inplace=0","support","1","yes","SYCL" "SYCL0","SOFT_MAX","type=f32,ne=[32,2,32,1],mask=1,sinks=1,m_prec=f32,nr23=[1,1],scale=0.100000,max_bias=8.000000,inplace=0","support","1","yes","SYCL" "SYCL0","SOFT_MAX","type=f32,ne=[32,2,32,1],mask=1,sinks=1,m_prec=f16,nr23=[1,1],scale=0.100000,max_bias=8.000000,inplace=0","support","1","yes","SYCL" +"SYCL0","SOFT_MAX","type=f32,ne=[200001,2,3,1],mask=1,sinks=1,m_prec=f32,nr23=[1,1],scale=0.100000,max_bias=8.000000,inplace=0","support","1","yes","SYCL" +"SYCL0","SOFT_MAX","type=f32,ne=[200001,2,3,1],mask=1,sinks=1,m_prec=f16,nr23=[1,1],scale=0.100000,max_bias=8.000000,inplace=0","support","1","yes","SYCL" +"SYCL0","SOFT_MAX","type=f32,ne=[200000,1,1,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0","support","1","yes","SYCL" +"SYCL0","SOFT_MAX","type=f32,ne=[200000,4,1,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0","support","1","yes","SYCL" +"SYCL0","SOFT_MAX","type=f32,ne=[643251,3,1,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0","support","1","yes","SYCL" "SYCL0","SOFT_MAX_BACK","type=f32,ne=[16,16,1,1],scale=1.000000,max_bias=0.000000","support","1","yes","SYCL" "SYCL0","SOFT_MAX_BACK","type=f32,ne=[15,15,1,1],scale=1.000000,max_bias=0.000000","support","1","yes","SYCL" "SYCL0","SOFT_MAX_BACK","type=f32,ne=[16,16,2,3],scale=1.000000,max_bias=0.000000","support","1","yes","SYCL" @@ -8969,6 +9079,7 @@ "SYCL0","ROPE","type=f32,ne_a=[128,40,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","1","yes","SYCL" "SYCL0","ROPE","type=f32,ne_a=[128,52,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","1","yes","SYCL" "SYCL0","ROPE","type=f32,ne_a=[128,64,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","1","yes","SYCL" +"SYCL0","ROPE","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","1","yes","SYCL" "SYCL0","ROPE","type=f32,ne_a=[64,1,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","1","yes","SYCL" "SYCL0","ROPE","type=f32,ne_a=[64,71,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","1","yes","SYCL" "SYCL0","ROPE","type=f32,ne_a=[64,8,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","1","yes","SYCL" @@ -8978,6 +9089,7 @@ "SYCL0","ROPE","type=f32,ne_a=[80,32,2,1],n_dims=20,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","1","yes","SYCL" "SYCL0","ROPE","type=f32,ne_a=[80,32,2,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","1","yes","SYCL" "SYCL0","ROPE","type=f32,ne_a=[80,32,4,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","1","yes","SYCL" +"SYCL0","ROPE","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","1","yes","SYCL" "SYCL0","ROPE","type=f32,ne_a=[128,12,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","1","yes","SYCL" "SYCL0","ROPE","type=f32,ne_a=[128,28,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","1","yes","SYCL" "SYCL0","ROPE","type=f32,ne_a=[128,12,2,1],n_dims=20,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","1","yes","SYCL" @@ -8988,11 +9100,13 @@ "SYCL0","ROPE","type=f32,ne_a=[128,28,2,1],n_dims=32,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","1","yes","SYCL" "SYCL0","ROPE","type=f32,ne_a=[80,16,2,1],n_dims=80,mode=24,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","1","yes","SYCL" "SYCL0","ROPE","type=f32,ne_a=[128,16,2,1],n_dims=128,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","1","yes","SYCL" +"SYCL0","ROPE","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","1","yes","SYCL" "SYCL0","ROPE","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","1","yes","SYCL" "SYCL0","ROPE","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","1","yes","SYCL" "SYCL0","ROPE","type=f32,ne_a=[128,40,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","1","yes","SYCL" "SYCL0","ROPE","type=f32,ne_a=[128,52,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","1","yes","SYCL" "SYCL0","ROPE","type=f32,ne_a=[128,64,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","1","yes","SYCL" +"SYCL0","ROPE","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","1","yes","SYCL" "SYCL0","ROPE","type=f32,ne_a=[64,1,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","1","yes","SYCL" "SYCL0","ROPE","type=f32,ne_a=[64,71,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","1","yes","SYCL" "SYCL0","ROPE","type=f32,ne_a=[64,8,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","1","yes","SYCL" @@ -9002,6 +9116,7 @@ "SYCL0","ROPE","type=f32,ne_a=[80,32,2,1],n_dims=20,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","1","yes","SYCL" "SYCL0","ROPE","type=f32,ne_a=[80,32,2,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","1","yes","SYCL" "SYCL0","ROPE","type=f32,ne_a=[80,32,4,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","1","yes","SYCL" +"SYCL0","ROPE","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","1","yes","SYCL" "SYCL0","ROPE","type=f32,ne_a=[128,12,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","1","yes","SYCL" "SYCL0","ROPE","type=f32,ne_a=[128,28,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","1","yes","SYCL" "SYCL0","ROPE","type=f32,ne_a=[128,12,2,1],n_dims=20,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","1","yes","SYCL" @@ -9012,11 +9127,13 @@ "SYCL0","ROPE","type=f32,ne_a=[128,28,2,1],n_dims=32,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","1","yes","SYCL" "SYCL0","ROPE","type=f32,ne_a=[80,16,2,1],n_dims=80,mode=24,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","1","yes","SYCL" "SYCL0","ROPE","type=f32,ne_a=[128,16,2,1],n_dims=128,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","1","yes","SYCL" +"SYCL0","ROPE","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","1","yes","SYCL" "SYCL0","ROPE","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","1","yes","SYCL" "SYCL0","ROPE","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","1","yes","SYCL" "SYCL0","ROPE","type=f32,ne_a=[128,40,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","1","yes","SYCL" "SYCL0","ROPE","type=f32,ne_a=[128,52,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","1","yes","SYCL" "SYCL0","ROPE","type=f32,ne_a=[128,64,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","1","yes","SYCL" +"SYCL0","ROPE","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","1","yes","SYCL" "SYCL0","ROPE","type=f32,ne_a=[64,1,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","1","yes","SYCL" "SYCL0","ROPE","type=f32,ne_a=[64,71,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","1","yes","SYCL" "SYCL0","ROPE","type=f32,ne_a=[64,8,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","1","yes","SYCL" @@ -9026,6 +9143,7 @@ "SYCL0","ROPE","type=f32,ne_a=[80,32,2,1],n_dims=20,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","1","yes","SYCL" "SYCL0","ROPE","type=f32,ne_a=[80,32,2,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","1","yes","SYCL" "SYCL0","ROPE","type=f32,ne_a=[80,32,4,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","1","yes","SYCL" +"SYCL0","ROPE","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","1","yes","SYCL" "SYCL0","ROPE","type=f32,ne_a=[128,12,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","1","yes","SYCL" "SYCL0","ROPE","type=f32,ne_a=[128,28,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","1","yes","SYCL" "SYCL0","ROPE","type=f32,ne_a=[128,12,2,1],n_dims=20,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","1","yes","SYCL" @@ -9036,11 +9154,13 @@ "SYCL0","ROPE","type=f32,ne_a=[128,28,2,1],n_dims=32,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","1","yes","SYCL" "SYCL0","ROPE","type=f32,ne_a=[80,16,2,1],n_dims=80,mode=24,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","1","yes","SYCL" "SYCL0","ROPE","type=f32,ne_a=[128,16,2,1],n_dims=128,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","1","yes","SYCL" +"SYCL0","ROPE","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","1","yes","SYCL" "SYCL0","ROPE","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","1","yes","SYCL" "SYCL0","ROPE","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","1","yes","SYCL" "SYCL0","ROPE","type=f32,ne_a=[128,40,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","1","yes","SYCL" "SYCL0","ROPE","type=f32,ne_a=[128,52,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","1","yes","SYCL" "SYCL0","ROPE","type=f32,ne_a=[128,64,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","1","yes","SYCL" +"SYCL0","ROPE","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","1","yes","SYCL" "SYCL0","ROPE","type=f32,ne_a=[64,1,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","1","yes","SYCL" "SYCL0","ROPE","type=f32,ne_a=[64,71,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","1","yes","SYCL" "SYCL0","ROPE","type=f32,ne_a=[64,8,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","1","yes","SYCL" @@ -9050,6 +9170,7 @@ "SYCL0","ROPE","type=f32,ne_a=[80,32,2,1],n_dims=20,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","1","yes","SYCL" "SYCL0","ROPE","type=f32,ne_a=[80,32,2,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","1","yes","SYCL" "SYCL0","ROPE","type=f32,ne_a=[80,32,4,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","1","yes","SYCL" +"SYCL0","ROPE","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","1","yes","SYCL" "SYCL0","ROPE","type=f32,ne_a=[128,12,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","1","yes","SYCL" "SYCL0","ROPE","type=f32,ne_a=[128,28,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","1","yes","SYCL" "SYCL0","ROPE","type=f32,ne_a=[128,12,2,1],n_dims=20,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","1","yes","SYCL" @@ -9060,6 +9181,7 @@ "SYCL0","ROPE","type=f32,ne_a=[128,28,2,1],n_dims=32,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","1","yes","SYCL" "SYCL0","ROPE","type=f32,ne_a=[80,16,2,1],n_dims=80,mode=24,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","1","yes","SYCL" "SYCL0","ROPE","type=f32,ne_a=[128,16,2,1],n_dims=128,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","1","yes","SYCL" +"SYCL0","ROPE","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","1","yes","SYCL" "SYCL0","ROPE","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","1","yes","SYCL" "SYCL0","ROPE","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","1","yes","SYCL" "SYCL0","ROPE","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","1","yes","SYCL" @@ -9185,6 +9307,7 @@ "SYCL0","ROPE_BACK","type=f32,ne_a=[128,40,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","SYCL" "SYCL0","ROPE_BACK","type=f32,ne_a=[128,52,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","SYCL" "SYCL0","ROPE_BACK","type=f32,ne_a=[128,64,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","SYCL" +"SYCL0","ROPE_BACK","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","SYCL" "SYCL0","ROPE_BACK","type=f32,ne_a=[64,1,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","SYCL" "SYCL0","ROPE_BACK","type=f32,ne_a=[64,71,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","SYCL" "SYCL0","ROPE_BACK","type=f32,ne_a=[64,8,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","SYCL" @@ -9194,6 +9317,7 @@ "SYCL0","ROPE_BACK","type=f32,ne_a=[80,32,2,1],n_dims=20,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","SYCL" "SYCL0","ROPE_BACK","type=f32,ne_a=[80,32,2,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","SYCL" "SYCL0","ROPE_BACK","type=f32,ne_a=[80,32,4,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","SYCL" +"SYCL0","ROPE_BACK","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","SYCL" "SYCL0","ROPE_BACK","type=f32,ne_a=[128,12,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","SYCL" "SYCL0","ROPE_BACK","type=f32,ne_a=[128,28,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","SYCL" "SYCL0","ROPE_BACK","type=f32,ne_a=[128,12,2,1],n_dims=20,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","SYCL" @@ -9204,11 +9328,13 @@ "SYCL0","ROPE_BACK","type=f32,ne_a=[128,28,2,1],n_dims=32,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","SYCL" "SYCL0","ROPE_BACK","type=f32,ne_a=[80,16,2,1],n_dims=80,mode=24,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","SYCL" "SYCL0","ROPE_BACK","type=f32,ne_a=[128,16,2,1],n_dims=128,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","SYCL" +"SYCL0","ROPE_BACK","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","SYCL" "SYCL0","ROPE_BACK","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","SYCL" "SYCL0","ROPE_BACK","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","SYCL" "SYCL0","ROPE_BACK","type=f32,ne_a=[128,40,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","SYCL" "SYCL0","ROPE_BACK","type=f32,ne_a=[128,52,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","SYCL" "SYCL0","ROPE_BACK","type=f32,ne_a=[128,64,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","SYCL" +"SYCL0","ROPE_BACK","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","SYCL" "SYCL0","ROPE_BACK","type=f32,ne_a=[64,1,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","SYCL" "SYCL0","ROPE_BACK","type=f32,ne_a=[64,71,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","SYCL" "SYCL0","ROPE_BACK","type=f32,ne_a=[64,8,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","SYCL" @@ -9218,6 +9344,7 @@ "SYCL0","ROPE_BACK","type=f32,ne_a=[80,32,2,1],n_dims=20,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","SYCL" "SYCL0","ROPE_BACK","type=f32,ne_a=[80,32,2,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","SYCL" "SYCL0","ROPE_BACK","type=f32,ne_a=[80,32,4,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","SYCL" +"SYCL0","ROPE_BACK","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","SYCL" "SYCL0","ROPE_BACK","type=f32,ne_a=[128,12,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","SYCL" "SYCL0","ROPE_BACK","type=f32,ne_a=[128,28,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","SYCL" "SYCL0","ROPE_BACK","type=f32,ne_a=[128,12,2,1],n_dims=20,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","SYCL" @@ -9228,11 +9355,13 @@ "SYCL0","ROPE_BACK","type=f32,ne_a=[128,28,2,1],n_dims=32,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","SYCL" "SYCL0","ROPE_BACK","type=f32,ne_a=[80,16,2,1],n_dims=80,mode=24,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","SYCL" "SYCL0","ROPE_BACK","type=f32,ne_a=[128,16,2,1],n_dims=128,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","SYCL" +"SYCL0","ROPE_BACK","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","SYCL" "SYCL0","ROPE_BACK","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","SYCL" "SYCL0","ROPE_BACK","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","SYCL" "SYCL0","ROPE_BACK","type=f32,ne_a=[128,40,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","SYCL" "SYCL0","ROPE_BACK","type=f32,ne_a=[128,52,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","SYCL" "SYCL0","ROPE_BACK","type=f32,ne_a=[128,64,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","SYCL" +"SYCL0","ROPE_BACK","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","SYCL" "SYCL0","ROPE_BACK","type=f32,ne_a=[64,1,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","SYCL" "SYCL0","ROPE_BACK","type=f32,ne_a=[64,71,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","SYCL" "SYCL0","ROPE_BACK","type=f32,ne_a=[64,8,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","SYCL" @@ -9242,6 +9371,7 @@ "SYCL0","ROPE_BACK","type=f32,ne_a=[80,32,2,1],n_dims=20,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","SYCL" "SYCL0","ROPE_BACK","type=f32,ne_a=[80,32,2,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","SYCL" "SYCL0","ROPE_BACK","type=f32,ne_a=[80,32,4,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","SYCL" +"SYCL0","ROPE_BACK","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","SYCL" "SYCL0","ROPE_BACK","type=f32,ne_a=[128,12,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","SYCL" "SYCL0","ROPE_BACK","type=f32,ne_a=[128,28,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","SYCL" "SYCL0","ROPE_BACK","type=f32,ne_a=[128,12,2,1],n_dims=20,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","SYCL" @@ -9252,11 +9382,13 @@ "SYCL0","ROPE_BACK","type=f32,ne_a=[128,28,2,1],n_dims=32,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","SYCL" "SYCL0","ROPE_BACK","type=f32,ne_a=[80,16,2,1],n_dims=80,mode=24,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","SYCL" "SYCL0","ROPE_BACK","type=f32,ne_a=[128,16,2,1],n_dims=128,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","SYCL" +"SYCL0","ROPE_BACK","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","SYCL" "SYCL0","ROPE_BACK","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","SYCL" "SYCL0","ROPE_BACK","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","SYCL" "SYCL0","ROPE_BACK","type=f32,ne_a=[128,40,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","SYCL" "SYCL0","ROPE_BACK","type=f32,ne_a=[128,52,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","SYCL" "SYCL0","ROPE_BACK","type=f32,ne_a=[128,64,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","SYCL" +"SYCL0","ROPE_BACK","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","SYCL" "SYCL0","ROPE_BACK","type=f32,ne_a=[64,1,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","SYCL" "SYCL0","ROPE_BACK","type=f32,ne_a=[64,71,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","SYCL" "SYCL0","ROPE_BACK","type=f32,ne_a=[64,8,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","SYCL" @@ -9266,6 +9398,7 @@ "SYCL0","ROPE_BACK","type=f32,ne_a=[80,32,2,1],n_dims=20,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","SYCL" "SYCL0","ROPE_BACK","type=f32,ne_a=[80,32,2,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","SYCL" "SYCL0","ROPE_BACK","type=f32,ne_a=[80,32,4,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","SYCL" +"SYCL0","ROPE_BACK","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","SYCL" "SYCL0","ROPE_BACK","type=f32,ne_a=[128,12,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","SYCL" "SYCL0","ROPE_BACK","type=f32,ne_a=[128,28,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","SYCL" "SYCL0","ROPE_BACK","type=f32,ne_a=[128,12,2,1],n_dims=20,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","SYCL" @@ -9276,6 +9409,7 @@ "SYCL0","ROPE_BACK","type=f32,ne_a=[128,28,2,1],n_dims=32,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","SYCL" "SYCL0","ROPE_BACK","type=f32,ne_a=[80,16,2,1],n_dims=80,mode=24,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","SYCL" "SYCL0","ROPE_BACK","type=f32,ne_a=[128,16,2,1],n_dims=128,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","SYCL" +"SYCL0","ROPE_BACK","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","SYCL" "SYCL0","ROPE_BACK","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","SYCL" "SYCL0","ROPE_BACK","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","SYCL" "SYCL0","ROPE_BACK","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","SYCL" @@ -9850,26 +9984,26 @@ "SYCL0","TOP_K","type=f32,ne=[2049,2,1,3],k=15,ties=0","support","0","no","SYCL" "SYCL0","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=nearest,transpose=0","support","1","yes","SYCL" "SYCL0","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=nearest,transpose=1","support","1","yes","SYCL" -"SYCL0","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=nearest,flags=none","support","1","yes","SYCL" -"SYCL0","UPSCALE","type=f32,ne=[5,7,11,13],ne_tgt=[2,5,7,11],mode=nearest,flags=none","support","1","yes","SYCL" +"SYCL0","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=nearest","support","1","yes","SYCL" +"SYCL0","UPSCALE","type=f32,ne=[5,7,11,13],ne_tgt=[2,5,7,11],mode=nearest","support","1","yes","SYCL" "SYCL0","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=bilinear,transpose=0","support","0","no","SYCL" "SYCL0","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=bilinear,transpose=1","support","0","no","SYCL" -"SYCL0","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=bilinear,flags=none","support","0","no","SYCL" -"SYCL0","UPSCALE","type=f32,ne=[5,7,11,13],ne_tgt=[2,5,7,11],mode=bilinear,flags=none","support","0","no","SYCL" +"SYCL0","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=bilinear","support","0","no","SYCL" +"SYCL0","UPSCALE","type=f32,ne=[5,7,11,13],ne_tgt=[2,5,7,11],mode=bilinear","support","0","no","SYCL" "SYCL0","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=bicubic,transpose=0","support","0","no","SYCL" "SYCL0","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=bicubic,transpose=1","support","0","no","SYCL" -"SYCL0","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=bicubic,flags=none","support","0","no","SYCL" -"SYCL0","UPSCALE","type=f32,ne=[5,7,11,13],ne_tgt=[2,5,7,11],mode=bicubic,flags=none","support","0","no","SYCL" -"SYCL0","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=513,transpose=0","support","0","no","SYCL" -"SYCL0","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=513,transpose=1","support","0","no","SYCL" -"SYCL0","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=bilinear,flags=none","support","0","no","SYCL" -"SYCL0","UPSCALE","type=f32,ne=[5,7,11,13],ne_tgt=[2,5,7,11],mode=bilinear,flags=none","support","0","no","SYCL" -"SYCL0","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=bilinear,flags=align_corners","support","0","no","SYCL" -"SYCL0","UPSCALE","type=f32,ne=[1,4,3,2],ne_tgt=[2,8,3,2],mode=bilinear,flags=align_corners","support","0","no","SYCL" -"SYCL0","UPSCALE","type=f32,ne=[4,1,3,2],ne_tgt=[1,1,3,2],mode=bilinear,flags=align_corners","support","0","no","SYCL" -"SYCL0","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=bicubic,flags=align_corners","support","0","no","SYCL" -"SYCL0","UPSCALE","type=f32,ne=[1,4,3,2],ne_tgt=[2,8,3,2],mode=bicubic,flags=align_corners","support","0","no","SYCL" -"SYCL0","UPSCALE","type=f32,ne=[4,1,3,2],ne_tgt=[1,1,3,2],mode=bicubic,flags=align_corners","support","0","no","SYCL" +"SYCL0","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=bicubic","support","0","no","SYCL" +"SYCL0","UPSCALE","type=f32,ne=[5,7,11,13],ne_tgt=[2,5,7,11],mode=bicubic","support","0","no","SYCL" +"SYCL0","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=bilinear|antialias,transpose=0","support","0","no","SYCL" +"SYCL0","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=bilinear|antialias,transpose=1","support","0","no","SYCL" +"SYCL0","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=bilinear|antialias","support","0","no","SYCL" +"SYCL0","UPSCALE","type=f32,ne=[5,7,11,13],ne_tgt=[2,5,7,11],mode=bilinear|antialias","support","0","no","SYCL" +"SYCL0","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=bilinear|align_corners","support","0","no","SYCL" +"SYCL0","UPSCALE","type=f32,ne=[1,4,3,2],ne_tgt=[2,8,3,2],mode=bilinear|align_corners","support","0","no","SYCL" +"SYCL0","UPSCALE","type=f32,ne=[4,1,3,2],ne_tgt=[1,1,3,2],mode=bilinear|align_corners","support","0","no","SYCL" +"SYCL0","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=bicubic|align_corners","support","0","no","SYCL" +"SYCL0","UPSCALE","type=f32,ne=[1,4,3,2],ne_tgt=[2,8,3,2],mode=bicubic|align_corners","support","0","no","SYCL" +"SYCL0","UPSCALE","type=f32,ne=[4,1,3,2],ne_tgt=[1,1,3,2],mode=bicubic|align_corners","support","0","no","SYCL" "SYCL0","SUM","type=f32,ne=[10,5,4,3]","support","1","yes","SYCL" "SYCL0","SUM_ROWS","type=f32,ne=[10,5,4,3],permute=0,slice=0","support","1","yes","SYCL" "SYCL0","SUM","type=f32,ne=[11,5,6,3],permute=[0,2,1,3]","support","0","no","SYCL" @@ -9892,8 +10026,9 @@ "SYCL0","GROUP_NORM","type=f32,ne=[64,64,320,1],num_groups=32,eps=0.000001","support","1","yes","SYCL" "SYCL0","GROUP_NORM","type=f32,ne=[9,9,1280,1],num_groups=32,eps=0.000001","support","1","yes","SYCL" "SYCL0","ACC","type=f32,ne_a=[256,17,1,1],ne_b=[256,16,1,1]","support","1","yes","SYCL" -"SYCL0","PAD","type=f32,ne_a=[512,512,1,1],pad_0=1,pad_1=1","support","1","yes","SYCL" -"SYCL0","PAD","type=f32,ne_a=[512,512,3,1],lp0=1,rp0=1,lp1=1,rp1=1,lp2=1,rp2=1,lp3=1,rp3=1,v=0","support","1","yes","SYCL" +"SYCL0","PAD","type=f32,ne_a=[512,512,1,1],pad_0=1,pad_1=1,circular=0","support","1","yes","SYCL" +"SYCL0","PAD","type=f32,ne_a=[33,17,2,1],pad_0=4,pad_1=3,circular=1","support","0","no","SYCL" +"SYCL0","PAD","type=f32,ne_a=[512,512,3,1],lp0=1,rp0=1,lp1=1,rp1=1,lp2=1,rp2=1,lp3=1,rp3=1,v=0,circular=0","support","1","yes","SYCL" "SYCL0","PAD_REFLECT_1D","type=f32,ne_a=[512,34,2,1],pad_0=10,pad_1=9","support","1","yes","SYCL" "SYCL0","PAD_REFLECT_1D","type=f32,ne_a=[3000,384,4,1],pad_0=10,pad_1=9","support","1","yes","SYCL" "SYCL0","ROLL","shift0=3,shift1=-2,shift3=1,shift4=-1","support","1","yes","SYCL" @@ -9915,6 +10050,7 @@ "SYCL0","CUMSUM","type=f32,ne=[2048,5,4,3]","support","0","no","SYCL" "SYCL0","CUMSUM","type=f32,ne=[242004,1,1,1]","support","0","no","SYCL" "SYCL0","CUMSUM","type=f32,ne=[375960,1,1,1]","support","0","no","SYCL" +"SYCL0","CUMSUM","type=f32,ne=[20481,4,1,1]","support","0","no","SYCL" "SYCL0","XIELU","type=f32,ne=[10,5,4,3]","support","0","no","SYCL" "SYCL0","TRI","type=f32,ne=[10,10,4,3],tri_type=3","support","0","no","SYCL" "SYCL0","TRI","type=f32,ne=[10,10,4,3],tri_type=2","support","0","no","SYCL" @@ -9924,19 +10060,41 @@ "SYCL0","FILL","type=f32,ne=[303,207,11,3],c=2.000000","support","0","no","SYCL" "SYCL0","FILL","type=f32,ne=[800,600,4,4],c=-152.000000","support","0","no","SYCL" "SYCL0","FILL","type=f32,ne=[2048,512,2,2],c=3.500000","support","0","no","SYCL" +"SYCL0","DIAG","type=f32,ne=[10,1,4,3]","support","0","no","SYCL" +"SYCL0","DIAG","type=f32,ne=[79,1,19,13]","support","0","no","SYCL" +"SYCL0","DIAG","type=f32,ne=[256,1,8,16]","support","0","no","SYCL" "SYCL0","SOLVE_TRI","type=f32,ne_lhs=[10,10,4,3],ne_rhs=[3,10,4,3]","support","0","no","SYCL" "SYCL0","SOLVE_TRI","type=f32,ne_lhs=[11,11,1,1],ne_rhs=[5,11,1,1]","support","0","no","SYCL" "SYCL0","SOLVE_TRI","type=f32,ne_lhs=[17,17,2,4],ne_rhs=[9,17,2,4]","support","0","no","SYCL" "SYCL0","SOLVE_TRI","type=f32,ne_lhs=[30,30,7,1],ne_rhs=[8,30,7,1]","support","0","no","SYCL" "SYCL0","SOLVE_TRI","type=f32,ne_lhs=[42,42,5,2],ne_rhs=[10,42,5,2]","support","0","no","SYCL" "SYCL0","SOLVE_TRI","type=f32,ne_lhs=[64,64,2,2],ne_rhs=[10,64,2,2]","support","0","no","SYCL" +"SYCL0","SOLVE_TRI","type=f32,ne_lhs=[64,64,2,2],ne_rhs=[64,64,2,2]","support","0","no","SYCL" +"SYCL0","SOLVE_TRI","type=f32,ne_lhs=[79,79,5,3],ne_rhs=[417,79,5,3]","support","0","no","SYCL" +"SYCL0","SOLVE_TRI","type=f32,ne_lhs=[128,128,4,2],ne_rhs=[32,128,4,2]","support","0","no","SYCL" +"SYCL0","SOLVE_TRI","type=f32,ne_lhs=[80,80,2,8],ne_rhs=[80,80,2,8]","support","0","no","SYCL" +"SYCL0","SOLVE_TRI","type=f32,ne_lhs=[80,80,2,8],ne_rhs=[79,80,2,8]","support","0","no","SYCL" +"SYCL0","SOLVE_TRI","type=f32,ne_lhs=[80,80,2,8],ne_rhs=[81,80,2,8]","support","0","no","SYCL" +"SYCL0","SOLVE_TRI","type=f32,ne_lhs=[80,80,8,8],ne_rhs=[80,80,8,8]","support","0","no","SYCL" +"SYCL0","SOLVE_TRI","type=f32,ne_lhs=[80,80,8,8],ne_rhs=[79,80,8,8]","support","0","no","SYCL" +"SYCL0","SOLVE_TRI","type=f32,ne_lhs=[80,80,8,8],ne_rhs=[81,80,8,8]","support","0","no","SYCL" +"SYCL0","SOLVE_TRI","type=f32,ne_lhs=[84,84,4,4],ne_rhs=[32,84,4,4]","support","0","no","SYCL" +"SYCL0","SOLVE_TRI","type=f32,ne_lhs=[95,95,8,8],ne_rhs=[40,95,8,8]","support","0","no","SYCL" "SYCL0","SOLVE_TRI","type=f32,ne_lhs=[100,100,4,4],ne_rhs=[41,100,4,4]","support","0","no","SYCL" "SYCL0","SOLVE_TRI","type=f32,ne_lhs=[128,128,4,4],ne_rhs=[31,128,4,4]","support","0","no","SYCL" -"SYCL0","SOLVE_TRI","type=f32,ne_lhs=[64,64,4,4],ne_rhs=[300,64,4,4]","support","0","no","SYCL" -"SYCL0","PAD","type=f32,ne_a=[512,512,1,1],lp0=0,rp0=1,lp1=0,rp1=1,lp2=0,rp2=0,lp3=0,rp3=0,v=0","support","1","yes","SYCL" -"SYCL0","PAD","type=f32,ne_a=[11,22,33,44],lp0=1,rp0=2,lp1=3,rp1=4,lp2=5,rp2=6,lp3=7,rp3=8,v=0","support","1","yes","SYCL" -"SYCL0","PAD","type=f32,ne_a=[512,512,1,1],lp0=0,rp0=1,lp1=0,rp1=1,lp2=0,rp2=0,lp3=0,rp3=0,v=1","support","0","no","SYCL" -"SYCL0","PAD","type=f32,ne_a=[11,22,33,44],lp0=1,rp0=2,lp1=3,rp1=4,lp2=5,rp2=6,lp3=7,rp3=8,v=1","support","0","no","SYCL" +"SYCL0","SOLVE_TRI","type=f32,ne_lhs=[128,128,4,4],ne_rhs=[32,128,4,4]","support","0","no","SYCL" +"SYCL0","SOLVE_TRI","type=f32,ne_lhs=[128,128,3,4],ne_rhs=[32,128,3,4]","support","0","no","SYCL" +"SYCL0","SOLVE_TRI","type=f32,ne_lhs=[128,128,4,1],ne_rhs=[32,128,4,1]","support","0","no","SYCL" +"SYCL0","SOLVE_TRI","type=f32,ne_lhs=[64,64,4,4],ne_rhs=[200,64,4,4]","support","0","no","SYCL" +"SYCL0","SOLVE_TRI","type=f32,ne_lhs=[64,64,4,4],ne_rhs=[384,64,4,4]","support","0","no","SYCL" +"SYCL0","PAD","type=f32,ne_a=[512,512,1,1],lp0=0,rp0=1,lp1=0,rp1=1,lp2=0,rp2=0,lp3=0,rp3=0,v=0,circular=0","support","1","yes","SYCL" +"SYCL0","PAD","type=f32,ne_a=[11,22,33,44],lp0=1,rp0=2,lp1=3,rp1=4,lp2=5,rp2=6,lp3=7,rp3=8,v=0,circular=0","support","1","yes","SYCL" +"SYCL0","PAD","type=f32,ne_a=[512,512,1,1],lp0=0,rp0=1,lp1=0,rp1=1,lp2=0,rp2=0,lp3=0,rp3=0,v=0,circular=1","support","0","no","SYCL" +"SYCL0","PAD","type=f32,ne_a=[11,22,33,44],lp0=1,rp0=2,lp1=3,rp1=4,lp2=5,rp2=6,lp3=7,rp3=8,v=0,circular=1","support","0","no","SYCL" +"SYCL0","PAD","type=f32,ne_a=[512,512,1,1],lp0=0,rp0=1,lp1=0,rp1=1,lp2=0,rp2=0,lp3=0,rp3=0,v=1,circular=0","support","0","no","SYCL" +"SYCL0","PAD","type=f32,ne_a=[11,22,33,44],lp0=1,rp0=2,lp1=3,rp1=4,lp2=5,rp2=6,lp3=7,rp3=8,v=1,circular=0","support","0","no","SYCL" +"SYCL0","PAD","type=f32,ne_a=[512,512,1,1],lp0=0,rp0=1,lp1=0,rp1=1,lp2=0,rp2=0,lp3=0,rp3=0,v=1,circular=1","support","0","no","SYCL" +"SYCL0","PAD","type=f32,ne_a=[11,22,33,44],lp0=1,rp0=2,lp1=3,rp1=4,lp2=5,rp2=6,lp3=7,rp3=8,v=1,circular=1","support","0","no","SYCL" "SYCL0","FLASH_ATTN_EXT","hsk=40,hsv=40,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" "SYCL0","FLASH_ATTN_EXT","hsk=40,hsv=40,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" "SYCL0","FLASH_ATTN_EXT","hsk=40,hsv=40,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" @@ -14097,86 +14255,86 @@ "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL" "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL" "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" @@ -14337,46 +14495,46 @@ "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" @@ -14537,46 +14695,46 @@ "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" @@ -14737,46 +14895,46 @@ "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" @@ -15017,86 +15175,86 @@ "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL" "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL" "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" @@ -15257,46 +15415,46 @@ "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" @@ -15457,46 +15615,46 @@ "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" @@ -15657,46 +15815,46 @@ "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" @@ -15857,46 +16015,46 @@ "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" @@ -16057,46 +16215,46 @@ "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" @@ -16257,46 +16415,46 @@ "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" @@ -16457,46 +16615,46 @@ "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" -"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL" +"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL" "SYCL0","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL" "SYCL0","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL" "SYCL0","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL" diff --git a/docs/speculative.md b/docs/speculative.md new file mode 100644 index 0000000000..8281eaa2d3 --- /dev/null +++ b/docs/speculative.md @@ -0,0 +1,120 @@ +# Speculative Decoding + +llama.cpp supports speculative decoding, a technique that can significantly accelerate token generation by predicting multiple tokens ahead of the main model. + +[Speculative decoding](https://en.wikipedia.org/wiki/Transformer_(deep_learning)#Speculative_decoding) leverages the fact that computing n tokens in a batch (as in prompt processing) is more efficient than computing n sequentially (as in response generation). By generating draft tokens quickly and then verifying them with the target model in a single batch, this approach can achieve substantial speedups when the draft predictions are frequently correct. + +## Implementations + +The `llama-server` application supports several implementations of speculative decoding: + +### Draft Model (`draft`) + +A much smaller model (called the _draft model_) generates drafts. +A draft model is the most used approach in speculative decoding. + +### n-gram Cache (`ngram-cache`) + +An n-gram is a sequence of n tokens. The n-gram cache implementation maintains statistics about short n-gram sequences. +A draft is computed using probabilities derived from these statistics. External statistics can also be loaded from files for improved accuracy. + +See: + +- #5479, #6828, #6848 + +### n-gram Map (`ngram-simple`, `ngram-map-*`) + +These implementations search the token history for patterns and use matching sequences as draft candidates. +They require no additional model but rely on patterns that have already appeared in the generated text. +An example to use this approach can be the rewriting of source code by a LLM. + +#### n-gram Map (`ngram-simple`) + +This implementation looks for the last n-gram in history that matches the current n-gram and creates a draft using the m tokens following the matched n-gram. It is the simplest self-speculative approach with minimal overhead. + +#### n-gram Map Key (`ngram-map-k`) + +This implementation looks for the current n-gram of size n (called the _key_) in the token history. If the key n-gram is followed by the same m tokens (called the _mgram_) multiple times, it creates a draft using these m tokens. This approach requires a minimum number of occurrences (argument `--spec-ngram-min-hits`) before generating drafts. + +The number of accepted tokens is stored for each used n-gram. + +#### n-gram Map Key-4-Values (`ngram-map-k4v`) + +This experimental implementation looks for the current n-gram of size n (called the _key_) in the token history. For each key, up to four _values_ (n-grams of size m, called _mgrams_) are tracked. An internal statistic counts the occurrences of each mgram after the key n-gram. If one mgram is significantly more frequent than the others, it is used as the draft. + +The number of accepted tokens is stored for each used n-gram. + +**Example:** Server options to be used if there are a lot of longer repetitions. +```bash +llama-server [...] --spec-type ngram-map-k4v --spec-ngram-size-n 8 --spec-ngram-size-m 8 --spec-ngram-min-hits 2 +``` + + +## Command-Line Options + +If a draft model is combined with a draftless decoding the draftless decoding has higher precedence. + +``` +--spec-type [none|ngram-cache|ngram-simple|ngram-map-k|ngram-map-k4v] + type of speculative decoding to use when no draft model is provided + (default: none) +--spec-ngram-size-n N ngram size N for ngram-simple/ngram-map speculative decoding, length + of lookup n-gram (default: 12) +--spec-ngram-size-m N ngram size M for ngram-simple/ngram-map speculative decoding, length + of draft m-gram (default: 48) +--spec-ngram-check-rate N ngram check rate for ngram-simple/ngram-map speculative decoding + (default: 1) +--spec-ngram-min-hits N minimum hits for ngram-map speculative decoding (default: 1) +``` + +### `--spec-type TYPE` + +Specifies a type of speculative decoding without draft model. + +| Type | Description | +|------|-------------| +| `none` | No speculative decoding (default) | +| `ngram-cache` | Use n-gram cache lookup | +| `ngram-simple` | Use simple n-gram pattern matching | +| `ngram-map-k` | Use n-gram pattern matching with n-gram-keys | +| `ngram-map-k4v` | Use n-gram pattern matching with n-gram-keys and up to four m-gram values (experimental) | + +**Example:** Server-instance used to refactor source code. +```bash +./llama-server [...] --spec-type ngram-simple +``` + +### `--spec-ngram-size-n N` + +Sets the size N of the lookup n-gram for n-gram map based speculative decoding. +The n-gram size N determines how many tokens in a row to look back when searching for matching patterns. + +### `--spec-ngram-size-m M` + +Sets the size M of the draft m-gram for n-gram map based speculative decoding. +The m-gram size determines how many tokens to draft when a match is found. +Larger values can provide more speedup but may reduce acceptance rate. + +### `--spec-ngram-check-rate R` + +This option aims at performance if the n-gram lookup in history is to costly. A lookup will be executed at every R tokens (default is 1, every token). + +### `--spec-ngram-min-hits H` + +This option defines how often a key has to appear in the token history to be used as a draft (default is 1). + +## Statistics +Each speculative decoding implementation prints statistics. + +``` +draft acceptance rate = 0.57576 ( 171 accepted / 297 generated) +statistics ngram_simple: #calls = 15, #gen drafts = 5, #acc drafts = 5, #gen tokens = 187, #acc tokens = 73 +statistics draft: #calls = 10, #gen drafts = 10, #acc drafts = 10, #gen tokens = 110, #acc tokens = 98 +``` + +- `#calls`: number of calls of this implementations +- `#gen drafts`: number of drafts generated by this implementation +- `#acc drafts`: number of drafts accepted (partially) by the main model +- `#gen tokens`: number of tokens generated by this implementation (including rejected tokens) +- `#acc tokens`: number of tokens accepted by the main model + diff --git a/examples/lookup/lookup-create.cpp b/examples/lookup/lookup-create.cpp index bb94a8fe06..f7b6ea1b19 100644 --- a/examples/lookup/lookup-create.cpp +++ b/examples/lookup/lookup-create.cpp @@ -32,9 +32,9 @@ int main(int argc, char ** argv){ common_ngram_cache ngram_cache; common_ngram_cache_update(ngram_cache, LLAMA_NGRAM_STATIC, LLAMA_NGRAM_STATIC, inp, inp.size(), true); - fprintf(stderr, "%s: hashing done, writing file to %s\n", __func__, params.lookup_cache_static.c_str()); + fprintf(stderr, "%s: hashing done, writing file to %s\n", __func__, params.speculative.lookup_cache_static.c_str()); - common_ngram_cache_save(ngram_cache, params.lookup_cache_static); + common_ngram_cache_save(ngram_cache, params.speculative.lookup_cache_static); return 0; } diff --git a/examples/lookup/lookup-stats.cpp b/examples/lookup/lookup-stats.cpp index 135f6fcab9..ae28b2e6e8 100644 --- a/examples/lookup/lookup-stats.cpp +++ b/examples/lookup/lookup-stats.cpp @@ -46,18 +46,18 @@ int main(int argc, char ** argv){ { const int64_t t_start_draft_us = ggml_time_us(); - if (!params.lookup_cache_static.empty()) { + if (!params.speculative.lookup_cache_static.empty()) { try { - ngram_cache_static = common_ngram_cache_load(params.lookup_cache_static); + ngram_cache_static = common_ngram_cache_load(params.speculative.lookup_cache_static); } catch (std::ifstream::failure const &) { - LOG_ERR("failed to open static lookup cache: %s", params.lookup_cache_static.c_str()); + LOG_ERR("failed to open static lookup cache: %s", params.speculative.lookup_cache_static.c_str()); exit(1); } } - if (!params.lookup_cache_dynamic.empty()) { + if (!params.speculative.lookup_cache_dynamic.empty()) { try { - ngram_cache_dynamic = common_ngram_cache_load(params.lookup_cache_dynamic); + ngram_cache_dynamic = common_ngram_cache_load(params.speculative.lookup_cache_dynamic); } catch (std::ifstream::failure const &) {} // if the file does not exist it will simply be created at the end of the program } diff --git a/examples/lookup/lookup.cpp b/examples/lookup/lookup.cpp index 27f159940a..8e73138a5f 100644 --- a/examples/lookup/lookup.cpp +++ b/examples/lookup/lookup.cpp @@ -51,18 +51,18 @@ int main(int argc, char ** argv){ const int64_t t_start_draft_us = ggml_time_us(); common_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, inp, inp.size(), false); - if (!params.lookup_cache_static.empty()) { + if (!params.speculative.lookup_cache_static.empty()) { try { - ngram_cache_static = common_ngram_cache_load(params.lookup_cache_static); + ngram_cache_static = common_ngram_cache_load(params.speculative.lookup_cache_static); } catch (std::ifstream::failure const &) { - LOG_ERR("failed to open static lookup cache: %s", params.lookup_cache_static.c_str()); + LOG_ERR("failed to open static lookup cache: %s", params.speculative.lookup_cache_static.c_str()); exit(1); } } - if (!params.lookup_cache_dynamic.empty()) { + if (!params.speculative.lookup_cache_dynamic.empty()) { try { - ngram_cache_dynamic = common_ngram_cache_load(params.lookup_cache_dynamic); + ngram_cache_dynamic = common_ngram_cache_load(params.speculative.lookup_cache_dynamic); } catch (std::ifstream::failure const &) {} // if the file does not exist it will simply be created at the end of the program } @@ -210,7 +210,7 @@ int main(int argc, char ** argv){ // Update dynamic ngram cache with context ngram cache and save it to disk: common_ngram_cache_merge(ngram_cache_dynamic, ngram_cache_context); - common_ngram_cache_save(ngram_cache_dynamic, params.lookup_cache_dynamic); + common_ngram_cache_save(ngram_cache_dynamic, params.speculative.lookup_cache_dynamic); LOG("\n\n"); diff --git a/examples/speculative-simple/speculative-simple.cpp b/examples/speculative-simple/speculative-simple.cpp index 8141052a22..d8b1f5a480 100644 --- a/examples/speculative-simple/speculative-simple.cpp +++ b/examples/speculative-simple/speculative-simple.cpp @@ -24,7 +24,7 @@ int main(int argc, char ** argv) { common_init(); - if (params.speculative.model.path.empty()) { + if (params.speculative.mparams_dft.path.empty()) { LOG_ERR("%s: --model-draft is required\n", __func__); return 1; } @@ -34,10 +34,8 @@ int main(int argc, char ** argv) { llama_numa_init(params.numa); llama_model * model_tgt = NULL; - //llama_model * model_dft = NULL; llama_context * ctx_tgt = NULL; - llama_context * ctx_dft = NULL; // load the target model auto llama_init_tgt = common_init_from_params(params); @@ -48,26 +46,38 @@ int main(int argc, char ** argv) { const llama_vocab * vocab = llama_model_get_vocab(model_tgt); // load the draft model - params.devices = params.speculative.devices; - params.model = params.speculative.model; - params.n_ctx = params.speculative.n_ctx; - params.n_batch = params.speculative.n_ctx > 0 ? params.speculative.n_ctx : params.n_batch; - params.n_gpu_layers = params.speculative.n_gpu_layers; + llama_model_ptr model_dft; - if (params.speculative.cpuparams.n_threads > 0) { - params.cpuparams.n_threads = params.speculative.cpuparams.n_threads; - } + // TODO: simplify this logic + { + const auto & params_spec = params.speculative; - params.cpuparams_batch.n_threads = params.speculative.cpuparams_batch.n_threads; - params.tensor_buft_overrides = params.speculative.tensor_buft_overrides; + auto params_dft = params; - auto llama_init_dft = common_init_from_params(params); + params_dft.n_parallel = 1; + params_dft.n_ctx = params_spec.n_ctx; + params_dft.n_batch = llama_n_ctx_seq(ctx_tgt); + params_dft.devices = params_spec.devices; + params_dft.model = params_spec.mparams_dft; + params_dft.n_gpu_layers = params_spec.n_gpu_layers; - //model_dft = llama_init_dft->model(); - ctx_dft = llama_init_dft->context(); + if (params_spec.cpuparams.n_threads > 0) { + params_dft.cpuparams.n_threads = params.speculative.cpuparams.n_threads; + params_dft.cpuparams_batch.n_threads = params.speculative.cpuparams_batch.n_threads; + } - if (!common_speculative_are_compatible(ctx_tgt, ctx_dft)) { - LOG_INF("the draft model '%s' is not compatible with the target model '%s'. tokens will be translated between the draft and target models.\n", params.speculative.model.path.c_str(), params.model.path.c_str()); + params_dft.tensor_buft_overrides = params.speculative.tensor_buft_overrides; + + auto mparams_dft = common_model_params_to_llama(params_dft); + + model_dft.reset(llama_model_load_from_file(params_dft.model.path.c_str(), mparams_dft)); + if (model_dft == nullptr) { + LOG_ERR("failed to load draft model, '%s'\n", params_dft.model.path.c_str()); + return 1; + } + + params.speculative.model_dft = model_dft.get(); + params.speculative.cparams_dft = common_context_params_to_llama(params_dft); } // Tokenize the prompt @@ -92,12 +102,6 @@ int main(int argc, char ** argv) { LOG("%s", common_token_to_piece(ctx_tgt, id).c_str()); } - // how many tokens to draft each time - int n_draft = params.speculative.n_max; - int n_draft_min = params.speculative.n_min; - - float p_min = params.speculative.p_min; - int n_predict = 0; int n_drafted = 0; int n_accept = 0; @@ -127,15 +131,11 @@ int main(int argc, char ** argv) { int n_past = inp.size() - 1; // init the speculator - struct common_speculative_params params_spec; - params_spec.n_draft = n_draft; - params_spec.n_reuse = llama_n_ctx(ctx_dft) - n_draft; - params_spec.p_min = p_min; + const auto & params_spec = params.speculative; - struct common_speculative * spec = common_speculative_init(ctx_tgt, ctx_dft); - for (auto &pair : params.speculative.replacements) { - common_speculative_add_replacement_tgt_dft(spec, pair.first.c_str(), pair.second.c_str()); - } + struct common_speculative * spec = common_speculative_init(params.speculative, ctx_tgt); + + common_speculative_begin(spec, prompt_tgt); llama_batch batch_tgt = llama_batch_init(llama_n_batch(ctx_tgt), 0, 1); @@ -151,7 +151,7 @@ int main(int argc, char ** argv) { // offloaded to a remote device. it doesn't even have to be based on an LLM. instead, it can provide tokens // from a cache or lookup tables. // - llama_tokens draft = common_speculative_gen_draft(spec, params_spec, prompt_tgt, id_last); + llama_tokens draft = common_speculative_draft(spec, params_spec, prompt_tgt, id_last); //LOG_DBG("draft: %s\n", string_from(ctx_dft, draft).c_str()); @@ -162,7 +162,7 @@ int main(int argc, char ** argv) { // evaluate the target model on [id_last, draft0, draft1, ..., draftN-1] { // do not waste time on small drafts - if (draft.size() < (size_t) n_draft_min) { + if (draft.size() < (size_t) params_spec.n_min) { draft.clear(); } @@ -240,7 +240,7 @@ int main(int argc, char ** argv) { LOG_INF("decoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_predict, (t_dec_end - t_dec_start) / 1e6f, n_predict / ((t_dec_end - t_dec_start) / 1e6f)); LOG_INF("\n"); - LOG_INF("n_draft = %d\n", n_draft); + LOG_INF("n_draft = %d\n", params_spec.n_max); LOG_INF("n_predict = %d\n", n_predict); LOG_INF("n_drafted = %d\n", n_drafted); LOG_INF("n_accept = %d\n", n_accept); @@ -249,8 +249,6 @@ int main(int argc, char ** argv) { LOG_INF("\n"); LOG_INF("draft:\n\n"); - llama_perf_context_print(ctx_dft); - LOG_INF("\n"); LOG_INF("target:\n\n"); common_perf_print(ctx_tgt, smpl); diff --git a/examples/speculative/speculative.cpp b/examples/speculative/speculative.cpp index 89d3249431..3e5cf5f46b 100644 --- a/examples/speculative/speculative.cpp +++ b/examples/speculative/speculative.cpp @@ -46,7 +46,7 @@ int main(int argc, char ** argv) { common_init(); - if (params.speculative.model.path.empty()) { + if (params.speculative.mparams_dft.path.empty()) { LOG_ERR("%s: --model-draft is required\n", __func__); return 1; } @@ -78,7 +78,7 @@ int main(int argc, char ** argv) { // load the draft model params.devices = params.speculative.devices; - params.model = params.speculative.model; + params.model = params.speculative.mparams_dft; params.n_gpu_layers = params.speculative.n_gpu_layers; if (params.speculative.cpuparams.n_threads > 0) { params.cpuparams.n_threads = params.speculative.cpuparams.n_threads; diff --git a/ggml/CMakeLists.txt b/ggml/CMakeLists.txt index 0176ca1ce9..b0b8e57898 100644 --- a/ggml/CMakeLists.txt +++ b/ggml/CMakeLists.txt @@ -228,6 +228,8 @@ option(GGML_WEBGPU_CPU_PROFILE "ggml: enable WebGPU profiling (CPU) option(GGML_WEBGPU_GPU_PROFILE "ggml: enable WebGPU profiling (GPU)" OFF) option(GGML_WEBGPU_JSPI "ggml: use JSPI for WebGPU" ON) option(GGML_ZDNN "ggml: use zDNN" OFF) +option(GGML_VIRTGPU "ggml: use the VirtGPU/Virglrenderer API Remoting frontend" OFF) +option(GGML_VIRTGPU_BACKEND "ggml: build the VirtGPU/Virglrenderer API Remoting backend" OFF) option(GGML_METAL "ggml: use Metal" ${GGML_METAL_DEFAULT}) option(GGML_METAL_NDEBUG "ggml: disable Metal debugging" OFF) option(GGML_METAL_SHADER_DEBUG "ggml: compile Metal with -fno-fast-math" OFF) @@ -320,6 +322,7 @@ set(GGML_PUBLIC_HEADERS include/ggml-opt.h include/ggml-metal.h include/ggml-rpc.h + include/ggml-virtgpu.h include/ggml-sycl.h include/ggml-vulkan.h include/ggml-webgpu.h diff --git a/ggml/include/ggml-virtgpu.h b/ggml/include/ggml-virtgpu.h new file mode 100644 index 0000000000..1cb4bd7a03 --- /dev/null +++ b/ggml/include/ggml-virtgpu.h @@ -0,0 +1,16 @@ +#pragma once + +#include "ggml.h" +#include "ggml-backend.h" + +#ifdef __cplusplus +extern "C" { +#endif + +#define GGML_REMOTING_FRONTEND_NAME "RemotingFrontend" + +GGML_BACKEND_API ggml_backend_reg_t ggml_backend_virtgpu_reg(); + +#ifdef __cplusplus +} +#endif diff --git a/ggml/src/CMakeLists.txt b/ggml/src/CMakeLists.txt index 6192a87046..260ad48f0e 100644 --- a/ggml/src/CMakeLists.txt +++ b/ggml/src/CMakeLists.txt @@ -451,6 +451,7 @@ ggml_add_backend(HIP) ggml_add_backend(METAL) ggml_add_backend(MUSA) ggml_add_backend(RPC) +ggml_add_backend(VirtGPU) ggml_add_backend(SYCL) ggml_add_backend(Vulkan) ggml_add_backend(WebGPU) diff --git a/ggml/src/ggml-backend-reg.cpp b/ggml/src/ggml-backend-reg.cpp index 6bee1bc4b4..dd991f262e 100644 --- a/ggml/src/ggml-backend-reg.cpp +++ b/ggml/src/ggml-backend-reg.cpp @@ -69,6 +69,10 @@ #include "ggml-rpc.h" #endif +#ifdef GGML_USE_VIRTGPU_FRONTEND +#include "ggml-virtgpu.h" +#endif + #ifdef GGML_USE_CANN #include "ggml-cann.h" #endif @@ -180,7 +184,12 @@ struct ggml_backend_registry { register_backend(ggml_backend_sycl_reg()); #endif #ifdef GGML_USE_VULKAN + // Add runtime disable check + if (getenv("GGML_DISABLE_VULKAN") == nullptr) { register_backend(ggml_backend_vk_reg()); + } else { + GGML_LOG_DEBUG("Vulkan backend disabled by GGML_DISABLE_VULKAN environment variable\n"); + } #endif #ifdef GGML_USE_WEBGPU register_backend(ggml_backend_webgpu_reg()); @@ -188,6 +197,10 @@ struct ggml_backend_registry { #ifdef GGML_USE_ZDNN register_backend(ggml_backend_zdnn_reg()); #endif +#ifdef GGML_USE_VIRTGPU_FRONTEND + register_backend(ggml_backend_virtgpu_reg()); +#endif + #ifdef GGML_USE_OPENCL register_backend(ggml_backend_opencl_reg()); #endif @@ -604,6 +617,7 @@ void ggml_backend_load_all_from_path(const char * dir_path) { ggml_backend_load_best("rpc", silent, dir_path); ggml_backend_load_best("sycl", silent, dir_path); ggml_backend_load_best("vulkan", silent, dir_path); + ggml_backend_load_best("virtgpu", silent, dir_path); ggml_backend_load_best("opencl", silent, dir_path); ggml_backend_load_best("hexagon", silent, dir_path); ggml_backend_load_best("musa", silent, dir_path); diff --git a/ggml/src/ggml-cpu/arch-fallback.h b/ggml/src/ggml-cpu/arch-fallback.h index 0a85a4cff3..427c1146e4 100644 --- a/ggml/src/ggml-cpu/arch-fallback.h +++ b/ggml/src/ggml-cpu/arch-fallback.h @@ -1,3 +1,4 @@ + #pragma once // Rename `_generic` functions if no native implementation is available. @@ -42,6 +43,7 @@ #define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K #define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K +#define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0 #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0 #define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0 @@ -53,6 +55,7 @@ #define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K #define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K +# define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0 #define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0 #define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0 @@ -73,6 +76,7 @@ #define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0 #define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K #define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K +#define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0 #define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0 #define ggml_gemv_q8_0_4x8_q8_0_generic ggml_gemv_q8_0_4x8_q8_0 @@ -80,6 +84,7 @@ #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0 #define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K #define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K +#define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0 #define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0 #define ggml_gemm_q8_0_4x8_q8_0_generic ggml_gemm_q8_0_4x8_q8_0 @@ -102,6 +107,7 @@ #define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K #define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K +#define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0 #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0 #define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0 @@ -113,6 +119,7 @@ #define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K #define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K +#define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0 #define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0 #define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0 @@ -136,6 +143,7 @@ #define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K #define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K +#define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0 #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0 #define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0 @@ -147,6 +155,7 @@ #define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K #define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K +#define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0 #define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0 #define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0 @@ -177,6 +186,7 @@ #define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K #define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K +#define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0 #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0 #define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0 @@ -187,6 +197,7 @@ #define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K #define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K +#define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0 #define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0 #define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0 @@ -216,6 +227,7 @@ #define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K #define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K +#define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0 #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0 #define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0 @@ -227,6 +239,7 @@ #define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K #define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K +#define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0 #define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0 #define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0 @@ -258,6 +271,7 @@ #define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K #define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K +#define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0 #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0 #define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0 @@ -269,6 +283,7 @@ #define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K #define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K +#define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0 #define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0 #define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0 diff --git a/ggml/src/ggml-cpu/arch/arm/repack.cpp b/ggml/src/ggml-cpu/arch/arm/repack.cpp index 883d862901..99bb70274c 100644 --- a/ggml/src/ggml-cpu/arch/arm/repack.cpp +++ b/ggml/src/ggml-cpu/arch/arm/repack.cpp @@ -1055,10 +1055,10 @@ void ggml_gemv_q5_K_8x8_q8_K(int n, // FUSED BIAS: Compute and subtract bias immediately // bias = (bsums_lo * mins_lo + bsums_hi * mins_hi) * sb_min - int32x4_t bias = vmull_s16(bsums_vec_lo, group_mins_lo); - bias = vmlal_s16(bias, bsums_vec_hi, group_mins_hi); + int32x4_t bias = vmull_s16(bsums_vec_lo, group_mins_lo); + bias = vmlal_s16(bias, bsums_vec_hi, group_mins_hi); float32x4_t bias_f32 = vcvtq_f32_s32(bias); - acc_f32[i] = vmlsq_f32(acc_f32[i], sb_min, bias_f32); + acc_f32[i] = vmlsq_f32(acc_f32[i], sb_min, bias_f32); } } // for sb } // for b @@ -1072,6 +1072,208 @@ void ggml_gemv_q5_K_8x8_q8_K(int n, ggml_gemv_q5_K_8x8_q8_K_generic(n, s, bs, vx, vy, nr, nc); } +void ggml_gemv_q6_K_8x8_q8_K(int n, + float * GGML_RESTRICT s, + size_t bs, + const void * GGML_RESTRICT vx, + const void * GGML_RESTRICT vy, + int nr, + int nc) { + constexpr int qk = QK_K; + const int nb = n / qk; + + constexpr int ncols_interleaved = 8; + constexpr int blocklen = 8; + + assert(n % qk == 0); + assert(nc % ncols_interleaved == 0); + + UNUSED(nb); + UNUSED(ncols_interleaved); + UNUSED(blocklen); + +#if defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD) + constexpr int col_pairs = ncols_interleaved / 2; + const uint8x16_t m4b = vdupq_n_u8(0x0f); + const uint8x16_t mask_lo = vdupq_n_u8(0x03); + const uint8x16_t mask_hi = vdupq_n_u8(0x30); + + // 1x8 tile = 2 x 4 + float32x4_t acc_f32[2]; + + const block_q8_K * GGML_RESTRICT q8_ptr = (const block_q8_K *) vy; + + for (int x = 0; x < nc / ncols_interleaved; x++) { + const block_q6_Kx8 * GGML_RESTRICT q6_ptr = (const block_q6_Kx8 *) vx + (x * nb); + + acc_f32[0] = vdupq_n_f32(0); + acc_f32[1] = vdupq_n_f32(0); + + for (int b = 0; b < nb; b++) { + float32x4_t q6_d_0 = vcvt_f32_f16(vld1_f16((const __fp16 *) q6_ptr[b].d)); // d0 d1 d2 d3 + float32x4_t q6_d_1 = vcvt_f32_f16(vld1_f16((const __fp16 *) q6_ptr[b].d + 4)); // d4 d5 d6 d7 + float32x4_t q8_d = vdupq_n_f32(q8_ptr[b].d); + float32x4_t sb_scale_0 = vmulq_f32(q6_d_0, q8_d); + float32x4_t sb_scale_1 = vmulq_f32(q6_d_1, q8_d); + + int32x2_t acc[col_pairs]; + for (int i = 0; i < col_pairs; i++) { + acc[i] = vdup_n_s32(0); + } + + // Load all 16 scales once and widen to int16 (Q6_K has 16 scales per block) + // Reused for bias and dequantization later + int16_t q6_scales[16 * 8]; + for (int i = 0; i < 16; i++) { + int16x8_t scales = vmovl_s8(vld1_s8(q6_ptr[b].scales + i * 8)); + vst1q_s16(q6_scales + i * 8, scales); + } + + // Compute bias per column using q8 bsums and preloaded scales to skip the -32 shift + int32x4_t bias_lo = vdupq_n_s32(0); + int32x4_t bias_hi = vdupq_n_s32(0); + + // Load bsums in chunks of 4 to process with vectorized operations + for (int i = 0; i < 16; i += 4) { + int16x4_t bsums_vec = vld1_s16(q8_ptr[b].bsums + i); + int16x4_t scales_lo_0 = vld1_s16(q6_scales + (i + 0) * 8); + int16x4_t scales_hi_0 = vld1_s16(q6_scales + (i + 0) * 8 + 4); + int16x4_t scales_lo_1 = vld1_s16(q6_scales + (i + 1) * 8); + int16x4_t scales_hi_1 = vld1_s16(q6_scales + (i + 1) * 8 + 4); + int16x4_t scales_lo_2 = vld1_s16(q6_scales + (i + 2) * 8); + int16x4_t scales_hi_2 = vld1_s16(q6_scales + (i + 2) * 8 + 4); + int16x4_t scales_lo_3 = vld1_s16(q6_scales + (i + 3) * 8); + int16x4_t scales_hi_3 = vld1_s16(q6_scales + (i + 3) * 8 + 4); + + bias_lo = vmlal_lane_s16(bias_lo, scales_lo_0, bsums_vec, 0); + bias_hi = vmlal_lane_s16(bias_hi, scales_hi_0, bsums_vec, 0); + bias_lo = vmlal_lane_s16(bias_lo, scales_lo_1, bsums_vec, 1); + bias_hi = vmlal_lane_s16(bias_hi, scales_hi_1, bsums_vec, 1); + bias_lo = vmlal_lane_s16(bias_lo, scales_lo_2, bsums_vec, 2); + bias_hi = vmlal_lane_s16(bias_hi, scales_hi_2, bsums_vec, 2); + bias_lo = vmlal_lane_s16(bias_lo, scales_lo_3, bsums_vec, 3); + bias_hi = vmlal_lane_s16(bias_hi, scales_hi_3, bsums_vec, 3); + } + bias_lo = vshlq_n_s32(bias_lo, 5); + bias_hi = vshlq_n_s32(bias_hi, 5); + + // Process two 128-value halves per superblock + for (int half = 0; half < 2; half++) { + const uint8_t * ql_base = q6_ptr[b].ql + half * 512; + const uint8_t * qh_base = q6_ptr[b].qh + half * 256; + + // A subblock (sb) is a set of weights that share the scale + // Since q6_K scales are per 16 elements + // num sbs -> 256 elements / (16 elements/scale * 2 elements/byte * 2 halves) + for (int sb = 0; sb < QK_K / 64; sb++) { + const int8_t * q8_base_l = q8_ptr[b].qs + half * 128 + sb * 16; + const int8_t * q8_base_h = q8_base_l + 64; + + // Load and duplicate q8 values (each register covers two interleaved columns of q6) + int8x16_t q8_l[2]; + int8x16_t q8_h[2]; + for (int i = 0; i < 2; i++) { + q8_l[i] = (int8x16_t) vld1q_dup_s64((const int64_t *) (q8_base_l + i * 8)); + q8_h[i] = (int8x16_t) vld1q_dup_s64((const int64_t *) (q8_base_h + i * 8)); + } + + // TODO: Test other qh repack patterns to reduce loads + const int ql_off_base = sb * QK_K / 2; + const int qh_off_base = ql_off_base & 255; // wraps after 256 bytes + + // Load 4 vectors at once (64 bytes each for ql_0, ql_1, qh_0, qh_1) + ggml_uint8x16x4_t q6_ql_0 = ggml_vld1q_u8_x4(ql_base + ql_off_base); + ggml_uint8x16x4_t q6_ql_1 = ggml_vld1q_u8_x4(ql_base + ql_off_base + 64); + ggml_uint8x16x4_t q6_qh_0 = ggml_vld1q_u8_x4(qh_base + qh_off_base); + ggml_uint8x16x4_t q6_qh_1 = ggml_vld1q_u8_x4(qh_base + qh_off_base + 64); + + // Adjust qh for subblocks 2 and 3 (shift right by 2) + if (sb > 1) { + q6_qh_0.val[0] = vshrq_n_u8(q6_qh_0.val[0], 2); + q6_qh_0.val[1] = vshrq_n_u8(q6_qh_0.val[1], 2); + q6_qh_0.val[2] = vshrq_n_u8(q6_qh_0.val[2], 2); + q6_qh_0.val[3] = vshrq_n_u8(q6_qh_0.val[3], 2); + q6_qh_1.val[0] = vshrq_n_u8(q6_qh_1.val[0], 2); + q6_qh_1.val[1] = vshrq_n_u8(q6_qh_1.val[1], 2); + q6_qh_1.val[2] = vshrq_n_u8(q6_qh_1.val[2], 2); + q6_qh_1.val[3] = vshrq_n_u8(q6_qh_1.val[3], 2); + } + + // Process column pairs (0-1, 2-3, 4-5, 6-7) + for (int cp = 0; cp < col_pairs; cp++) { + const uint8x16_t q6_qs_cp_0_l = q6_ql_0.val[cp]; + const uint8x16_t q6_qs_cp_1_l = q6_ql_1.val[cp]; + const uint8x16_t q6_qs_cp_0_h = q6_qh_0.val[cp]; + const uint8x16_t q6_qs_cp_1_h = q6_qh_1.val[cp]; + + // Extract high 2 bits for upper nibble reconstruction + const uint8x16_t q6_qs_cp_0_hh = vandq_u8(q6_qs_cp_0_h, mask_hi); + const uint8x16_t q6_qs_cp_1_hh = vandq_u8(q6_qs_cp_1_h, mask_hi); + + // q6 = (low4 | high2<<4), without -32 bias (handled via bsums) + const int8x16_t q6_l0 = vreinterpretq_s8_u8( + vsliq_n_u8(vandq_u8(q6_qs_cp_0_l, m4b), vandq_u8(q6_qs_cp_0_h, mask_lo), 4)); + const int8x16_t q6_l1 = vreinterpretq_s8_u8( + vsliq_n_u8(vandq_u8(q6_qs_cp_1_l, m4b), vandq_u8(q6_qs_cp_1_h, mask_lo), 4)); + const int8x16_t q6_h0 = + vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6_qs_cp_0_l, 4), q6_qs_cp_0_hh)); + const int8x16_t q6_h1 = + vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6_qs_cp_1_l, 4), q6_qs_cp_1_hh)); + + int32x4_t sb_acc_l = vdupq_n_s32(0); + sb_acc_l = vdotq_s32(sb_acc_l, q6_l0, q8_l[0]); + sb_acc_l = vdotq_s32(sb_acc_l, q6_l1, q8_l[1]); + + int32x4_t sb_acc_h = vdupq_n_s32(0); + sb_acc_h = vdotq_s32(sb_acc_h, q6_h0, q8_h[0]); + sb_acc_h = vdotq_s32(sb_acc_h, q6_h1, q8_h[1]); + + // Pairwise add to get per-column sums: [col0, col1] + int32x2_t sum_l = vpadd_s32(vget_low_s32(sb_acc_l), vget_high_s32(sb_acc_l)); + int32x2_t sum_h = vpadd_s32(vget_low_s32(sb_acc_h), vget_high_s32(sb_acc_h)); + + const int scale_idx_l = half * 8 + sb; + const int scale_idx_h = half * 8 + sb + 4; + + // Access scales using array indexing (scales are interleaved by column) + const int32x2_t scale_vec_l = { (int32_t) q6_scales[scale_idx_l * 8 + cp * 2], + (int32_t) q6_scales[scale_idx_l * 8 + cp * 2 + 1] }; + const int32x2_t scale_vec_h = { (int32_t) q6_scales[scale_idx_h * 8 + cp * 2], + (int32_t) q6_scales[scale_idx_h * 8 + cp * 2 + 1] }; + + // Accumulate scaled results + acc[cp] = vmla_s32(acc[cp], sum_l, scale_vec_l); + acc[cp] = vmla_s32(acc[cp], sum_h, scale_vec_h); + } + } + } // for half + + // Bias correction + acc[0] = vsub_s32(acc[0], vget_low_s32(bias_lo)); + acc[1] = vsub_s32(acc[1], vget_high_s32(bias_lo)); + acc[2] = vsub_s32(acc[2], vget_low_s32(bias_hi)); + acc[3] = vsub_s32(acc[3], vget_high_s32(bias_hi)); + + // Apply superblock scale (no mins for q6_K) + // acc[cp] has [c0, c1] + float32x2_t w_01 = vmul_f32(vcvt_f32_s32(acc[0]), vget_low_f32(sb_scale_0)); + float32x2_t w_23 = vmul_f32(vcvt_f32_s32(acc[1]), vget_high_f32(sb_scale_0)); + float32x2_t w_45 = vmul_f32(vcvt_f32_s32(acc[2]), vget_low_f32(sb_scale_1)); + float32x2_t w_67 = vmul_f32(vcvt_f32_s32(acc[3]), vget_high_f32(sb_scale_1)); + + acc_f32[0] = vaddq_f32(acc_f32[0], vcombine_f32(w_01, w_23)); + acc_f32[1] = vaddq_f32(acc_f32[1], vcombine_f32(w_45, w_67)); + } // for b + + int base = x * ncols_interleaved; + vst1q_f32(s + base, acc_f32[0]); + vst1q_f32(s + base + 4, acc_f32[1]); + } // for x + return; +#endif // defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD) + ggml_gemv_q6_K_8x8_q8_K_generic(n, s, bs, vx, vy, nr, nc); +} + void ggml_gemv_q8_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, @@ -2946,16 +3148,17 @@ void ggml_gemm_q4_K_8x8_q8_K(int n, // Scales[i] corresponds to column i const int scale_offset = cp * 2; - for (int blk = 0; blk < 2; blk++) { - const int32x4_t block_scale = { - (int32_t) q4sb_scales[blk][scale_offset], - (int32_t) q4sb_scales[blk][scale_offset], - (int32_t) q4sb_scales[blk][scale_offset + 1], - (int32_t) q4sb_scales[blk][scale_offset + 1], - }; - acc[cp] = vmlaq_s32(acc[cp], sb_acc[blk], block_scale); - acc[cp + 4] = vmlaq_s32(acc[cp + 4], sb_acc[blk + 2], block_scale); - } + const int32_t scale_00 = q4sb_scales[0][scale_offset]; + const int32_t scale_01 = q4sb_scales[0][scale_offset + 1]; + const int32_t scale_10 = q4sb_scales[1][scale_offset]; + const int32_t scale_11 = q4sb_scales[1][scale_offset + 1]; + const int32x4_t block_scale_0 = vcombine_s32(vdup_n_s32(scale_00), vdup_n_s32(scale_01)); + const int32x4_t block_scale_1 = vcombine_s32(vdup_n_s32(scale_10), vdup_n_s32(scale_11)); + + acc[cp] = vmlaq_s32(acc[cp], sb_acc[0], block_scale_0); + acc[cp + 4] = vmlaq_s32(acc[cp + 4], sb_acc[2], block_scale_0); + acc[cp] = vmlaq_s32(acc[cp], sb_acc[1], block_scale_1); + acc[cp + 4] = vmlaq_s32(acc[cp + 4], sb_acc[3], block_scale_1); } // Multiply Acc bsum + mins @@ -3146,8 +3349,8 @@ void ggml_gemm_q5_K_8x8_q8_K(int n, const int8x16_t qs_lo_0 = vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(qs_cp_0, m4b), hbit_lo_0, 4)); int32x4_t acc_0 = sb_acc[0]; acc_0 = vmmlaq_s32(acc_0, qs_lo_0, q8s[0][0]); - int32x4_t acc_2 = sb_acc[2]; - acc_2 = vmmlaq_s32(acc_2, qs_lo_0, q8s[1][0]); + int32x4_t acc_2 = sb_acc[2]; + acc_2 = vmmlaq_s32(acc_2, qs_lo_0, q8s[1][0]); const int8x16_t qs_hi_0 = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(qs_cp_0, 4), hbit_hi_0)); int32x4_t acc_1 = sb_acc[1]; acc_1 = vmmlaq_s32(acc_1, qs_hi_0, q8s[0][4]); @@ -3271,6 +3474,223 @@ void ggml_gemm_q5_K_8x8_q8_K(int n, ggml_gemm_q5_K_8x8_q8_K_generic(n, s, bs, vx, vy, nr, nc); } +void ggml_gemm_q6_K_8x8_q8_K(int n, + float * GGML_RESTRICT s, + size_t bs, + const void * GGML_RESTRICT vx, + const void * GGML_RESTRICT vy, + int nr, + int nc) { + constexpr int qk = QK_K; + const int nb = n / qk; + + constexpr int ncols_interleaved = 8; + constexpr int blocklen = 8; + + assert(n % qk == 0); + assert(nr % 4 == 0); + assert(nc % ncols_interleaved == 0); + + UNUSED(nb); + UNUSED(ncols_interleaved); + UNUSED(blocklen); + +#if defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8) + constexpr int q8_k_blocklen = 4; + const uint8x16_t m4b = vdupq_n_u8(0x0f); + const uint8x16_t mask_lo = vdupq_n_u8(0x03); + const uint8x16_t mask_hi = vdupq_n_u8(0x30); + const int8x16_t m32s = vdupq_n_s8(32); + + // 8 accumulators: 4 q8 rows × 2 col groups (0-3, 4-7) + float32x4_t acc_f32[blocklen]; + + for (int y = 0; y < nr / q8_k_blocklen; y++) { + const block_q8_Kx4 * GGML_RESTRICT q8_ptr = (const block_q8_Kx4 *) vy + (y * nb); + + for (int x = 0; x < nc / ncols_interleaved; x++) { + const block_q6_Kx8 * GGML_RESTRICT q6_ptr = (const block_q6_Kx8 *) vx + (x * nb); + + for (int i = 0; i < blocklen; i++) { + acc_f32[i] = vdupq_n_f32(0); + } + + for (int b = 0; b < nb; b++) { + int32x4_t acc[8]; // rows 01 stored in [0][1][2][3], rows 23 stored in [4][5][6][7] + for (int i = 0; i < 8; i++) { + acc[i] = vdupq_n_s32(0); + } + + // Q6_K has simple 8-bit scales, 16 per block (one per 16 values) + // Reused for bias and dequantization later + int16_t q6_scales[16 * 8]; + for (int i = 0; i < 16; ++i) { + int16x8_t s16 = vmovl_s8(vld1_s8(q6_ptr[b].scales + i * 8)); + vst1q_s16(q6_scales + i * 8, s16); + } + + // Process two 128-value halves per superblock + for (int half = 0; half < 2; half++) { + + const uint8_t * ql_base = q6_ptr[b].ql + half * 512; + const uint8_t * qh_base = q6_ptr[b].qh + half * 256; + + // A subblock (sb) is a set of weights that share the scale + // Since q6_K scales are per 16 elements + // num sbs -> 256 elements / (16 elements/scale * 2 elements/byte * 2 halves) + for (int sb = 0; sb < QK_K / 64; sb++) { + // Q6_K weight index increasing by 64 instead of 32 requires + // loading various q8 memory regions + const int8_t * q8_base_l = q8_ptr[b].qs + half * 512 + sb * 64; + const int8_t * q8_base_h = q8_ptr[b].qs + half * 512 + 256 + sb * 64; + + int8x16_t q8_l_01[2]; + int8x16_t q8_l_23[2]; + for (int i = 0; i < 2; i++) { + const int offset = i * 32; + q8_l_01[i] = vld1q_s8(q8_base_l + offset); // 0..7 & 8..15 (r01) + q8_l_23[i] = vld1q_s8(q8_base_l + offset + 16); // 0..7 & 8..15 (r23) + } + + int8x16_t q8_h_01[2]; + int8x16_t q8_h_23[2]; + for (int i = 0; i < 2; i++) { + const int offset = i * 32; + q8_h_01[i] = vld1q_s8(q8_base_h + offset); + q8_h_23[i] = vld1q_s8(q8_base_h + offset + 16); + } + + const int ql_off_base = sb * QK_K / 2; + + uint8x16_t q6_ql_0[4]; + uint8x16_t q6_ql_1[4]; + for (int k = 0; k < 4; k++) { + q6_ql_0[k] = vld1q_u8(ql_base + ql_off_base + 16 * k); + q6_ql_1[k] = vld1q_u8(ql_base + ql_off_base + 64 + 16 * k); + } + + const int qh_off_base = (sb * QK_K / 2) & 255; // wrap after 256 bytes + uint8x16_t q6_qh_0[4]; + uint8x16_t q6_qh_1[4]; + for (int k = 0; k < 4; k++) { + q6_qh_0[k] = vld1q_u8(qh_base + qh_off_base + 16 * k); + q6_qh_1[k] = vld1q_u8(qh_base + qh_off_base + 64 + 16 * k); + } + + // Adjust for the proper high bits (Sb 2 and 3) + if (sb > 1) { + for (int k = 0; k < 4; k++) { + q6_qh_0[k] = vshrq_n_u8(q6_qh_0[k], 2); + q6_qh_1[k] = vshrq_n_u8(q6_qh_1[k], 2); + } + } + + // Process column pairs (0-1, 2-3, 4-5, 6-7) + for (int cp = 0; cp < ncols_interleaved / 2; cp++) { + const uint8x16_t q6_qs_cp_0_l = q6_ql_0[cp]; + const uint8x16_t q6_qs_cp_1_l = q6_ql_1[cp]; + const uint8x16_t q6_qs_cp_0_h = q6_qh_0[cp]; + const uint8x16_t q6_qs_cp_1_h = q6_qh_1[cp]; + + // Extract high 2 bits for upper nibble reconstruction + const uint8x16_t q6_qs_cp_0_hh = vandq_u8(q6_qs_cp_0_h, mask_hi); + const uint8x16_t q6_qs_cp_1_hh = vandq_u8(q6_qs_cp_1_h, mask_hi); + + // q6 = (low4 | high2<<4) - 32 + // Use vsliq_n_u8 to combine shift-left-insert in one instruction (like Q5_K) + const int8x16_t q6_l0 = vsubq_s8( + vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(q6_qs_cp_0_l, m4b), vandq_u8(q6_qs_cp_0_h, mask_lo), 4)), + m32s); + const int8x16_t q6_l1 = vsubq_s8( + vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(q6_qs_cp_1_l, m4b), vandq_u8(q6_qs_cp_1_h, mask_lo), 4)), + m32s); + const int8x16_t q6_h0 = vsubq_s8( + vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6_qs_cp_0_l, 4), q6_qs_cp_0_hh)), m32s); + const int8x16_t q6_h1 = vsubq_s8( + vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6_qs_cp_1_l, 4), q6_qs_cp_1_hh)), m32s); + + // row pair 0, base_l + int32x4_t sb_acc_0l = vmmlaq_s32(vdupq_n_s32(0), q6_l0, q8_l_01[0]); + sb_acc_0l = vmmlaq_s32(sb_acc_0l, q6_l1, q8_l_01[1]); + // row pair 0, base_h + int32x4_t sb_acc_0h = vmmlaq_s32(vdupq_n_s32(0), q6_h0, q8_h_01[0]); + sb_acc_0h = vmmlaq_s32(sb_acc_0h, q6_h1, q8_h_01[1]); + // row pair 1, base_l + int32x4_t sb_acc_1l = vmmlaq_s32(vdupq_n_s32(0), q6_l0, q8_l_23[0]); + sb_acc_1l = vmmlaq_s32(sb_acc_1l, q6_l1, q8_l_23[1]); + // row pair 1, base_h + int32x4_t sb_acc_1h = vmmlaq_s32(vdupq_n_s32(0), q6_h0, q8_h_23[0]); + sb_acc_1h = vmmlaq_s32(sb_acc_1h, q6_h1, q8_h_23[1]); + + const int scale_idx_l = half * 8 + sb; + const int scale_idx_h = half * 8 + sb + 4; + + const int32x4_t scale_vec_l = { + q6_scales[scale_idx_l * 8 + cp * 2 + 0], + q6_scales[scale_idx_l * 8 + cp * 2 + 0], + q6_scales[scale_idx_l * 8 + cp * 2 + 1], + q6_scales[scale_idx_l * 8 + cp * 2 + 1], + }; + const int32x4_t scale_vec_h = { + q6_scales[scale_idx_h * 8 + cp * 2 + 0], + q6_scales[scale_idx_h * 8 + cp * 2 + 0], + q6_scales[scale_idx_h * 8 + cp * 2 + 1], + q6_scales[scale_idx_h * 8 + cp * 2 + 1], + }; + + acc[cp] = vmlaq_s32(acc[cp], sb_acc_0l, scale_vec_l); + acc[cp] = vmlaq_s32(acc[cp], sb_acc_0h, scale_vec_h); + acc[cp + 4] = vmlaq_s32(acc[cp + 4], sb_acc_1l, scale_vec_l); + acc[cp + 4] = vmlaq_s32(acc[cp + 4], sb_acc_1h, scale_vec_h); + } + } + } // for half + + // Reorder i8mm output to match memory layout + for (int i = 0; i < 8; i++) { + int32x2x2_t aux = vzip_s32(vget_low_s32(acc[i]), vget_high_s32(acc[i])); + acc[i] = vcombine_s32(aux.val[0], aux.val[1]); + } + int32x4_t reorder_acc[8] = { + vcombine_s32(vget_low_s32(acc[0]), vget_low_s32(acc[1])), + vcombine_s32(vget_low_s32(acc[2]), vget_low_s32(acc[3])), + vcombine_s32(vget_high_s32(acc[0]), vget_high_s32(acc[1])), + vcombine_s32(vget_high_s32(acc[2]), vget_high_s32(acc[3])), + vcombine_s32(vget_low_s32(acc[4]), vget_low_s32(acc[5])), + vcombine_s32(vget_low_s32(acc[6]), vget_low_s32(acc[7])), + vcombine_s32(vget_high_s32(acc[4]), vget_high_s32(acc[5])), + vcombine_s32(vget_high_s32(acc[6]), vget_high_s32(acc[7])), + }; + + // Apply superblock scale (no mins for q6_K) + for (int i = 0; i < q8_k_blocklen; i++) { + for (int j = 0; j < 2; j++) { + float32x4_t q8_d = vdupq_n_f32(q8_ptr[b].d[i]); + float32x4_t q6_d = vcvt_f32_f16(vld1_f16((const __fp16 *) (q6_ptr[b].d + j * 4))); + const float32x4_t scale = vmulq_f32(q6_d, q8_d); + + acc_f32[2 * i + j] = + vmlaq_f32(acc_f32[2 * i + j], vcvtq_f32_s32(reorder_acc[2 * i + j]), scale); + } + } + } // for b + + // Store results + for (int i = 0; i < q8_k_blocklen; i++) { + int row = y * q8_k_blocklen + i; + for (int j = 0; j < 2; j++) { + int col = x * ncols_interleaved + j * 4; + int offset = row * bs + col; + vst1q_f32(s + offset, acc_f32[2 * i + j]); + } + } + } // for x + } // for y + return; +#endif // defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8) + ggml_gemm_q6_K_8x8_q8_K_generic(n, s, bs, vx, vy, nr, nc); +} + void ggml_gemm_q8_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, diff --git a/ggml/src/ggml-cpu/common.h b/ggml/src/ggml-cpu/common.h index 6adca5437f..1057b5bb15 100644 --- a/ggml/src/ggml-cpu/common.h +++ b/ggml/src/ggml-cpu/common.h @@ -6,6 +6,9 @@ #include "ggml-impl.h" #include "simd-mappings.h" +#define GGML_FA_TILE_Q 32 +#define GGML_FA_TILE_KV 16 + #ifdef __cplusplus #include @@ -84,4 +87,9 @@ static std::pair get_thread_range(const struct ggml_compute_pa return {ir0, ir1}; } +struct ggml_fa_tile_config { + static constexpr size_t Q = GGML_FA_TILE_Q; + static constexpr size_t KV = GGML_FA_TILE_KV; +}; + #endif diff --git a/ggml/src/ggml-cpu/ggml-cpu.c b/ggml/src/ggml-cpu/ggml-cpu.c index 4c7a75e768..b1de2ae871 100644 --- a/ggml/src/ggml-cpu/ggml-cpu.c +++ b/ggml/src/ggml-cpu/ggml-cpu.c @@ -14,6 +14,7 @@ #include "vec.h" #include "ops.h" #include "ggml.h" +#include "common.h" #if defined(_MSC_VER) || defined(__MINGW32__) #include // using malloc.h with MSC/MINGW @@ -2866,10 +2867,12 @@ struct ggml_cplan ggml_graph_plan( } break; case GGML_OP_FLASH_ATTN_EXT: { - const int64_t ne10 = node->src[1]->ne[0]; // DK - const int64_t ne20 = node->src[2]->ne[0]; // DV + const int64_t DK = node->src[1]->ne[0]; + const int64_t DV = node->src[2]->ne[0]; - cur = sizeof(float)*(1*ne10 + 2*ne20)*n_tasks; // 1x head size K + 2x head size V (per thread) + // Tiled flash attention scratch (tile sizes defined in common.h) + // Per-thread: Q_q + KQ + mask + VKQ32 + V32 + padding + cur = sizeof(float)*(GGML_FA_TILE_Q*DK + 2*GGML_FA_TILE_Q*GGML_FA_TILE_KV + GGML_FA_TILE_Q*DV + GGML_FA_TILE_KV*DV)*n_tasks; } break; case GGML_OP_FLASH_ATTN_BACK: { diff --git a/ggml/src/ggml-cpu/llamafile/sgemm.cpp b/ggml/src/ggml-cpu/llamafile/sgemm.cpp index 7dc36d4f8a..8f980c16b9 100644 --- a/ggml/src/ggml-cpu/llamafile/sgemm.cpp +++ b/ggml/src/ggml-cpu/llamafile/sgemm.cpp @@ -1797,10 +1797,27 @@ class tinyBLAS_Q0_AVX { } \ } \ +template +struct mma_instr; + +template<> +struct mma_instr { + static inline void outer_product(acc_t *acc, vec_t a, vec_t b) { + __builtin_mma_xvbf16ger2pp(acc, a, b); + } +}; + +template<> +struct mma_instr { + static inline void outer_product(acc_t *acc, vec_t a, vec_t b) { + __builtin_mma_xvf16ger2pp(acc, a, b); + } +}; + template -class tinyBLAS_BF16_PPC { +class tinyBLAS_HP16_PPC { public: - tinyBLAS_BF16_PPC(int64_t k, + tinyBLAS_HP16_PPC(int64_t k, const TA *A, int64_t lda, const TB *B, int64_t ldb, TC *C, int64_t ldc, @@ -2118,8 +2135,8 @@ class tinyBLAS_BF16_PPC { packNormal((A+(ii*lda)+l), lda, 4, 8, (uint8_t*)vec_A); packNormal((B+(jj*ldb)+l), ldb, 8, 8, (uint8_t*)vec_B); for (int x = 0; x < 4; x++) { - __builtin_mma_xvbf16ger2pp(&acc_0, vec_A[x], vec_B[x]); - __builtin_mma_xvbf16ger2pp(&acc_1, vec_A[x], vec_B[x+4]); + mma_instr::outer_product(&acc_0, vec_A[x], vec_B[x]); + mma_instr::outer_product(&acc_1, vec_A[x], vec_B[x+4]); } } SAVE_ACC(&acc_0, ii, jj); @@ -2135,8 +2152,8 @@ class tinyBLAS_BF16_PPC { packNormal((A+(ii*lda)+l), lda, 8, 8, (uint8_t*)vec_A); packNormal((B+(jj*ldb)+l), ldb, 8, 4, (uint8_t*)vec_B); for (int x = 0; x < 4; x++) { - __builtin_mma_xvbf16ger2pp(&acc_0, vec_A[x], vec_B[x]); - __builtin_mma_xvbf16ger2pp(&acc_1, vec_A[x+4], vec_B[x]); + mma_instr::outer_product(&acc_0, vec_A[x], vec_B[x]); + mma_instr::outer_product(&acc_1, vec_A[x], vec_B[x+4]); } } SAVE_ACC(&acc_0, ii, jj); @@ -2155,10 +2172,10 @@ class tinyBLAS_BF16_PPC { packNormal(A+(ii*lda)+l, lda, 8, 8, (uint8_t*)vec_A); packNormal(B+(jj*ldb)+l, ldb, 8, 8, (uint8_t*)vec_B); for (int x = 0; x < 4; x++) { - __builtin_mma_xvbf16ger2pp(&acc_0, vec_A[x], vec_B[x]); - __builtin_mma_xvbf16ger2pp(&acc_1, (vec_t)vec_A[x], (vec_t)vec_B[x+4]); - __builtin_mma_xvbf16ger2pp(&acc_2, (vec_t)vec_A[x+4], (vec_t)vec_B[x]); - __builtin_mma_xvbf16ger2pp(&acc_3, (vec_t)vec_A[x+4], (vec_t)vec_B[x+4]); + mma_instr::outer_product(&acc_0, vec_A[x], vec_B[x]); + mma_instr::outer_product(&acc_1, vec_A[x], vec_B[x+4]); + mma_instr::outer_product(&acc_2, vec_A[x+4], vec_B[x]); + mma_instr::outer_product(&acc_3, vec_A[x+4], vec_B[x+4]); } } @@ -2189,7 +2206,7 @@ class tinyBLAS_BF16_PPC { packNormal(A+(ii*lda)+l, lda, RM, 4, (uint8_t*)vec_A); packNormal(B+(jj*ldb)+l, ldb, RN, 4, (uint8_t*)vec_B); for (int x = 0; x<2; x++) { - __builtin_mma_xvbf16ger2pp(&acc_0, vec_A[x], vec_B[x]); + mma_instr::outer_product(&acc_0, vec_A[x], vec_B[x]); } } __builtin_mma_disassemble_acc(vec_C, &acc_0); @@ -2224,8 +2241,8 @@ class tinyBLAS_BF16_PPC { packNormal(A+(ii*lda)+l, lda, RM, 8, (uint8_t*)vec_A); packNormal(B+(jj*ldb)+l, ldb, RN, 8, (uint8_t*)vec_B); for (int x = 0; x<4; x++) { - __builtin_mma_xvbf16ger2pp(&acc_0, vec_A[x], vec_B[x]); - __builtin_mma_xvbf16ger2pp(&acc_1, vec_A[x], vec_B[x+4]); + mma_instr::outer_product(&acc_0, vec_A[x], vec_B[x]); + mma_instr::outer_product(&acc_1, vec_A[x], vec_B[x+4]); } } __builtin_mma_disassemble_acc(vec_C, &acc_0); @@ -3418,16 +3435,19 @@ bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64 return tb.matmul(m, n); } #elif defined(__MMA__) - if ((k % 8)) - return false; - if(Btype == GGML_TYPE_BF16) { - tinyBLAS_BF16_PPC tb{ k, - (const ggml_bf16_t *)A, lda, - (const ggml_bf16_t *)B, ldb, - (float *)C, ldc, - params->ith, params->nth}; - tb.matmul(m, n); - return true; + if (k % 8) { + return false; + } + + if (Btype == GGML_TYPE_BF16) { + tinyBLAS_HP16_PPC tb{ k, + (const ggml_bf16_t *)A, lda, + (const ggml_bf16_t *)B, ldb, + (float *)C, ldc, + params->ith, params->nth }; + + tb.matmul(m, n); + return true; } #elif defined(__riscv_zvfbfwma) #if LMUL == 1 @@ -3516,6 +3536,21 @@ bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64 #endif return tb.matmul(m, n); } +#elif defined(__MMA__) + if (k % 8) { + return false; + } + + if (Btype == GGML_TYPE_F16) { + tinyBLAS_HP16_PPC tb{ k, + (const ggml_fp16_t *)A, lda, + (const ggml_fp16_t *)B, ldb, + (float *)C, ldc, + params->ith, params->nth }; + + tb.matmul(m, n); + return true; + } #endif return false; } diff --git a/ggml/src/ggml-cpu/ops.cpp b/ggml/src/ggml-cpu/ops.cpp index 387e2fe42c..48c8964361 100644 --- a/ggml/src/ggml-cpu/ops.cpp +++ b/ggml/src/ggml-cpu/ops.cpp @@ -8164,6 +8164,7 @@ static void ggml_compute_forward_flash_attn_ext_f16_one_chunk( // online softmax / attention // loop over n_kv and n_head_kv // ref: https://arxiv.org/pdf/2112.05682.pdf + for (int64_t ic = 0; ic < nek1; ++ic) { const float mv = mp ? slope*GGML_CPU_FP16_TO_FP32(mp[ic]) : 0.0f; if (mv == -INFINITY) { @@ -8271,6 +8272,280 @@ static void ggml_compute_forward_flash_attn_ext_f16_one_chunk( } } +static void ggml_compute_forward_flash_attn_ext_tiled( + const ggml_compute_params * params, + ggml_tensor * dst, + int ir0, int ir1) { + const ggml_tensor * q = dst->src[0]; + const ggml_tensor * k = dst->src[1]; + const ggml_tensor * v = dst->src[2]; + const ggml_tensor * mask = dst->src[3]; + const ggml_tensor * sinks = dst->src[4]; + + GGML_TENSOR_LOCALS(int64_t, neq, q, ne) + GGML_TENSOR_LOCALS(size_t, nbq, q, nb) + GGML_TENSOR_LOCALS(int64_t, nek, k, ne) + GGML_TENSOR_LOCALS(size_t, nbk, k, nb) + GGML_TENSOR_LOCALS(int64_t, nev, v, ne) + GGML_TENSOR_LOCALS(size_t, nbv, v, nb) + GGML_TENSOR_LOCALS(int64_t, ne, dst, ne) + GGML_TENSOR_LOCALS(size_t, nb, dst, nb) + + const int64_t DK = nek0; + const int64_t DV = nev0; + const int64_t N = neq1; + + GGML_ASSERT(ne0 == DV); + GGML_ASSERT(ne2 == N); + + // input tensor rows must be contiguous + GGML_ASSERT(nbq0 == ggml_type_size(q->type)); + GGML_ASSERT(nbk0 == ggml_type_size(k->type)); + GGML_ASSERT(nbv0 == ggml_type_size(v->type)); + + GGML_ASSERT(neq0 == DK); + GGML_ASSERT(nek0 == DK); + GGML_ASSERT(nev0 == DV); + + GGML_ASSERT(neq1 == N); + + // dst cannot be transposed or permuted + GGML_ASSERT(nb0 == sizeof(float)); + GGML_ASSERT(nb0 <= nb1); + GGML_ASSERT(nb1 <= nb2); + GGML_ASSERT(nb2 <= nb3); + + GGML_ASSERT(k->type == v->type); + const ggml_type kv_type = k->type; + + const auto * kv_type_traits_cpu = ggml_get_type_traits_cpu(kv_type); + const ggml_from_float_t kv_from_float = kv_type_traits_cpu->from_float; + const ggml_vec_dot_t kv_vec_dot = kv_type_traits_cpu->vec_dot; + const size_t kv_type_size = ggml_type_size(kv_type); + + // broadcast factors + const int64_t rk2 = neq2/nek2; + const int64_t rk3 = neq3/nek3; + + const int64_t rv2 = neq2/nev2; + const int64_t rv3 = neq3/nev3; + + float scale = 1.0f; + float max_bias = 0.0f; + float logit_softcap = 0.0f; + + memcpy(&scale, (float *) dst->op_params + 0, sizeof(float)); + memcpy(&max_bias, (float *) dst->op_params + 1, sizeof(float)); + memcpy(&logit_softcap, (float *) dst->op_params + 2, sizeof(float)); + + if (logit_softcap != 0) { + scale /= logit_softcap; + } + + const uint32_t n_head = neq2; + const uint32_t n_head_log2 = 1u << (uint32_t) floor(log2(n_head)); + + const float m0 = powf(2.0f, -(max_bias ) / n_head_log2); + const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2); + + int ith = params->ith; + + static constexpr int Q_TILE_SZ = ggml_fa_tile_config::Q; + static constexpr int KV_TILE_SZ = ggml_fa_tile_config::KV; + + GGML_ASSERT(nek1 % KV_TILE_SZ == 0 && "KV sequence length must be divisible by KV_TILE_SZ"); + + int ir = ir0; + while (ir < ir1) { + // q indices for the start of this tile + const int iq3 = ir/(neq2*neq1); + const int iq2 = (ir - iq3*neq2*neq1)/neq1; + const int iq1 = (ir - iq3*neq2*neq1 - iq2*neq1); + + // Number of valid rows in this tile: + // - limited by tile size (Q_TILE_SZ) + // - limited by chunk boundary (ir1 - ir) + // - limited by head boundary (neq1 - iq1) to avoid crossing into next head + const int tile_rows = MIN(Q_TILE_SZ, MIN((int)(ir1 - ir), (int)(neq1 - iq1))); + GGML_ASSERT(tile_rows > 0); + + const uint32_t h = iq2; // head index + const float slope = (max_bias > 0.0f) ? h < n_head_log2 ? powf(m0, h + 1) : powf(m1, 2*(h - n_head_log2) + 1) : 1.0f; + + float S[Q_TILE_SZ]; + float M[Q_TILE_SZ]; + + for (int i = 0 ; i < Q_TILE_SZ; ++i) { + S[i] = 0.; + M[i] = -INFINITY; + } + + // Per-thread scratch layout: + // Q_q: Q_TILE_SZ * DK (converted Q tile in KV type) + // KQ: Q_TILE_SZ * KV_TILE_SZ (attention scores in float) + // mask: Q_TILE_SZ * KV_TILE_SZ (mask in float) + // VKQ32: Q_TILE_SZ * DV (FP32 output accumulator) + // V32: KV_TILE_SZ * DV (F32 buffer for V tile - used for f166 conversion) + float * base = (float *) params->wdata + ith*(Q_TILE_SZ*DK + 2*Q_TILE_SZ*KV_TILE_SZ + Q_TILE_SZ*DV + KV_TILE_SZ*DV + CACHE_LINE_SIZE_F32); + + void * Q_q = base; + float * KQ = (float *)((char *)base + Q_TILE_SZ * DK * sizeof(float)); + float * mask32 = KQ + Q_TILE_SZ * KV_TILE_SZ; + float * VKQ32 = mask32 + Q_TILE_SZ * KV_TILE_SZ; + float * V32 = VKQ32 + Q_TILE_SZ * DV; // F32 buffer for V tile + + memset(VKQ32, 0, Q_TILE_SZ * DV * sizeof(float)); + memset(mask32, 0, Q_TILE_SZ * KV_TILE_SZ * sizeof(float)); + + // k indices + const int ik3 = iq3 / rk3; + const int ik2 = iq2 / rk2; + + // v indices + const int iv3 = iq3 / rv3; + const int iv2 = iq2 / rv2; + + for (int tq = 0; tq < tile_rows; tq++) { + const float * pq = (const float *) ((char *) q->data + ((iq1 + tq)*nbq1 + iq2*nbq2 + iq3*nbq3)); + kv_from_float(pq, (char *)Q_q + tq * DK * kv_type_size, DK); + } + // Zero-pad remaining rows + for (int tq = tile_rows; tq < Q_TILE_SZ; tq++) { + memset((char *)Q_q + tq * DK * kv_type_size, 0, DK * kv_type_size); + } + + for (int64_t ic = 0; ic < nek1; ic += KV_TILE_SZ) { + + // skip the tile entirely if all the masks are -inf + if (mask) { + bool can_skip = true; + for (int tq = 0; tq < tile_rows; tq++) { + const ggml_fp16_t * mp_row = (const ggml_fp16_t *)((const char *) mask->data + (iq1 + tq)*mask->nb[1] + (iq2%mask->ne[2])*mask->nb[2] + (iq3%mask->ne[3])*mask->nb[3]); + for (int tk = 0; tk < KV_TILE_SZ; tk++) { + mask32[tq * KV_TILE_SZ + tk] = slope * GGML_CPU_FP16_TO_FP32(mp_row[ic + tk]); + if (mask32[tq * KV_TILE_SZ + tk] != -INFINITY) { + can_skip = false; + } + } + } + + if (can_skip) { + continue; + } + } + + for (int tq = 0; tq < Q_TILE_SZ; tq++) { + const void * q_row = (const char *)Q_q + tq * DK * kv_type_size; + for (int tk = 0; tk < KV_TILE_SZ; tk++) { + const void * k_row = (const char *) k->data + ((ic + tk)*nbk1 + ik2*nbk2 + ik3*nbk3); + float s; + kv_vec_dot(DK, &s, 0, k_row, 0, q_row, 0, 1); + KQ[tq * KV_TILE_SZ + tk] = s * scale; + } + } + + if (logit_softcap != 0.0f) { + ggml_vec_tanh_f32(Q_TILE_SZ * KV_TILE_SZ, KQ, KQ); + ggml_vec_scale_f32(Q_TILE_SZ * KV_TILE_SZ, KQ, logit_softcap); + } + + if (mask) { + ggml_vec_add_f32(tile_rows * KV_TILE_SZ, KQ, KQ, mask32); + } + + bool skip[Q_TILE_SZ] = {}; + + for (int tq = 0; tq < Q_TILE_SZ; tq++) { + float * kq_row = KQ + tq * KV_TILE_SZ; + + float tile_max; + ggml_vec_max_f32(KV_TILE_SZ, &tile_max, kq_row); + + if (tile_max == -INFINITY) { + skip[tq] = true; + continue; + } + + const float Mold = M[tq]; + const float Mnew = fmaxf(Mold, tile_max); + + if (Mnew > Mold) { + const float ms = expf(Mold - Mnew); + ggml_vec_scale_f32(DV, VKQ32 + tq * DV, ms); + S[tq] *= ms; + } + M[tq] = Mnew; + + + S[tq] += ggml_vec_soft_max_f32(KV_TILE_SZ, kq_row, kq_row, Mnew); + } + + // Convert V tile to F32 first (if F16), then do MAD + // On x86, ggml_vec_mad_f16 internall converts F16<->F32 on every load/store, so pre-converting is faster. + // TODO: on ARM, native f16 should be faster + if (kv_type == GGML_TYPE_F16) { + for (int tk = 0; tk < KV_TILE_SZ; tk++) { + const ggml_fp16_t * v_row = (const ggml_fp16_t *)((const char *) v->data + ((ic + tk)*nbv1 + iv2*nbv2 + iv3*nbv3)); + ggml_fp16_to_fp32_row(v_row, V32 + tk * DV, DV); + } + for (int tq = 0; tq < Q_TILE_SZ; tq++) { + if (skip[tq]) continue; + float * vkq_row = VKQ32 + tq * DV; + for (int tk = 0; tk < KV_TILE_SZ; tk++) { + const float p = KQ[tq * KV_TILE_SZ + tk]; + ggml_vec_mad_f32(DV, vkq_row, V32 + tk * DV, p); + } + } + } else { + for (int tq = 0; tq < Q_TILE_SZ; tq++) { + if (skip[tq]) continue; + float * vkq_row = VKQ32 + tq * DV; + for (int tk = 0; tk < KV_TILE_SZ; tk++) { + const float p = KQ[tq * KV_TILE_SZ + tk]; + const float * v_row = (const float *)((const char *) v->data + ((ic + tk)*nbv1 + iv2*nbv2 + iv3*nbv3)); + ggml_vec_mad_f32(DV, vkq_row, v_row, p); + } + } + } + } + + // sinks (apply only to valid rows in the tile) + if (sinks) { + const float s = ((float *)((char *) sinks->data))[h]; + + for (int tq = 0; tq < tile_rows; tq++) { + float ms = 1.0f; + float vs = 1.0f; + + if (s > M[tq]) { + ms = expf(M[tq] - s); + ggml_vec_scale_f32(DV, VKQ32 + tq * DV, ms); + } else { + vs = expf(s - M[tq]); + } + + S[tq] = S[tq] * ms + vs; + } + } + + for (int tq = 0; tq < tile_rows; tq++) { + // V /= S + const float S_inv = S[tq] == 0.0f ? 0.0f : 1.0f / S[tq]; + ggml_vec_scale_f32(DV, VKQ32 + tq * DV, S_inv); + + // dst indices + const int i1 = iq1 + tq; + const int i2 = iq2; + const int i3 = iq3; + + // permute(0, 2, 1, 3) + memcpy((char *) dst->data + (i3*ne2*ne1 + i2 + i1*ne1)*nb1, VKQ32 + tq * DV, nb1); + } + + ir += tile_rows; + } +} + static void ggml_compute_forward_flash_attn_ext_f16( const ggml_compute_params * params, ggml_tensor * dst) { @@ -8343,6 +8618,15 @@ static void ggml_compute_forward_flash_attn_ext_f16( // The number of elements in each chunk const int64_t dr = (nr + nchunk - 1) / nchunk; + static constexpr int64_t KV_TILE_SZ = ggml_fa_tile_config::KV; + static constexpr int64_t Q_TILE_SZ = ggml_fa_tile_config::Q; + const bool kv_is_f32_or_f16 = (k->type == GGML_TYPE_F32 || k->type == GGML_TYPE_F16); + const bool use_tiled = (q->type == GGML_TYPE_F32 && + kv_is_f32_or_f16 && + k->type == v->type && + nek1 % KV_TILE_SZ == 0 && + neq1 >= Q_TILE_SZ); // Only use tiled for batch >= tile size + // The first chunk comes from our thread_id, the rest will get auto-assigned. int current_chunk = ith; @@ -8350,7 +8634,11 @@ static void ggml_compute_forward_flash_attn_ext_f16( const int64_t ir0 = dr * current_chunk; const int64_t ir1 = MIN(ir0 + dr, nr); - ggml_compute_forward_flash_attn_ext_f16_one_chunk(params, dst, ir0, ir1); + if (use_tiled) { + ggml_compute_forward_flash_attn_ext_tiled(params, dst, ir0, ir1); + } else { + ggml_compute_forward_flash_attn_ext_f16_one_chunk(params, dst, ir0, ir1); + } current_chunk = ggml_threadpool_chunk_add(params->threadpool, 1); } diff --git a/ggml/src/ggml-cpu/repack.cpp b/ggml/src/ggml-cpu/repack.cpp index 19e021e59a..24e8ab4618 100644 --- a/ggml/src/ggml-cpu/repack.cpp +++ b/ggml/src/ggml-cpu/repack.cpp @@ -703,6 +703,97 @@ void ggml_gemv_q5_K_8x8_q8_K_generic(int n, } } + +void ggml_gemv_q6_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { + constexpr int qk = QK_K; + const int nb = n / qk; + const int ncols_interleaved = 8; + const int blocklen = 8; + + assert(n % qk == 0); + assert(nc % ncols_interleaved == 0); + + UNUSED(bs); + UNUSED(nr); + + float sumf[8]; + + const block_q8_K * a_ptr = (const block_q8_K *) vy; + for (int x = 0; x < nc / ncols_interleaved; x++) { + const block_q6_Kx8 * b_ptr = (const block_q6_Kx8 *) vx + (x * nb); + + for (int j = 0; j < ncols_interleaved; j++) { + sumf[j] = 0.0f; + } + + for (int l = 0; l < nb; l++) { + + + for (int k = 0; k < 16; k++) { + // k = 0.. 7 weights 0-63 low, 64-127 high + // k = 8..15 weights 128-191 low, 192-255 high + const int base_l = (k / 8) * 128 + (k % 8) * 8; + const int base_h = base_l + 64; + + const int scale_idx_l = base_l / 16; + const int scale_idx_h = base_h / 16; + + // Bit shift cycles 0,2,4,6 for each 32-value group within a 128-value half + const int qh_shift_l = ((base_l % 128) / 32) * 2; + const int qh_shift_h = ((base_h % 128) / 32) * 2; + + // qh_half: offset to the correct 32-byte half (0 or 32) + const int qh_half_l = (base_l / 128) * 32; + const int qh_half_h = (base_h / 128) * 32; + + for (int j = 0; j < ncols_interleaved; j++) { + // Interleaved scales + const int8_t scale_l = b_ptr[l].scales[scale_idx_l * 8 + j]; + const int8_t scale_h = b_ptr[l].scales[scale_idx_h * 8 + j]; + + int sumi_l = 0; + int sumi_h = 0; + + for (int i = 0; i < blocklen; i++) { + const int ql_pos = k * 64 + j * 8 + i; + const int l_4 = b_ptr[l].ql[ql_pos] & 0xF; + const int hi_4 = (b_ptr[l].ql[ql_pos] >> 4) & 0xF; + + // qh indexing with 8-byte interleaving (like q5_K) + const int qh_byte_l = qh_half_l + ((base_l + i) % 32); + const int qh_chunk_l = qh_byte_l / 8; + const int qh_pos_l = qh_byte_l % 8; + const int qh_offset_l = qh_chunk_l * 64 + j * 8 + qh_pos_l; + const int hi_2_l = (b_ptr[l].qh[qh_offset_l] >> qh_shift_l) & 0x3; + + const int qh_byte_h = qh_half_h + ((base_h + i) % 32); + const int qh_chunk_h = qh_byte_h / 8; + const int qh_pos_h = qh_byte_h % 8; + const int qh_offset_h = qh_chunk_h * 64 + j * 8 + qh_pos_h; + const int hi_2_h = (b_ptr[l].qh[qh_offset_h] >> qh_shift_h) & 0x3; + + const int q_l = ((hi_2_l << 4) | l_4) - 32; + const int q_h = ((hi_2_h << 4) | hi_4) - 32; + + const int8_t a_l = a_ptr[l].qs[base_l + i]; + const int8_t a_h = a_ptr[l].qs[base_h + i]; + + sumi_l += q_l * a_l; + sumi_h += q_h * a_h; + } + + sumf[j] += + (sumi_l * scale_l + sumi_h * scale_h) * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d; + } + } + } + + for (int j = 0; j < ncols_interleaved; j++) { + s[x * ncols_interleaved + j] = sumf[j]; + } + } +} + void ggml_gemv_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { const int qk = QK8_0; const int nb = n / qk; @@ -1133,15 +1224,7 @@ void ggml_gemm_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, assert (nr % 4 == 0); assert (nc % ncols_interleaved == 0); - UNUSED(s); UNUSED(bs); - UNUSED(vx); - UNUSED(vy); - UNUSED(nr); - UNUSED(nc); - UNUSED(nb); - UNUSED(ncols_interleaved); - UNUSED(blocklen); float sumf[4][8]; float sum_minf[4][8]; @@ -1402,6 +1485,111 @@ void ggml_gemm_q5_K_8x8_q8_K_generic(int n, } } +void ggml_gemm_q6_K_8x8_q8_K_generic(int n, + float * GGML_RESTRICT s, + size_t bs, + const void * GGML_RESTRICT vx, + const void * GGML_RESTRICT vy, + int nr, + int nc) { + const int qk = QK_K; + const int nb = n / qk; + const int ncols_interleaved = 8; + const int blocklen = 8; + + assert(n % qk == 0); + assert(nr % 4 == 0); + assert(nc % ncols_interleaved == 0); + + UNUSED(bs); + + float sumf[4][8]; + + for (int y = 0; y < nr / 4; y++) { + const block_q8_Kx4 * a_ptr = (const block_q8_Kx4 *) vy + (y * nb); + for (int x = 0; x < nc / ncols_interleaved; x++) { + const block_q6_Kx8 * b_ptr = (const block_q6_Kx8 *) vx + (x * nb); + + for (int m = 0; m < 4; m++) { + for (int j = 0; j < ncols_interleaved; j++) { + sumf[m][j] = 0.0f; + } + } + + for (int l = 0; l < nb; l++) { + for (int k = 0; k < 16; k++) { + // k = 0.. 7 weights 0-63 low, 64-127 high + // k = 8..15 weights 128-191 low, 192-255 high + const int base_l = (k / 8) * 128 + (k % 8) * 8; + const int base_h = base_l + 64; + + const int scale_idx_l = base_l / 16; + const int scale_idx_h = base_h / 16; + + // Bit shift cycles 0,2,4,6 for each 32-value group within a 128-value half + const int qh_shift_l = ((base_l % 128) / 32) * 2; + const int qh_shift_h = ((base_h % 128) / 32) * 2; + + // qh_half: offset to the correct 32-byte half (0 or 32) + const int qh_half_l = (base_l / 128) * 32; + const int qh_half_h = (base_h / 128) * 32; + + // Activation base indices for q8_Kx4 interleaved format + // Layout: 128-value halves (k/8), then 8-value sub-blocks (k%8) with stride 32 + const int q8_base = (k / 8) * 512 + (k % 8) * 32; + + for (int m = 0; m < 4; m++) { + for (int j = 0; j < ncols_interleaved; j++) { + // Interleaved scales + const int8_t scale_l = b_ptr[l].scales[scale_idx_l * 8 + j]; + const int8_t scale_h = b_ptr[l].scales[scale_idx_h * 8 + j]; + + int sumi_l = 0; + int sumi_h = 0; + + for (int i = 0; i < blocklen; i++) { + const int ql_pos = k * 64 + j * 8 + i; + const int l_4 = b_ptr[l].ql[ql_pos] & 0xF; + const int hi_4 = (b_ptr[l].ql[ql_pos] >> 4) & 0xF; + + const int qh_idx_l = qh_half_l + ((base_l + i) % 32); + const int qh_chunk_l = qh_idx_l / 8; + const int qh_pos_l = qh_idx_l % 8; + const int qh_offset_l = qh_chunk_l * 64 + j * 8 + qh_pos_l; + const int hi_2_l = (b_ptr[l].qh[qh_offset_l] >> qh_shift_l) & 0x3; + + const int qh_idx_h = qh_half_h + ((base_h + i) % 32); + const int qh_chunk_h = qh_idx_h / 8; + const int qh_pos_h = qh_idx_h % 8; + const int qh_offset_h = qh_chunk_h * 64 + j * 8 + qh_pos_h; + const int hi_2_h = (b_ptr[l].qh[qh_offset_h] >> qh_shift_h) & 0x3; + + const int q_l = ((hi_2_l << 4) | l_4) - 32; + const int q_h = ((hi_2_h << 4) | hi_4) - 32; + + const int8_t q8_l = a_ptr[l].qs[q8_base + m * 8 + i]; + const int8_t q8_h = a_ptr[l].qs[q8_base + m * 8 + i + 256]; + + sumi_l += q_l * q8_l; + sumi_h += q_h * q8_h; + } + + sumf[m][j] += (sumi_l * scale_l + sumi_h * scale_h) * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * + a_ptr[l].d[m]; + } + } + } + } + + for (int m = 0; m < 4; m++) { + for (int j = 0; j < ncols_interleaved; j++) { + s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j]; + } + } + } + } +} + void ggml_gemm_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { const int qk = QK8_0; const int nb = n / qk; @@ -1801,8 +1989,7 @@ static block_q2_Kx8 make_block_q2_Kx8(block_q2_K * in, unsigned int blck_size_in // Every 16 byte is packed such that it contains scales and mins for corresponding sub blocks from Q2_K structure // For eg - First 16 bytes contains 16 scales and 16 mins - each of first and second sub blocks from different Q2_K structures - for(int i = 0; i < 128; i++){ - + for (int i = 0; i < 128; i++) { // Index for selecting which q2k super block int src1 = (i % 16) / 2; // Index for selecting scale @@ -1902,6 +2089,52 @@ static block_q5_Kx8 make_block_q5_Kx8(block_q5_K * in, unsigned int blck_size_in return out; } +static block_q6_Kx8 make_block_q6_Kx8(block_q6_K * in, unsigned int blck_size_interleave) { + block_q6_Kx8 out; + constexpr int n_blocks = 8; // Kx8 + for (int i = 0; i < n_blocks; i++) { + out.d[i] = in[i].d; + } + + const int end_ls = QK_K * 4 / blck_size_interleave; + // Interleave Q6_K quants by taking 8 bytes at a time + for (int i = 0; i < end_ls; ++i) { + int src_id = i % n_blocks; + int src_offset = (i / n_blocks) * blck_size_interleave; + int dst_offset = i * blck_size_interleave; + + uint64_t elem_ls; + memcpy(&elem_ls, &in[src_id].ql[src_offset], sizeof(uint64_t)); + memcpy(&out.ql[dst_offset], &elem_ls, sizeof(uint64_t)); + } + + // Interleave high bits using same 8-byte pattern as low bits + const int end_hs = end_ls / 2; + for (int i = 0; i < end_hs; ++i) { + int src_id = i % n_blocks; + int src_offset = (i / n_blocks) * blck_size_interleave; + int dst_offset = i * blck_size_interleave; + + uint64_t elem_hs; + memcpy(&elem_hs, &in[src_id].qh[src_offset], sizeof(uint64_t)); + memcpy(&out.qh[dst_offset], &elem_hs, sizeof(uint64_t)); + } + + // The below logic is designed so as to unpack and rearrange scales in Q6_K + // The output Q6_Kx8 structure interleaves the 8 bit scales in the same fashion as the quants + // Q6_K structure has an 8-bit scale per 16 elements -> 16 scales + // scales: [0 bl0 0 bl1 ... 0 bl7][1 bl0 ... 1 bl7] ... [15 bl0 ... 15 bl7] (bl = block) + constexpr int n_scales = QK_K / 16; + + for (int i = 0; i < n_blocks; i++) { + for (int j = 0; j < n_scales; j++) { + out.scales[j * n_blocks + i] = in[i].scales[j]; + } + } + + return out; +} + static int repack_q4_0_to_q4_0_4_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) { GGML_ASSERT(t->type == GGML_TYPE_Q4_0); GGML_ASSERT(interleave_block == 4 || interleave_block == 8); @@ -1983,7 +2216,7 @@ static int repack_q2_K_to_q2_K_8_bl(struct ggml_tensor * t, int interleave_block for (int b = 0; b < nrow; b += nrows_interleaved) { for (int64_t x = 0; x < nblocks; x++) { - for (int i = 0; i < nrows_interleaved; i++ ) { + for (int i = 0; i < nrows_interleaved; i++) { dst_tmp[i] = src[x + i * nblocks]; } *dst++ = make_block_q2_Kx8(dst_tmp, interleave_block); @@ -2027,6 +2260,35 @@ static int repack_q5_K_to_q5_K_8_bl(struct ggml_tensor * t, return 0; } +static int repack_q6_K_to_q6_K_8_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) { + GGML_ASSERT(t->type == GGML_TYPE_Q6_K); + GGML_ASSERT(interleave_block == 8); + constexpr int nrows_interleaved = 8; + + block_q6_Kx8 * dst = (block_q6_Kx8 *)t->data; + const block_q6_K * src = (const block_q6_K *) data; + block_q6_K dst_tmp[8]; + int nrow = ggml_nrows(t); + int nblocks = t->ne[0] / QK_K; + + GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_q6_K)); + + if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % 8 != 0) { + return -1; + } + + for (int b = 0; b < nrow; b += nrows_interleaved) { + for (int64_t x = 0; x < nblocks; x++) { + for (int i = 0; i < nrows_interleaved; i++) { + dst_tmp[i] = src[x + i * nblocks]; + } + *dst++ = make_block_q6_Kx8(dst_tmp, interleave_block); + } + src += nrows_interleaved * nblocks; + } + return 0; +} + static int repack_q4_0_to_q4_0_8_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) { GGML_ASSERT(t->type == GGML_TYPE_Q4_0); GGML_ASSERT(interleave_block == 8); @@ -2249,6 +2511,10 @@ template <> int repack(struct ggml_tensor * t, const void * da return repack_q5_K_to_q5_K_8_bl(t, 8, data, data_size); } +template <> int repack(struct ggml_tensor * t, const void * data, size_t data_size) { + return repack_q6_K_to_q6_K_8_bl(t, 8, data, data_size); +} + template <> int repack(struct ggml_tensor * t, const void * data, size_t data_size) { return repack_iq4_nl_to_iq4_nl_4_bl(t, 4, data, data_size); } @@ -2286,7 +2552,14 @@ template <> void gemv(int n, float * s, size_t ggml_gemv_q4_0_8x8_q8_0(n, s, bs, vx, vy, nr, nc); } -template <> void gemv(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { +template <> +void gemv(int n, + float * s, + size_t bs, + const void * vx, + const void * vy, + int nr, + int nc) { ggml_gemv_q2_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc); } @@ -2302,6 +2575,10 @@ template <> void gemv(int n, float * s, size_t ggml_gemv_q5_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc); } +template <> void gemv(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { + ggml_gemv_q6_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc); +} + template <> void gemv(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { ggml_gemv_iq4_nl_4x4_q8_0(n, s, bs, vx, vy, nr, nc); } @@ -2330,7 +2607,14 @@ template <> void gemm(int n, float * s, size_t ggml_gemm_q4_0_4x8_q8_0(n, s, bs, vx, vy, nr, nc); } -template <> void gemm(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { +template <> +void gemm(int n, + float * s, + size_t bs, + const void * vx, + const void * vy, + int nr, + int nc) { ggml_gemm_q4_0_8x8_q8_0(n, s, bs, vx, vy, nr, nc); } @@ -2350,6 +2634,10 @@ template <> void gemm(int n, float * s, size_t ggml_gemm_q5_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc); } +template <> void gemm(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { + ggml_gemm_q6_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc); +} + template <> void gemm(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { ggml_gemm_iq4_nl_4x4_q8_0(n, s, bs, vx, vy, nr, nc); } @@ -2714,20 +3002,19 @@ template (ne00, - (float *)((char *) dst->data + (i1 * nb1 + i2 * nb2)) + src0_cur_start, ne01, - src0_cur + src0_cur_start * nb01, - src1_col, 1, src0_cur_end - src0_cur_start); + gemv( + ne00, (float *) ((char *) dst->data + (i1 * nb1 + i2 * nb2)) + src0_cur_start, ne01, + src0_cur + src0_cur_start * nb01, src1_col, 1, src0_cur_end - src0_cur_start); } } #undef MMID_MATRIX_ROW @@ -2743,7 +3030,6 @@ template q4_0_4x4_q8_0; static const ggml::cpu::repack::tensor_traits q4_0_4x8_q8_0; @@ -2756,6 +3042,9 @@ static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(cons // instance for Q5_K static const ggml::cpu::repack::tensor_traits q5_K_8x8_q8_K; + // instance for Q6_K + static const ggml::cpu::repack::tensor_traits q6_K_8x8_q8_K; + // instance for Q2 static const ggml::cpu::repack::tensor_traits q2_K_8x8_q8_K; @@ -2812,6 +3101,12 @@ static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(cons return &q5_K_8x8_q8_K; } } + } else if (cur->type == GGML_TYPE_Q6_K) { + if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) { + if (cur->ne[1] % 8 == 0) { + return &q6_K_8x8_q8_K; + } + } } else if (cur->type == GGML_TYPE_IQ4_NL) { if (ggml_cpu_has_avx2()) { if (cur->ne[1] % 8 == 0) { diff --git a/ggml/src/ggml-cpu/repack.h b/ggml/src/ggml-cpu/repack.h index da87103157..855320eeeb 100644 --- a/ggml/src/ggml-cpu/repack.h +++ b/ggml/src/ggml-cpu/repack.h @@ -65,6 +65,16 @@ struct block_q5_Kx8 { static_assert(sizeof(block_q5_Kx8) == sizeof(ggml_half) * 16 + K_SCALE_SIZE * 8 + QK_K * 5, "wrong q5_K block size/padding"); +struct block_q6_Kx8 { + ggml_half d[8]; + int8_t scales[QK_K / 16 * 8]; + uint8_t ql[QK_K / 2 * 8]; // low bits of 6-bit quants (groups of 2) + uint8_t qh[QK_K / 4 * 8]; // high bits of 6-bit quants (groups of 4) +}; + +static_assert(sizeof(block_q6_Kx8) == sizeof(ggml_half) * 8 + QK_K / 16 * 8 + 3 * QK_K / 4 * 8, + "wrong q6_K block size/padding"); + struct block_q8_Kx4 { float d[4]; // delta int8_t qs[QK_K * 4]; // quants @@ -95,13 +105,14 @@ void ggml_quantize_mat_q8_0_4x4(const float * GGML_RESTRICT x, void * GGML_RESTR void ggml_quantize_mat_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k); void ggml_quantize_mat_q8_K_4x4(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k); void ggml_quantize_mat_q8_K_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k); -void ggml_gemv_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemv_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemv_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); +void ggml_gemv_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemv_q4_K_8x4_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemv_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemv_q5_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); +void ggml_gemv_q6_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemv_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemv_iq4_nl_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); @@ -111,6 +122,7 @@ void ggml_gemm_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo void ggml_gemm_q4_K_8x4_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemm_q5_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); +void ggml_gemm_q6_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemm_iq4_nl_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemv_q8_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); @@ -130,6 +142,7 @@ void ggml_gemv_q2_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, void ggml_gemv_q4_K_8x4_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemv_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemv_q5_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); +void ggml_gemv_q6_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemv_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemv_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemm_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); @@ -139,6 +152,7 @@ void ggml_gemm_q2_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, void ggml_gemm_q4_K_8x4_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemm_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemm_q5_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); +void ggml_gemm_q6_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemm_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemm_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemv_q8_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); diff --git a/ggml/src/ggml-cuda/common.cuh b/ggml/src/ggml-cuda/common.cuh index 09a491a836..3335f443ae 100644 --- a/ggml/src/ggml-cuda/common.cuh +++ b/ggml/src/ggml-cuda/common.cuh @@ -53,6 +53,7 @@ // While BW spans CC 1000, 1100 & 1200, we are integrating Tensor Core instructions available to 1200 family, see // https://docs.nvidia.com/cutlass/media/docs/cpp/blackwell_functionality.html#blackwell-sm120-gemms #define GGML_CUDA_CC_BLACKWELL 1200 +#define GGML_CUDA_CC_DGX_SPARK 1210 #define GGML_CUDA_CC_RUBIN 1300 #define GGML_CUDA_CC_OFFSET_AMD 0x1000000 #define GGML_CUDA_CC_OFFSET_MTHREADS 0x0100000 diff --git a/ggml/src/ggml-cuda/fattn-common.cuh b/ggml/src/ggml-cuda/fattn-common.cuh index 40c7725784..b6a7460da8 100644 --- a/ggml/src/ggml-cuda/fattn-common.cuh +++ b/ggml/src/ggml-cuda/fattn-common.cuh @@ -629,8 +629,8 @@ static __global__ void flash_attn_mask_to_KV_max( template // D == head size __launch_bounds__(D, 1) static __global__ void flash_attn_stream_k_fixup( - float * __restrict__ dst, const float2 * __restrict__ dst_fixup, const int ne01, const int ne02, const int ne03, const int ne11, - const int nbatch_fa) { + float * __restrict__ dst, const float2 * __restrict__ dst_fixup, const int ne01, const int ne02, const int ne03, + const int ne11, const int ne12, const int nbatch_fa) { constexpr int ncols = ncols1*ncols2; const int bidx0 = blockIdx.x; @@ -641,11 +641,14 @@ static __global__ void flash_attn_stream_k_fixup( const float * dst_fixup_data = ((const float *) dst_fixup) + gridDim.x*(2*2*ncols); - const int iter_k = (ne11 + (nbatch_fa - 1)) / nbatch_fa; - const int iter_j = (ne01 + (ncols1 - 1)) / ncols1; + const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix. - const int kbc0 = int64_t(bidx0 + 0)*(iter_k*iter_j*(ne02/ncols2)*ne03) / gridDim.x; - const int kbc0_stop = int64_t(bidx0 + 1)*(iter_k*iter_j*(ne02/ncols2)*ne03) / gridDim.x; + const int iter_k = (ne11 + (nbatch_fa - 1)) / nbatch_fa; + const int iter_j = (ne01 + (ncols1 - 1)) / ncols1; + const int iter_z_gqa = (gqa_ratio + (ncols2 - 1)) / ncols2; + + const int kbc0 = int64_t(bidx0 + 0)*(iter_k*iter_j*iter_z_gqa*ne12*ne03) / gridDim.x; + const int kbc0_stop = int64_t(bidx0 + 1)*(iter_k*iter_j*iter_z_gqa*ne12*ne03) / gridDim.x; const bool did_not_have_any_data = kbc0 == kbc0_stop; const bool wrote_beginning_of_tile = kbc0 % iter_k == 0; @@ -654,15 +657,19 @@ static __global__ void flash_attn_stream_k_fixup( return; } - const int sequence = kbc0 / (iter_k*iter_j*(ne02/ncols2)); - const int head = (kbc0 - iter_k*iter_j*(ne02/ncols2)*sequence) / (iter_k*iter_j); - const int jt = (kbc0 - iter_k*iter_j*(ne02/ncols2)*sequence - iter_k*iter_j*head) / iter_k; // j index of current tile. + // z_KV == K/V head index, zt_gqa = Q head start index per K/V head, jt = token position start index + const int sequence = kbc0 /(iter_k*iter_j*iter_z_gqa*ne12); + const int z_KV = (kbc0 - iter_k*iter_j*iter_z_gqa*ne12 * sequence)/(iter_k*iter_j*iter_z_gqa); + const int zt_gqa = (kbc0 - iter_k*iter_j*iter_z_gqa*ne12 * sequence - iter_k*iter_j*iter_z_gqa * z_KV)/(iter_k*iter_j); + const int jt = (kbc0 - iter_k*iter_j*iter_z_gqa*ne12 * sequence - iter_k*iter_j*iter_z_gqa * z_KV - iter_k*iter_j * zt_gqa) / iter_k; - if (jt*ncols1 + j >= ne01) { + const int zt_Q = z_KV*gqa_ratio + zt_gqa*ncols2; // Global Q head start index. + + if (jt*ncols1 + j >= ne01 || zt_gqa*ncols2 + c >= gqa_ratio) { return; } - dst += sequence*ne02*ne01*D + jt*ne02*(ncols1*D) + head*(ncols2*D) + (j*ne02 + c)*D + tid; + dst += sequence*ne02*ne01*D + jt*ne02*(ncols1*D) + zt_Q*D + (j*ne02 + c)*D + tid; // Load the partial result that needs a fixup: float dst_val = 0.0f; @@ -681,7 +688,7 @@ static __global__ void flash_attn_stream_k_fixup( int bidx = bidx0 - 1; int kbc_stop = kbc0; while(true) { - const int kbc = int64_t(bidx)*(iter_k*iter_j*(ne02/ncols2)*ne03) / gridDim.x; + const int kbc = int64_t(bidx)*(iter_k*iter_j*iter_z_gqa*ne12*ne03) / gridDim.x; if (kbc == kbc_stop) { // Did not have any data. bidx--; kbc_stop = kbc; @@ -782,7 +789,7 @@ void launch_fattn( const ggml_tensor * K = dst->src[1]; const ggml_tensor * V = dst->src[2]; - const bool V_is_K_view = V->op == GGML_OP_VIEW && V->src[0] == K && V->data == K->data; + const bool V_is_K_view = V->view_src && (V->view_src == K || (V->view_src == K->view_src && V->view_offs == K->view_offs)); const ggml_tensor * mask = dst->src[3]; const ggml_tensor * sinks = dst->src[4]; @@ -882,8 +889,10 @@ void launch_fattn( } } - const int ntiles_x = ((Q->ne[1] + ncols1 - 1) / ncols1); - const int ntiles_total = ntiles_x * (Q->ne[2] / ncols2) * Q->ne[3]; + const int ntiles_x = ((Q->ne[1] + ncols1 - 1) / ncols1); + const int gqa_ratio = Q->ne[2] / K->ne[2]; + const int ntiles_z_gqa = ((gqa_ratio + ncols2 - 1) / ncols2); + const int ntiles_total = ntiles_x * ntiles_z_gqa * K->ne[2] * Q->ne[3]; // Optional optimization where the mask is scanned to determine whether part of the calculation can be skipped. // Only worth the overhead if there is at lease one FATTN_KQ_STRIDE x FATTN_KQ_STRIDE square to be skipped or @@ -958,7 +967,7 @@ void launch_fattn( blocks_num.x = ntiles_x; blocks_num.y = parallel_blocks; - blocks_num.z = (Q->ne[2]/ncols2)*Q->ne[3]; + blocks_num.z = ntiles_z_gqa*K->ne[2]*Q->ne[3]; if (parallel_blocks > 1) { dst_tmp.alloc(parallel_blocks*ggml_nelements(KQV)); @@ -1012,7 +1021,7 @@ void launch_fattn( flash_attn_stream_k_fixup <<>> - ((float *) KQV->data, dst_tmp_meta.ptr, Q->ne[1], Q->ne[2], Q->ne[3], K->ne[1], nbatch_fa); + ((float *) KQV->data, dst_tmp_meta.ptr, Q->ne[1], Q->ne[2], Q->ne[3], K->ne[1], K->ne[2], nbatch_fa); } } else if (parallel_blocks > 1) { const dim3 block_dim_combine(DV, 1, 1); diff --git a/ggml/src/ggml-cuda/fattn-mma-f16.cuh b/ggml/src/ggml-cuda/fattn-mma-f16.cuh index 3e7d67b40d..0b8ef90794 100644 --- a/ggml/src/ggml-cuda/fattn-mma-f16.cuh +++ b/ggml/src/ggml-cuda/fattn-mma-f16.cuh @@ -933,6 +933,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile( const float logit_softcap, const uint3 ne01, const int ne02, + const int gqa_ratio, const int ne11, const int stride_Q1, const int stride_Q2, @@ -940,6 +941,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile( const int stride_V, const int stride_mask, const int jt, + const int zt_gqa, const int kb0_start, const int kb0_stop) { #if defined(VOLTA_MMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || (defined(AMD_WMMA_AVAILABLE) && defined(RDNA4)) @@ -1022,7 +1024,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile( const int j = jc / ncols2; const int c = jc % ncols2; - if (jt*ncols1 + j < int(ne01.z)) { + if ((ncols1 == 1 || jt*ncols1 + j < int(ne01.z)) && (ncols2 == 1 || zt_gqa*ncols2 + c < gqa_ratio)) { #pragma unroll for (int k0 = k0_start; k0 < k0_stop; k0 += stride_k) { const int k = k0 + (stride_k == WARP_SIZE ? threadIdx.x : threadIdx.x % stride_k); @@ -1408,7 +1410,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile( const int j_dst = jc_dst / ncols2; const int c_dst = jc_dst % ncols2; - if (!is_fixup && jt*ncols1 + j_dst >= int(ne01.z)) { + if (!is_fixup && ((ncols1 > 1 && jt*ncols1 + j_dst >= int(ne01.z)) || (ncols2 > 1 && zt_gqa*ncols2 + c_dst >= gqa_ratio))) { continue; } @@ -1447,7 +1449,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile( } #else GGML_UNUSED_VARS(Q_f2, K_h2, V_h2, mask_h, sinks_f, dstk, dstk_fixup, - scale, slope, logit_softcap, ne01, ne02, + scale, slope, logit_softcap, ne01, ne02, gqa_ratio, stride_Q1, stride_Q2, stride_K, stride_V, stride_mask, jt, kb0_start, kb0_stop); NO_DEVICE_CODE; @@ -1520,12 +1522,13 @@ static __global__ void flash_attn_ext_f16( const int stride_V = V_is_K_view ? stride_K : nb21 / sizeof(half2); - const int iter_k = (ne11 + (nbatch_fa - 1)) / nbatch_fa; - const int iter_j = (ne01.z + (ncols1 - 1)) / ncols1; + const int iter_k = (ne11 + (nbatch_fa - 1)) / nbatch_fa; + const int iter_j = (ne01.z + (ncols1 - 1)) / ncols1; + const int iter_z_gqa = (gqa_ratio + (ncols2 - 1)) / ncols2; // kbc == k block continuous, current index in continuous ijk space. - int kbc = int64_t(blockIdx.x + 0)*(iter_k*iter_j*(ne02/ncols2)*ne03) / gridDim.x; - const int kbc_stop = int64_t(blockIdx.x + 1)*(iter_k*iter_j*(ne02/ncols2)*ne03) / gridDim.x; + int kbc = int64_t(blockIdx.x + 0)*(iter_k*iter_j*iter_z_gqa*ne12*ne03) / gridDim.x; + const int kbc_stop = int64_t(blockIdx.x + 1)*(iter_k*iter_j*iter_z_gqa*ne12*ne03) / gridDim.x; // If the seams of 2 CUDA blocks fall within an output tile their results need to be combined. // For this we need to track both the block that starts the tile (needs_fixup) and the block that finishes the tile (is_fixup). @@ -1536,22 +1539,24 @@ static __global__ void flash_attn_ext_f16( int kb0_stop = min(iter_k, kb0_start + kbc_stop - kbc); while (kbc < kbc_stop && kb0_stop == iter_k) { - const int sequence = kbc / (iter_k*iter_j*(ne02/ncols2)); - const int zt = (kbc - iter_k*iter_j*(ne02/ncols2)*sequence) / (iter_k*iter_j); // head in units of ncols2 - const int jt = (kbc - iter_k*iter_j*(ne02/ncols2)*sequence - iter_k*iter_j*zt) / iter_k; // j index of current tile. + // z_KV == K/V head index, zt_gqa = Q head start index per K/V head, jt = token position start index + const int sequence = kbc /(iter_k*iter_j*iter_z_gqa*ne12); + const int z_KV = (kbc - iter_k*iter_j*iter_z_gqa*ne12 * sequence)/(iter_k*iter_j*iter_z_gqa); + const int zt_gqa = (kbc - iter_k*iter_j*iter_z_gqa*ne12 * sequence - iter_k*iter_j*iter_z_gqa * z_KV)/(iter_k*iter_j); + const int jt = (kbc - iter_k*iter_j*iter_z_gqa*ne12 * sequence - iter_k*iter_j*iter_z_gqa * z_KV - iter_k*iter_j * zt_gqa) / iter_k; - const int head0 = zt * ncols2; + const int zt_Q = z_KV*gqa_ratio + zt_gqa*ncols2; // Global Q head start index. - const float2 * Q_f2 = (const float2 *) (Q + nb03*sequence + nb02* head0); - const half2 * K_h2 = (const half2 *) (K + nb13*sequence + nb12*(head0 / gqa_ratio)); + const float2 * Q_f2 = (const float2 *) (Q + nb03*sequence + nb02*zt_Q); + const half2 * K_h2 = (const half2 *) (K + nb13*sequence + nb12*z_KV); const half * mask_h = ncols2 == 1 && !mask ? nullptr : (const half *) (mask + nb33*(sequence % ne33)); - float2 * dstk = ((float2 *) dst) + (sequence*ne01.z*ne02 + head0) * (DV/2); + float2 * dstk = ((float2 *) dst) + (sequence*ne01.z*ne02 + zt_Q) * (DV/2); - const half2 * V_h2 = V_is_K_view ? K_h2 : (const half2 *) (V + nb23*sequence + nb22*(head0 / gqa_ratio)); - const float * sinks_f = sinks ? (const float *) sinks + head0 : nullptr; + const half2 * V_h2 = V_is_K_view ? K_h2 : (const half2 *) (V + nb23*sequence + nb22*z_KV); + const float * sinks_f = sinks ? (const float *) sinks + zt_Q : nullptr; - const float slope = ncols2 == 1 ? get_alibi_slope(max_bias, head0, n_head_log2, m0, m1) : 1.0f; + const float slope = ncols2 == 1 ? get_alibi_slope(max_bias, zt_Q, n_head_log2, m0, m1) : 1.0f; if (KV_max) { kb0_stop = min(kb0_stop, KV_max[sequence*iter_j + jt] / nbatch_fa); @@ -1561,12 +1566,12 @@ static __global__ void flash_attn_ext_f16( constexpr bool needs_fixup = false; // CUDA block is working on an entire tile. flash_attn_ext_f16_process_tile (Q_f2, K_h2, V_h2, mask_h, sinks_f, dstk, dst_meta, scale, slope, logit_softcap, - ne01, ne02, ne11, stride_Q1, stride_Q2, stride_K, stride_V, stride_mask, jt, kb0_start, kb0_stop); + ne01, ne02, gqa_ratio, ne11, stride_Q1, stride_Q2, stride_K, stride_V, stride_mask, jt, zt_gqa, kb0_start, kb0_stop); } else { constexpr bool needs_fixup = true; // CUDA block is missing the beginning of a tile. flash_attn_ext_f16_process_tile (Q_f2, K_h2, V_h2, mask_h, sinks_f, dstk, dst_meta, scale, slope, logit_softcap, - ne01, ne02, ne11, stride_Q1, stride_Q2, stride_K, stride_V, stride_mask, jt, kb0_start, kb0_stop); + ne01, ne02, gqa_ratio, ne11, stride_Q1, stride_Q2, stride_K, stride_V, stride_mask, jt, zt_gqa, kb0_start, kb0_stop); } kbc += iter_k; @@ -1580,22 +1585,24 @@ static __global__ void flash_attn_ext_f16( return; } - const int sequence = kbc / (iter_k*iter_j*(ne02/ncols2)); - const int zt = (kbc - iter_k*iter_j*(ne02/ncols2)*sequence) / (iter_k*iter_j); // head in units of ncols2 - const int jt = (kbc - iter_k*iter_j*(ne02/ncols2)*sequence - iter_k*iter_j*zt) / iter_k; // j index of current tile. + // z_KV == K/V head index, zt_gqa = Q head start index per K/V head, jt = token position start index. + const int sequence = kbc /(iter_k*iter_j*iter_z_gqa*ne12); + const int z_KV = (kbc - iter_k*iter_j*iter_z_gqa*ne12 * sequence)/(iter_k*iter_j*iter_z_gqa); + const int zt_gqa = (kbc - iter_k*iter_j*iter_z_gqa*ne12 * sequence - iter_k*iter_j*iter_z_gqa * z_KV)/(iter_k*iter_j); + const int jt = (kbc - iter_k*iter_j*iter_z_gqa*ne12 * sequence - iter_k*iter_j*iter_z_gqa * z_KV - iter_k*iter_j * zt_gqa) / iter_k; - const int head0 = zt * ncols2; + const int zt_Q = z_KV*gqa_ratio + zt_gqa*ncols2; // Global Q head start index. - const float2 * Q_f2 = (const float2 *) (Q + nb03*sequence + nb02* head0); - const half2 * K_h2 = (const half2 *) (K + nb13*sequence + nb12*(head0 / gqa_ratio)); + const float2 * Q_f2 = (const float2 *) (Q + nb03*sequence + nb02*zt_Q); + const half2 * K_h2 = (const half2 *) (K + nb13*sequence + nb12*z_KV); const half * mask_h = ncols2 == 1 && !mask ? nullptr : (const half *) (mask + nb33*(sequence % ne33)); - float2 * dstk = ((float2 *) dst) + (sequence*ne01.z*ne02 + head0) * (DV/2); + float2 * dstk = ((float2 *) dst) + (sequence*ne01.z*ne02 + zt_Q) * (DV/2); - const half2 * V_h2 = V_is_K_view ? K_h2 : (const half2 *) (V + nb23*sequence + nb22*(head0 / gqa_ratio)); - const float * sinks_f = sinks ? (const float *) sinks + head0 : nullptr; + const half2 * V_h2 = V_is_K_view ? K_h2 : (const half2 *) (V + nb23*sequence + nb22*z_KV); + const float * sinks_f = sinks ? (const float *) sinks + zt_Q : nullptr; - const float slope = ncols2 == 1 ? get_alibi_slope(max_bias, head0, n_head_log2, m0, m1) : 1.0f; + const float slope = ncols2 == 1 ? get_alibi_slope(max_bias, zt_Q, n_head_log2, m0, m1) : 1.0f; if (KV_max) { kb0_stop = min(kb0_stop, KV_max[sequence*iter_j + jt] / nbatch_fa); @@ -1605,7 +1612,7 @@ static __global__ void flash_attn_ext_f16( constexpr bool needs_fixup = false; flash_attn_ext_f16_process_tile (Q_f2, K_h2, V_h2, mask_h, sinks_f, dstk, dst_meta, scale, slope, logit_softcap, - ne01, ne02, ne11, stride_Q1, stride_Q2, stride_K, stride_V, stride_mask, jt, kb0_start, kb0_stop); + ne01, ne02, gqa_ratio, ne11, stride_Q1, stride_Q2, stride_K, stride_V, stride_mask, jt, zt_gqa, kb0_start, kb0_stop); #else GGML_UNUSED_VARS(Q, K, V, mask, sinks, KV_max, dst, dst_meta, scale, max_bias, m0, m1, n_head_log2, logit_softcap, @@ -1739,3 +1746,5 @@ extern DECL_FATTN_MMA_F16_CASE(576, 512, 4, 16); extern DECL_FATTN_MMA_F16_CASE(576, 512, 4, 4); extern DECL_FATTN_MMA_F16_CASE(576, 512, 8, 4); extern DECL_FATTN_MMA_F16_CASE(576, 512, 16, 4); +extern DECL_FATTN_MMA_F16_CASE(576, 512, 1, 32); +extern DECL_FATTN_MMA_F16_CASE(576, 512, 2, 32); diff --git a/ggml/src/ggml-cuda/fattn.cu b/ggml/src/ggml-cuda/fattn.cu index ba2b96bc32..195904ee20 100644 --- a/ggml/src/ggml-cuda/fattn.cu +++ b/ggml/src/ggml-cuda/fattn.cu @@ -18,9 +18,11 @@ static void ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1(ggml_backend_cuda_con } } - if ((turing_mma_available(cc) || amd_wmma_available(cc)) && Q->ne[1] <= 16/ncols2) { - ggml_cuda_flash_attn_ext_mma_f16_case(ctx, dst); - return; + if constexpr (ncols2 <= 16) { + if ((turing_mma_available(cc) || amd_wmma_available(cc)) && Q->ne[1] <= 16/ncols2) { + ggml_cuda_flash_attn_ext_mma_f16_case(ctx, dst); + return; + } } if (ggml_cuda_highest_compiled_arch(cc) == GGML_CUDA_CC_TURING || amd_wmma_available(cc) || Q->ne[1] <= 32/ncols2) { @@ -33,6 +35,7 @@ static void ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1(ggml_backend_cuda_con template static void ggml_cuda_flash_attn_ext_mma_f16_switch_ncols2(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc; const ggml_tensor * KQV = dst; const ggml_tensor * Q = dst->src[0]; const ggml_tensor * K = dst->src[1]; @@ -60,17 +63,38 @@ static void ggml_cuda_flash_attn_ext_mma_f16_switch_ncols2(ggml_backend_cuda_con GGML_ASSERT(Q->ne[2] % K->ne[2] == 0); const int gqa_ratio = Q->ne[2] / K->ne[2]; - if (use_gqa_opt && gqa_ratio % 8 == 0) { + // On Volta the GQA optimizations aren't as impactful vs. minimizing wasted compute: + if (cc == GGML_CUDA_CC_VOLTA) { + if (use_gqa_opt && gqa_ratio % 8 == 0) { + ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1(ctx, dst); + return; + } + + if (use_gqa_opt && gqa_ratio % 4 == 0) { + ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1(ctx, dst); + return; + } + + if (use_gqa_opt && gqa_ratio % 2 == 0) { + ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1(ctx, dst); + return; + } + + ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1(ctx, dst); + return; + } + + if (use_gqa_opt && gqa_ratio > 4) { ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1(ctx, dst); return; } - if (use_gqa_opt && gqa_ratio % 4 == 0) { + if (use_gqa_opt && gqa_ratio > 2) { ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1(ctx, dst); return; } - if (use_gqa_opt && gqa_ratio % 2 == 0) { + if (use_gqa_opt && gqa_ratio > 1) { ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1(ctx, dst); return; } @@ -79,6 +103,7 @@ static void ggml_cuda_flash_attn_ext_mma_f16_switch_ncols2(ggml_backend_cuda_con } static void ggml_cuda_flash_attn_ext_mma_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc; const ggml_tensor * KQV = dst; const ggml_tensor * Q = dst->src[0]; const ggml_tensor * K = dst->src[1]; @@ -121,8 +146,46 @@ static void ggml_cuda_flash_attn_ext_mma_f16(ggml_backend_cuda_context & ctx, gg GGML_ASSERT(Q->ne[2] % K->ne[2] == 0); const int gqa_ratio = Q->ne[2] / K->ne[2]; - GGML_ASSERT(gqa_ratio % 4 == 0); - if (gqa_ratio % 16 == 0) { + if (gqa_ratio == 20) { // GLM 4.7 Flash + if (cc >= GGML_CUDA_CC_DGX_SPARK) { + if (Q->ne[1] <= 8) { + ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<576, 512, 16>(ctx, dst); + break; + } + ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<576, 512, 4>(ctx, dst); + break; + } + if (cc >= GGML_CUDA_CC_BLACKWELL) { + if (Q->ne[1] <= 4 && K->ne[1] >= 65536) { + ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<576, 512, 16>(ctx, dst); + break; + } + ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<576, 512, 4>(ctx, dst); + break; + } + if (cc >= GGML_CUDA_CC_ADA_LOVELACE) { + if (Q->ne[1] <= 4) { + ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<576, 512, 16>(ctx, dst); + break; + } + ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<576, 512, 4>(ctx, dst); + break; + } + if (cc >= GGML_CUDA_CC_TURING) { + if (Q->ne[1] <= 4) { + if (K->ne[1] <= 16384) { + ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<576, 512, 16>(ctx, dst); + break; + } + ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<576, 512, 32>(ctx, dst); + break; + } + ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<576, 512, 4>(ctx, dst); + break; + } + // Volta: + ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<576, 512, 4>(ctx, dst); + } else if (gqa_ratio % 16 == 0) { ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<576, 512, 16>(ctx, dst); } else { ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<576, 512, 4>(ctx, dst); @@ -234,7 +297,7 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const // The effective batch size for the kernel can be increased by gqa_ratio. // The kernel versions without this optimization are also used for ALiBi, if there is no mask, or if the KV cache is not padded, - bool gqa_opt_applies = gqa_ratio % 2 == 0 && mask && max_bias == 0.0f && K->ne[1] % FATTN_KQ_STRIDE == 0; + bool gqa_opt_applies = gqa_ratio >= 2 && mask && max_bias == 0.0f && K->ne[1] % FATTN_KQ_STRIDE == 0; for (const ggml_tensor * t : {Q, K, V, mask}) { if (t == nullptr || ggml_is_quantized(t->type)) { continue; @@ -247,7 +310,7 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const } } - const bool V_is_K_view = V->op == GGML_OP_VIEW && V->src[0] == K && V->data == K->data; + const bool V_is_K_view = V->view_src && (V->view_src == K || (V->view_src == K->view_src && V->view_offs == K->view_offs)); const int cc = ggml_cuda_info().devices[device].cc; @@ -268,7 +331,7 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const if (V->ne[0] != 512) { return BEST_FATTN_KERNEL_NONE; } - if (!gqa_opt_applies || gqa_ratio % 4 != 0) { + if (!gqa_opt_applies) { return BEST_FATTN_KERNEL_NONE; } if (!V_is_K_view) { diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu index 99f0919a51..76d0f12550 100644 --- a/ggml/src/ggml-cuda/ggml-cuda.cu +++ b/ggml/src/ggml-cuda/ggml-cuda.cu @@ -3080,63 +3080,166 @@ static bool ggml_cuda_should_fuse_rope_set_rows(const ggml_tensor * rope, return true; } -static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx, std::initializer_list ops, std::initializer_list unary_ops) { +static bool ggml_cuda_topk_moe_fusion(const struct ggml_cgraph * cgraph, int node_idx, ggml_cuda_topk_moe_args & args) { + args.sigmoid = false; + args.softmax = false; + args.delayed_softmax = false; + args.prob_bias = false; + args.norm = false; + + const int n_nodes = cgraph->n_nodes; + ggml_tensor ** nodes = cgraph->nodes; + + if (nodes[node_idx]->op == GGML_OP_SOFT_MAX) { + args.softmax = true; + } + + if (nodes[node_idx]->op == GGML_OP_UNARY) { + if (ggml_get_unary_op(nodes[node_idx]) != GGML_UNARY_OP_SIGMOID) { + return false; + } + args.sigmoid = true; + } + + if (nodes[node_idx]->op == GGML_OP_ARGSORT) { + args.delayed_softmax = true; + } + + node_idx++; + + if (args.sigmoid || args.softmax) { + // SOFTMAX -> RESHAPE + if (node_idx >= n_nodes || nodes[node_idx]->op != GGML_OP_RESHAPE || + nodes[node_idx]->src[0] != nodes[node_idx - 1]) { + return false; + } + ggml_tensor * probs_reshaped = nodes[node_idx]; + node_idx++; + + if (node_idx >= n_nodes) { + return false; + } + + // src of bias add is the unreshaped probs (-2 instead of -1) + if (nodes[node_idx]->op == GGML_OP_ADD && nodes[node_idx]->src[0] == nodes[node_idx - 2]) { + args.prob_bias = true; + node_idx++; + } + // RESHAPE/ADD -> ARGSORT + if (node_idx >= n_nodes || nodes[node_idx]->op != GGML_OP_ARGSORT) { + return false; + } + + if (args.prob_bias && nodes[node_idx]->src[0] != nodes[node_idx - 1]) { + return false; + } else if (!args.prob_bias && nodes[node_idx]->src[0] != nodes[node_idx - 2]) { + return false; + } + + node_idx++; + + // ARGSORT-> VIEW + if (node_idx >= n_nodes || nodes[node_idx]->op != GGML_OP_VIEW || + nodes[node_idx]->src[0] != nodes[node_idx - 1]) { + return false; + } + node_idx++; + + if (node_idx >= n_nodes || nodes[node_idx]->op != GGML_OP_GET_ROWS) { + return false; + } + + // GET_ROWS + if (nodes[node_idx]->src[0] != probs_reshaped || nodes[node_idx]->src[1] != nodes[node_idx - 1]) { + return false; + } + node_idx++; + } else if (args.delayed_softmax) { + if (node_idx - 2 < 0) { + return false; + } + ggml_tensor * probs_reshaped = nodes[node_idx - 2]; + + // VIEW->ARGSORT + if (node_idx >= n_nodes || nodes[node_idx]->op != GGML_OP_VIEW || + nodes[node_idx]->src[0] != nodes[node_idx - 1]) { + return false; + } + node_idx++; + + // GET_ROWS + if (node_idx >= n_nodes || nodes[node_idx]->src[1] != nodes[node_idx - 1] || + nodes[node_idx]->src[0] != probs_reshaped) { + return false; + } + node_idx++; + + static const std::vector remaining_ops = { GGML_OP_RESHAPE, GGML_OP_SOFT_MAX, GGML_OP_RESHAPE }; + + for (const ggml_op op : remaining_ops) { + if (node_idx >= n_nodes || nodes[node_idx]->op != op || nodes[node_idx]->src[0] != nodes[node_idx - 1]) { + return false; + } + node_idx++; + } + } + + // At this point we can check for norm + scale. Everything is now at least valid till the norm + if (node_idx >= n_nodes) { + return true; + } + + if (nodes[node_idx]->op == GGML_OP_RESHAPE) { + //check RESHAPE->SUM_ROWS->CLAMP->DIV->RESHAPE + static const std::vector norm_ops = { GGML_OP_RESHAPE, GGML_OP_SUM_ROWS, GGML_OP_CLAMP }; + + args.norm = true; + for (const ggml_op op : norm_ops) { + if (nodes[node_idx]->op == op && nodes[node_idx]->src[0] == nodes[node_idx - 1]) { + node_idx++; + } else { + args.norm = false; + return true; + } + } + + // DIV <- CLAMP, RESHAPE + if (nodes[node_idx]->op != GGML_OP_DIV || nodes[node_idx]->src[1] != nodes[node_idx - 1] || + nodes[node_idx]->src[0] != nodes[node_idx - 3]) { + args.norm = false; + return true; + } + node_idx++; + + if (nodes[node_idx]->op != GGML_OP_RESHAPE || nodes[node_idx]->src[0] != nodes[node_idx - 1]) { + args.norm = false; + return true; + } + + node_idx++; + } + + if (nodes[node_idx]->op == GGML_OP_SCALE && nodes[node_idx]->src[0] == nodes[node_idx - 1]) { + args.scale = true; + } + + return true; +} + +static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, + int node_idx, + std::initializer_list ops, + std::initializer_list unary_ops) { #ifndef NDEBUG const size_t num_unary = std::count(ops.begin(), ops.end(), GGML_OP_UNARY); GGML_ASSERT(unary_ops.size() == num_unary); #endif - //TODO: remove special case once ggml_can_fuse can handle empty nodes - std::initializer_list topk_moe_ops = - ggml_cuda_topk_moe_ops(/*with_norm*/ false, /*delayed_softmax=*/false); - std::initializer_list topk_moe_ops_with_norm = - ggml_cuda_topk_moe_ops(/*with_norm=*/true, /*delayed_softmax=*/false); - std::initializer_list topk_moe_ops_delayed_softmax = - ggml_cuda_topk_moe_ops(/*with_norm=*/false, /*delayed_softmax=*/true); - const auto is_equal = [](const std::initializer_list & list1, const std::initializer_list & list2) { return std::equal(list1.begin(), list1.end(), list2.begin(), list2.end()); }; - if (is_equal(topk_moe_ops_with_norm, ops) && - ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 3, node_idx + 9 })) { - ggml_tensor * softmax = cgraph->nodes[node_idx]; - ggml_tensor * weights = cgraph->nodes[node_idx + 9]; - ggml_tensor * get_rows = cgraph->nodes[node_idx + 4]; - ggml_tensor * argsort = cgraph->nodes[node_idx + 2]; - int n_expert = cgraph->nodes[node_idx]->src[0]->ne[0]; - - if (ggml_cuda_should_use_topk_moe(softmax, weights, get_rows, argsort, nullptr, n_expert)) { - return true; - } - } - - if (is_equal(topk_moe_ops, ops) && ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 3, node_idx + 4 })) { - ggml_tensor * softmax = cgraph->nodes[node_idx]; - ggml_tensor * weights = cgraph->nodes[node_idx + 4]; - ggml_tensor * get_rows = cgraph->nodes[node_idx + 4]; - ggml_tensor * argsort = cgraph->nodes[node_idx + 2]; - int n_expert = cgraph->nodes[node_idx]->src[0]->ne[0]; - - if (ggml_cuda_should_use_topk_moe(softmax, weights, get_rows, argsort, nullptr, n_expert)) { - return true; - } - } - - if (is_equal(topk_moe_ops_delayed_softmax, ops) && - ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 1, node_idx + 5 })) { - ggml_tensor * softmax = cgraph->nodes[node_idx + 4]; - ggml_tensor * weights = cgraph->nodes[node_idx + 5]; - ggml_tensor * get_rows = cgraph->nodes[node_idx + 2]; - ggml_tensor * argsort = cgraph->nodes[node_idx + 0]; - int n_expert = cgraph->nodes[node_idx]->src[0]->ne[0]; - - if (ggml_cuda_should_use_topk_moe(softmax, weights, get_rows, argsort, nullptr, n_expert)) { - return true; - } - } - std::initializer_list mul_mat_bias_glu_ops = { GGML_OP_MUL_MAT, GGML_OP_ADD, GGML_OP_MUL_MAT, GGML_OP_ADD, GGML_OP_GLU }; std::initializer_list mul_mat_id_bias_glu_ops = { GGML_OP_MUL_MAT_ID, GGML_OP_ADD_ID, GGML_OP_MUL_MAT_ID, GGML_OP_ADD_ID, GGML_OP_GLU }; @@ -3398,35 +3501,75 @@ static void ggml_cuda_graph_evaluate_and_capture(ggml_backend_cuda_context * cud // start of fusion operations static bool disable_fusion = (getenv("GGML_CUDA_DISABLE_FUSION") != nullptr); if (!disable_fusion) { + ggml_cuda_topk_moe_args args; - if (ggml_cuda_can_fuse(cgraph, i, ggml_cuda_topk_moe_ops(/*with norm*/ true), {})) { - ggml_tensor * weights = cgraph->nodes[i + 9]; - ggml_tensor * selected_experts = cgraph->nodes[i + 3]; - ggml_tensor * clamp = cgraph->nodes[i + 7]; - ggml_cuda_op_topk_moe(*cuda_ctx, node->src[0], weights, selected_experts, /*with norm*/ true, - /*delayed softmax*/ false, clamp); - i += 9; - continue; - } + if (cgraph->nodes[i]->op == GGML_OP_UNARY || cgraph->nodes[i]->op == GGML_OP_SOFT_MAX || + cgraph->nodes[i]->op == GGML_OP_ARGSORT) { + const bool can_fuse = ggml_cuda_topk_moe_fusion(cgraph, i, args); - if (ggml_cuda_can_fuse(cgraph, i, ggml_cuda_topk_moe_ops(/*with norm*/ false), {})) { - ggml_tensor * weights = cgraph->nodes[i + 4]; - ggml_tensor * selected_experts = cgraph->nodes[i + 3]; - ggml_cuda_op_topk_moe(*cuda_ctx, node->src[0], weights, selected_experts, /*with norm*/ false, - /*delayed softmax*/ false); - i += 4; - continue; - } + std::vector ops; - if (ggml_cuda_can_fuse(cgraph, i, - ggml_cuda_topk_moe_ops(/*with norm*/ false, /*delayed softmax*/ true), {})) { - ggml_tensor * weights = cgraph->nodes[i + 5]; - ggml_tensor * ids = cgraph->nodes[i + 1]; + if (can_fuse) { + const ggml_tensor * logits = node->src[0]; + ggml_tensor * weights = nullptr; + ggml_tensor * ids = nullptr; + const ggml_tensor * bias = nullptr; + const ggml_tensor * clamp = nullptr; + const ggml_tensor * scale = nullptr; - ggml_cuda_op_topk_moe(*cuda_ctx, node->src[0], weights, ids, /*with norm*/ false, - /*delayed_softmax*/ true); - i += 5; - continue; + if (!args.delayed_softmax) { + ggml_op gating_op = args.sigmoid ? GGML_OP_UNARY : GGML_OP_SOFT_MAX; + int out_nodes[2]; // nodes which can't be elided + + if (args.prob_bias) { + bias = cgraph->nodes[i + 2]->src[1]; + ops.insert(ops.end(), { gating_op, GGML_OP_RESHAPE, GGML_OP_ADD, GGML_OP_ARGSORT, + GGML_OP_VIEW, GGML_OP_GET_ROWS }); + out_nodes[0] = i + 4; + ids = cgraph->nodes[i + 4]; + } else { + ops.insert(ops.end(), { gating_op, GGML_OP_RESHAPE, GGML_OP_ARGSORT, GGML_OP_VIEW, + GGML_OP_GET_ROWS }); + out_nodes[0] = i + 3; + ids = cgraph->nodes[i + 3]; + } + + if (args.norm) { + ops.insert(ops.end(), { GGML_OP_RESHAPE, GGML_OP_SUM_ROWS, GGML_OP_CLAMP, + GGML_OP_DIV, GGML_OP_RESHAPE }); + clamp = cgraph->nodes[i + ops.size() - 3]; + } + if (args.scale) { + ops.insert(ops.end(), { GGML_OP_SCALE }); + scale = cgraph->nodes[i + ops.size() - 1]; + } + + weights = cgraph->nodes[i + ops.size() - 1]; + out_nodes[1] = i + ops.size() - 1; + + if (ggml_can_fuse_subgraph(cgraph, i, ops.size(), ops.data(), out_nodes, 2) && + ggml_cuda_should_use_topk_moe(node, logits, weights, ids)) { + ggml_cuda_op_topk_moe(*cuda_ctx, logits, weights, ids, clamp, scale, bias, args); + i += ops.size() - 1; + continue; + } + } else if (!args.norm && !args.prob_bias) { + //special case gpt-oss, no norm, no bias. + ops.insert(ops.end(), { GGML_OP_ARGSORT, GGML_OP_VIEW, GGML_OP_GET_ROWS, + GGML_OP_RESHAPE, GGML_OP_SOFT_MAX, GGML_OP_RESHAPE }); + weights = cgraph->nodes[i + 5]; + ids = cgraph->nodes[i + 1]; + const ggml_tensor * softmax = cgraph->nodes[i + 4]; + + int out_nodes[2] = { i + 1, i + 5 }; + if (ggml_can_fuse_subgraph(cgraph, i, ops.size(), ops.data(), out_nodes, 2) && + ggml_cuda_should_use_topk_moe(softmax, logits, weights, ids)) { + ggml_cuda_op_topk_moe(*cuda_ctx, logits, weights, ids, clamp, scale, bias, args); + i += ops.size() - 1; + continue; + } + } + } } if (ggml_cuda_can_fuse(cgraph, i, { GGML_OP_ROPE, GGML_OP_VIEW, GGML_OP_SET_ROWS }, {})) { @@ -4876,6 +5019,16 @@ ggml_backend_reg_t ggml_backend_cuda_reg() { static std::mutex mutex; std::lock_guard lock(mutex); if (!initialized) { + // Set CUDA_SCALE_LAUNCH_QUEUES before any CUDA API call to improve multi-GPU pipeline parallelism performance + // PR: https://github.com/ggml-org/llama.cpp/pull/19042 + if (getenv("CUDA_SCALE_LAUNCH_QUEUES") == nullptr) { +#ifdef _WIN32 + _putenv_s("CUDA_SCALE_LAUNCH_QUEUES", "4x"); +#else + setenv("CUDA_SCALE_LAUNCH_QUEUES", "4x", 0); // don't overwrite if already set +#endif // _WIN32 + } + ggml_backend_cuda_reg_context * ctx = new ggml_backend_cuda_reg_context; const int min_batch_size = getenv("GGML_OP_OFFLOAD_MIN_BATCH") ? atoi(getenv("GGML_OP_OFFLOAD_MIN_BATCH")) : 32; diff --git a/ggml/src/ggml-cuda/mma.cuh b/ggml/src/ggml-cuda/mma.cuh index 42085d1002..dd45d6c78f 100644 --- a/ggml/src/ggml-cuda/mma.cuh +++ b/ggml/src/ggml-cuda/mma.cuh @@ -333,7 +333,33 @@ namespace ggml_cuda_mma { static __device__ __forceinline__ int get_j(const int l) { if constexpr (I == 16 && J == 8) { - return 4 * (threadIdx.x / 16) + l; + return ne * (threadIdx.x / 16) + l; + } else { + NO_DEVICE_CODE; + return -1; + } + } +#elif defined(AMD_MFMA_AVAILABLE) + static constexpr int ne = I * J / 64; + half2 x[ne] = {{0.0f, 0.0f}}; + + static constexpr __device__ bool supported() { + if (I == 16 && J == 8) return true; + return false; + } + + static __device__ __forceinline__ int get_i(const int l) { + if constexpr (I == 16 && J == 8) { + return threadIdx.x % 16; + } else { + NO_DEVICE_CODE; + return -1; + } + } + + static __device__ __forceinline__ int get_j(const int l) { + if constexpr (I == 16 && J == 8) { + return ne * (threadIdx.x / 16) + l; } else { NO_DEVICE_CODE; return -1; @@ -391,7 +417,22 @@ namespace ggml_cuda_mma { static constexpr data_layout dl = DATA_LAYOUT_I_MAJOR; #if defined(AMD_WMMA_AVAILABLE) - static constexpr int ne = I * J / 32; + static constexpr int ne = tile::ne; + nv_bfloat162 x[ne] = {{0.0f, 0.0f}}; + + static constexpr __device__ bool supported() { + return tile::supported(); + } + + static __device__ __forceinline__ int get_i(const int l) { + return tile::get_i(l); + } + + static __device__ __forceinline__ int get_j(const int l) { + return tile::get_j(l); + } +#elif defined(AMD_MFMA_AVAILABLE) + static constexpr int ne = tile::ne; nv_bfloat162 x[ne] = {{0.0f, 0.0f}}; static constexpr __device__ bool supported() { @@ -945,6 +986,32 @@ namespace ggml_cuda_mma { #endif // AMPERE_MMA_AVAILABLE } + template + static __device__ __forceinline__ void mma( + tile<16, 16, float, dl_d> & D, const tile<16, 8, float, dl_ab> & A, const tile<16, 8, float, dl_ab> & B) { +#ifdef AMD_MFMA_AVAILABLE + using floatx4_t = __attribute__((ext_vector_type(4))) float; + floatx4_t& acc_frag = reinterpret_cast(D.x[0]); +#if defined(CDNA3) + using floatx2_t = __attribute__((ext_vector_type(2))) float; + const floatx2_t& a_frag = reinterpret_cast(A.x[0]); + const floatx2_t& b_frag = reinterpret_cast(B.x[0]); + acc_frag = __builtin_amdgcn_mfma_f32_16x16x8_xf32(a_frag, b_frag, acc_frag, 0, 0, 0); +#elif defined(CDNA2) || defined(CDNA1) +#pragma unroll + for (int i = 0; i < 2; ++i) { + acc_frag = __builtin_amdgcn_mfma_f32_16x16x4f32(A.x[i], B.x[i], acc_frag, 0, 0, 0); + } +#else + GGML_UNUSED_VARS(D, A, B); + NO_DEVICE_CODE; +#endif // defined(CDNA3) +#else + GGML_UNUSED_VARS(D, A, B); + NO_DEVICE_CODE; +#endif // AMD_MFMA_AVAILABLE + } + static __device__ __forceinline__ void mma_block_scaled(tile<16, 8, float> & D, const tile<16, 8, int> & A, const tile<8, 8, int> & B, @@ -1054,6 +1121,13 @@ namespace ggml_cuda_mma { GGML_UNUSED_VARS(D, A, B); NO_DEVICE_CODE; #endif // RDNA4 +#elif defined(AMD_MFMA_AVAILABLE) + using halfx4_t = __attribute__((ext_vector_type(4))) _Float16; + using floatx4_t = __attribute__((ext_vector_type(4))) float; + floatx4_t& acc_frag = reinterpret_cast(D.x[0]); + const halfx4_t& a_frag = reinterpret_cast(A.x[0]); + const halfx4_t& b_frag = reinterpret_cast(B.x[0]); + acc_frag = __builtin_amdgcn_mfma_f32_16x16x16f16(a_frag, b_frag, acc_frag, 0, 0, 0); #else GGML_UNUSED_VARS(D, A, B); NO_DEVICE_CODE; @@ -1081,11 +1155,31 @@ namespace ggml_cuda_mma { #else GGML_UNUSED_VARS(D, A, B); NO_DEVICE_CODE; -#endif // RDNA4 +#endif // defined(RDNA4) +#elif defined(AMD_MFMA_AVAILABLE) + using floatx4_t = __attribute__((ext_vector_type(4))) float; + floatx4_t& acc_frag = reinterpret_cast(D.x[0]); +#if defined(CDNA3) || defined(CDNA2) + using bf16x4_t = __attribute__((ext_vector_type(4))) __bf16; + const bf16x4_t& a_frag = reinterpret_cast(A.x[0]); + const bf16x4_t& b_frag = reinterpret_cast(B.x[0]); + acc_frag = __builtin_amdgcn_mfma_f32_16x16x16bf16_1k(a_frag, b_frag, acc_frag, 0, 0, 0); +#elif defined(CDNA1) +#pragma unroll + for (int i = 0; i < 2; ++i) { + using bf16x2_t = __attribute__((ext_vector_type(2))) __bf16; + const bf16x2_t& a_frag = reinterpret_cast(A.x[i]); + const bf16x2_t& b_frag = reinterpret_cast(B.x[i]); + acc_frag = __builtin_amdgcn_mfma_f32_16x16x8bf16(a_frag, b_frag, acc_frag, 0, 0, 0); + } #else GGML_UNUSED_VARS(D, A, B); NO_DEVICE_CODE; -#endif // AMPERE_MMA_AVAILABLE +#endif // defined(CDNA3) || defined(CDNA2) +#else + GGML_UNUSED_VARS(D, A, B); + NO_DEVICE_CODE; +#endif // defined(AMD_WMMA_AVAILABLE) } template diff --git a/ggml/src/ggml-cuda/mmf.cu b/ggml/src/ggml-cuda/mmf.cu index 6643f243b1..aad4c34aa6 100644 --- a/ggml/src/ggml-cuda/mmf.cu +++ b/ggml/src/ggml-cuda/mmf.cu @@ -2,6 +2,13 @@ #include "mmf.cuh" #include "mmid.cuh" +static __forceinline__ int mmf_get_rows_per_block(const int cc) { + if (GGML_CUDA_CC_IS_CDNA(cc)) { + return MMF_ROWS_PER_BLOCK_CDNA; + } else { + return MMF_ROWS_PER_BLOCK; + } +} void ggml_cuda_mul_mat_f(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst) { GGML_ASSERT( src1->type == GGML_TYPE_F32); @@ -89,28 +96,32 @@ void ggml_cuda_mul_mat_f(ggml_backend_cuda_context & ctx, const ggml_tensor * sr ids_info_ptr = &ids_info; } + const int device = ggml_cuda_get_device(); + const int cc = ggml_cuda_info().devices[device].cc; + const int rows_per_block = mmf_get_rows_per_block(cc); + switch (src0->type) { case GGML_TYPE_F32: { const float * src0_d = (const float *) src0->data; constexpr int vals_per_T = 1; - mul_mat_f_switch_cols_per_block( - src0_d, src1_d, ids_d, dst_d, ne00/vals_per_T, ne01, ncols_dst, s01/vals_per_T, stride_col_y/vals_per_T, stride_col_dst, + mul_mat_f_switch_rows_per_block( + rows_per_block, src0_d, src1_d, ids_d, dst_d, ne00/vals_per_T, ne01, ncols_dst, s01/vals_per_T, stride_col_y/vals_per_T, stride_col_dst, ids_s0, ids_s1, ne02, nchannels_y, nchannels_dst, s02/vals_per_T, stride_channel_y, stride_channel_dst, ne03, ne3, s03/vals_per_T, s13, s3, ctx.stream(), ids_info_ptr); } break; case GGML_TYPE_F16: { const half2 * src0_d = (const half2 *) src0->data; constexpr int vals_per_T = 2; - mul_mat_f_switch_cols_per_block( - src0_d, src1_d, ids_d, dst_d, ne00/vals_per_T, ne01, ncols_dst, s01/vals_per_T, stride_col_y/vals_per_T, stride_col_dst, + mul_mat_f_switch_rows_per_block( + rows_per_block, src0_d, src1_d, ids_d, dst_d, ne00/vals_per_T, ne01, ncols_dst, s01/vals_per_T, stride_col_y/vals_per_T, stride_col_dst, ids_s0, ids_s1, ne02, nchannels_y, nchannels_dst, s02/vals_per_T, stride_channel_y, stride_channel_dst, ne03, ne3, s03/vals_per_T, s13, s3, ctx.stream(), ids_info_ptr); } break; case GGML_TYPE_BF16: { const nv_bfloat162 * src0_d = (const nv_bfloat162 *) src0->data; constexpr int vals_per_T = 2; - mul_mat_f_switch_cols_per_block( - src0_d, src1_d, ids_d, dst_d, ne00/vals_per_T, ne01, ncols_dst, s01/vals_per_T, stride_col_y/vals_per_T, stride_col_dst, + mul_mat_f_switch_rows_per_block( + rows_per_block, src0_d, src1_d, ids_d, dst_d, ne00/vals_per_T, ne01, ncols_dst, s01/vals_per_T, stride_col_y/vals_per_T, stride_col_dst, ids_s0, ids_s1, ne02, nchannels_y, nchannels_dst, s02/vals_per_T, stride_channel_y, stride_channel_dst, ne03, ne3, s03/vals_per_T, s13, s3, ctx.stream(), ids_info_ptr); } break; @@ -140,7 +151,11 @@ bool ggml_cuda_should_use_mmf(enum ggml_type type, int cc, int warp_size, const return false; } } - if (src0_ne[1] % MMF_ROWS_PER_BLOCK != 0) { + if (src0_ne[1] % mmf_get_rows_per_block(cc) != 0) { + return false; + } + + if (GGML_CUDA_CC_IS_CDNA3(cc) && type == GGML_TYPE_BF16) { return false; } @@ -153,6 +168,11 @@ bool ggml_cuda_should_use_mmf(enum ggml_type type, int cc, int warp_size, const } else { if (GGML_CUDA_CC_IS_RDNA3_0(cc) && src1_ncols > 8) { return false; + } else if (GGML_CUDA_CC_IS_CDNA2(cc) && (type == GGML_TYPE_F16 || type == GGML_TYPE_BF16)) { + //TODO: truse CDNA2 as CDNA1, tune the perf when CDNA2 is available. + return false; + } else if (GGML_CUDA_CC_IS_CDNA1(cc) && (type == GGML_TYPE_F16 || type == GGML_TYPE_BF16)) { + return false; } else if (src1_ncols > 16) { return false; } @@ -160,11 +180,11 @@ bool ggml_cuda_should_use_mmf(enum ggml_type type, int cc, int warp_size, const switch (type) { case GGML_TYPE_F32: - return ampere_mma_available(cc); + return ampere_mma_available(cc) || amd_mfma_available(cc); case GGML_TYPE_F16: - return volta_mma_available(cc) || turing_mma_available(cc) || amd_wmma_available(cc); + return volta_mma_available(cc) || turing_mma_available(cc) || amd_wmma_available(cc) || amd_mfma_available(cc); case GGML_TYPE_BF16: - return ampere_mma_available(cc) || amd_wmma_available(cc); + return ampere_mma_available(cc) || amd_wmma_available(cc) || amd_mfma_available(cc); default: return false; } diff --git a/ggml/src/ggml-cuda/mmf.cuh b/ggml/src/ggml-cuda/mmf.cuh index e36730948f..c2a8d54c95 100644 --- a/ggml/src/ggml-cuda/mmf.cuh +++ b/ggml/src/ggml-cuda/mmf.cuh @@ -7,6 +7,31 @@ using namespace ggml_cuda_mma; #define MMF_ROWS_PER_BLOCK 32 +#define MMF_ROWS_PER_BLOCK_CDNA 64 + +static __forceinline__ int64_t mmf_get_max_block_size(int cc) { + if (GGML_CUDA_CC_IS_CDNA(cc)) { + return 512; + } else { + return 256; + } +} + +static __forceinline__ int mmf_get_padding(int cc) { + if (GGML_CUDA_CC_IS_CDNA(cc)) { + return 2; + } else { + return 4; + } +} + +static constexpr __device__ int mmf_get_padding() { +#if defined(AMD_MFMA_AVAILABLE) + return 2; +#else + return 4; +#endif // defined(AMD_MFMA_AVAILABLE) +} struct mmf_ids_data { const int32_t * ids_src_compact = nullptr; @@ -29,23 +54,25 @@ static __global__ void mul_mat_f( const int channel_ratio, const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst, const int sample_ratio, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst) { // TODO: handle this in a consistent and simpler way after AMD MFMA support has been added -#if (!defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)) || defined(AMD_WMMA_AVAILABLE) +#if defined(VOLTA_MMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) || defined(AMD_MFMA_AVAILABLE) #if defined(AMD_WMMA_AVAILABLE) - // Special case for tf32, just dummy mma layout as wmma doesn't support it. - constexpr bool is_tf32 = std::is_same_v; - constexpr int tile_B_I = is_tf32 ? 8 : 16; - constexpr int tile_C_J = is_tf32 ? 8 : 16; - constexpr data_layout ab_layout = is_tf32 ? DATA_LAYOUT_I_MAJOR : get_input_data_layout(); - typedef tile<16, 8, T, ab_layout> tile_A; - typedef tile tile_B; - typedef tile<16, tile_C_J, float, DATA_LAYOUT_J_MAJOR> tile_C; + if constexpr (!(std::is_same_v || std::is_same_v) || rows_per_block != MMF_ROWS_PER_BLOCK) {NO_DEVICE_CODE;} else { + typedef tile<16, 8, T, get_input_data_layout()> tile_A; + typedef tile<16, 8, T, get_input_data_layout()> tile_B; + typedef tile<16, 16, float, DATA_LAYOUT_J_MAJOR> tile_C; +#elif defined(AMD_MFMA_AVAILABLE) + if constexpr (rows_per_block != MMF_ROWS_PER_BLOCK_CDNA) {NO_DEVICE_CODE;} else { + typedef tile<16, 8, T, DATA_LAYOUT_I_MAJOR> tile_A; + typedef tile<16, 8, T, DATA_LAYOUT_I_MAJOR> tile_B; + typedef tile<16, 16, float, DATA_LAYOUT_J_MAJOR> tile_C; #else #ifdef VOLTA_MMA_AVAILABLE - if constexpr (!std::is_same_v) {NO_DEVICE_CODE;} else { + if constexpr (!std::is_same_v || rows_per_block != MMF_ROWS_PER_BLOCK) {NO_DEVICE_CODE;} else { typedef tile<32, 4, T, DATA_LAYOUT_I_MAJOR> tile_A; typedef tile< 8, 4, T, DATA_LAYOUT_I_MAJOR_MIRRORED> tile_B; typedef tile<32, 8, float, DATA_LAYOUT_I_MAJOR> tile_C; #else + if constexpr (rows_per_block != MMF_ROWS_PER_BLOCK) {NO_DEVICE_CODE;} else { typedef tile<16, 8, T> tile_A; typedef tile<8, 8, T> tile_B; typedef tile<16, 8, float> tile_C; @@ -57,7 +84,7 @@ static __global__ void mul_mat_f( } constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - constexpr int tile_k_padded = warp_size + 4; + constexpr int tile_k_padded = warp_size + mmf_get_padding(); constexpr int ntA = rows_per_block / tile_A::I; constexpr int ntB = (cols_per_block + tile_B::I - 1) / tile_B::I; @@ -198,7 +225,7 @@ static __global__ void mul_mat_f( } float * buf_iw = (float *) compute_base; - constexpr int kiw = nwarps*rows_per_block + 4; + constexpr int kiw = nwarps*rows_per_block + mmf_get_padding(); if (nwarps > 1) { __syncthreads(); @@ -228,27 +255,34 @@ static __global__ void mul_mat_f( return; } - float sum = 0.0f; - static_assert(rows_per_block == warp_size, "need loop/check"); + float sum[rows_per_block/warp_size] = {0.0f}; + static_assert((rows_per_block % warp_size) == 0, "rows_per_block must be a multiple of warp_size."); #pragma unroll for (int i0 = 0; i0 < nwarps*rows_per_block; i0 += rows_per_block) { - const int i = i0 + threadIdx.x; +#pragma unroll + for (int i1 = 0; i1 < sizeof(sum)/sizeof(sum[0]); ++i1) { + const int i = i0 + i1*warp_size + threadIdx.x; - sum += buf_iw[j*kiw + i]; + sum[i1] += buf_iw[j*kiw + i]; + } } if constexpr (!has_ids) { - dst[j*stride_col_dst + row0 + threadIdx.x] = sum; +#pragma unroll + for (int i0 = 0; i0 < sizeof(sum)/sizeof(sum[0]); ++i0) { + dst[j*stride_col_dst + row0 + i0*warp_size + threadIdx.x] = sum[i0]; + } } else { const int slot = (j < cols_per_block) ? slot_map[j] : -1; if (slot >= 0 && (col_base + j) < ncols_dst_total) { - dst[slot*stride_channel_dst + j*stride_col_dst + row0 + threadIdx.x] = sum; +#pragma unroll + for (int i0 = 0; i0 < sizeof(sum)/sizeof(sum[0]); ++i0) { + dst[slot*stride_channel_dst + j*stride_col_dst + row0 + i0*warp_size + threadIdx.x] = sum[i0]; + } } } } -#ifdef VOLTA_MMA_AVAILABLE } -#endif //VOLTA_MMA_AVAILABLE #else GGML_UNUSED_VARS(x, y, ids, dst, ncols, ncols_dst_total, nchannels_dst, stride_row, stride_col_y, stride_col_dst, @@ -256,7 +290,7 @@ static __global__ void mul_mat_f( channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst, sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst); NO_DEVICE_CODE; -#endif // (!defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)) || defined(AMD_WMMA_AVAILABLE) +#endif // defined(VOLTA_MMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) || defined(AMD_MFMA_AVAILABLE) } //This kernel is for larger batch sizes of mul_mat_id @@ -271,23 +305,25 @@ static __global__ void mul_mat_f_ids( const int sample_ratio, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst, const uint3 sis1_fd, const uint3 nch_fd) { // TODO: handle this in a consistent and simpler way after AMD MFMA support has been added -#if (!defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)) || defined(AMD_WMMA_AVAILABLE) +#if defined(VOLTA_MMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) || defined(AMD_MFMA_AVAILABLE) #if defined(AMD_WMMA_AVAILABLE) - // Special case for tf32, just dummy mma layout as wmma doesn't support it. - constexpr bool is_tf32 = std::is_same_v; - constexpr int tile_B_I = is_tf32 ? 8 : 16; - constexpr int tile_C_J = is_tf32 ? 8 : 16; - constexpr data_layout ab_layout = is_tf32 ? DATA_LAYOUT_I_MAJOR : get_input_data_layout(); - typedef tile<16, 8, T, ab_layout> tile_A; - typedef tile tile_B; - typedef tile<16, tile_C_J, float, DATA_LAYOUT_J_MAJOR> tile_C; + if constexpr (!(std::is_same_v || std::is_same_v) || rows_per_block != MMF_ROWS_PER_BLOCK) {NO_DEVICE_CODE;} else { + typedef tile<16, 8, T, get_input_data_layout()> tile_A; + typedef tile<16, 8, T, get_input_data_layout()> tile_B; + typedef tile<16, 16, float, DATA_LAYOUT_J_MAJOR> tile_C; +#elif defined(AMD_MFMA_AVAILABLE) + if constexpr (rows_per_block != MMF_ROWS_PER_BLOCK_CDNA) {NO_DEVICE_CODE;} else { + typedef tile<16, 8, T, DATA_LAYOUT_I_MAJOR> tile_A; + typedef tile<16, 8, T, DATA_LAYOUT_I_MAJOR> tile_B; + typedef tile<16, 16, float, DATA_LAYOUT_J_MAJOR> tile_C; #else #ifdef VOLTA_MMA_AVAILABLE - if constexpr (!std::is_same_v) {NO_DEVICE_CODE;} else { + if constexpr (!std::is_same_v || rows_per_block != MMF_ROWS_PER_BLOCK) {NO_DEVICE_CODE;} else { typedef tile<32, 4, T, DATA_LAYOUT_I_MAJOR> tile_A; typedef tile< 8, 4, T, DATA_LAYOUT_I_MAJOR_MIRRORED> tile_B; typedef tile<32, 8, float, DATA_LAYOUT_I_MAJOR> tile_C; #else + if constexpr (rows_per_block != MMF_ROWS_PER_BLOCK) {NO_DEVICE_CODE;} else { typedef tile<16, 8, T> tile_A; typedef tile<8, 8, T> tile_B; typedef tile<16, 8, float> tile_C; @@ -300,7 +336,7 @@ static __global__ void mul_mat_f_ids( constexpr int warp_size = ggml_cuda_get_physical_warp_size(); - constexpr int tile_k_padded = warp_size + 4; + constexpr int tile_k_padded = warp_size + mmf_get_padding(); constexpr int ntA = rows_per_block / tile_A::I; constexpr int ntB = (cols_per_block + tile_B::I - 1) / tile_B::I; @@ -467,7 +503,7 @@ static __global__ void mul_mat_f_ids( } float * buf_iw = (float *) compute_base; - constexpr int kiw = nwarps*rows_per_block + 4; + constexpr int kiw = nwarps*rows_per_block + mmf_get_padding(); if (nwarps > 1) { __syncthreads(); @@ -497,13 +533,16 @@ static __global__ void mul_mat_f_ids( return; } - float sum = 0.0f; - static_assert(rows_per_block == warp_size, "need loop/check"); + float sum[rows_per_block/warp_size] = {0.0f}; + static_assert((rows_per_block % warp_size) == 0, "rows_per_block must be a multiple of warp_size."); #pragma unroll for (int i0 = 0; i0 < nwarps*rows_per_block; i0 += rows_per_block) { - const int i = i0 + threadIdx.x; +#pragma unroll + for (int i1 = 0; i1 < sizeof(sum)/sizeof(sum[0]); ++i1) { + const int i = i0 + i1*warp_size + threadIdx.x; - sum += buf_iw[j*kiw + i]; + sum[i1] += buf_iw[j * kiw + i]; + } } const int global_j = col_base + j; @@ -513,23 +552,24 @@ static __global__ void mul_mat_f_ids( const int token = (int) qrm.x; if (token < ncols_dst_total) { const int slot = (int) qrm.y; - dst[slot*stride_channel_dst + token*stride_col_dst + row0 + threadIdx.x] = sum; +#pragma unroll + for (int i0 = 0; i0 < sizeof(sum)/sizeof(sum[0]); ++i0) { + dst[slot * stride_channel_dst + token * stride_col_dst + row0 + i0*warp_size + threadIdx.x] = sum[i0]; + } } } } -#ifdef VOLTA_MMA_AVAILABLE } -#endif // VOLTA_MMA_AVAILABLE #else GGML_UNUSED_VARS(x, y, ids_src_compact, ids_dst_compact, expert_bounds, dst, ncols, ncols_dst_total, nchannels_dst, stride_row, stride_col_y, stride_col_dst, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst, sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, sis1_fd, nch_fd); NO_DEVICE_CODE; -#endif // (!defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)) || defined(AMD_WMMA_AVAILABLE) +#endif // defined(VOLTA_MMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) || defined(AMD_MFMA_AVAILABLE) } -template +template static inline void mul_mat_f_switch_ids( const T * x, const float * y, const int32_t * ids, float * dst, const int64_t ncols_x, const int64_t ncols_dst, const int64_t nchannels_dst, @@ -553,7 +593,7 @@ static inline void mul_mat_f_switch_ids( const uint3 sis1_fd = ids_data->sis1 > 0 ? init_fastdiv_values((uint32_t) ids_data->sis1) : make_uint3(0, 0, 1); const uint3 nch_fd = init_fastdiv_values((uint32_t) nchannels_dst); - mul_mat_f_ids<<>> + mul_mat_f_ids<<>> (x, y, ids_data->ids_src_compact, ids_data->ids_dst_compact, ids_data->expert_bounds_dev, dst, ncols_x, ncols_dst, nchannels_dst, stride_row, stride_col_y, stride_col_dst, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst, @@ -564,19 +604,19 @@ static inline void mul_mat_f_switch_ids( dim3 block_nums_ids = block_nums; block_nums_ids.y *= col_tiles; - mul_mat_f<<>> + mul_mat_f<<>> (x, y, ids, dst, ncols_x, ncols_dst, nchannels_dst, stride_row, stride_col_y, stride_col_dst, stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst, sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst); } else { - mul_mat_f<<>> + mul_mat_f<<>> (x, y, ids, dst, ncols_x, cols_per_block, nchannels_dst, stride_row, stride_col_y, stride_col_dst, stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst, sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst); } } -template +template void mul_mat_f_cuda( const T * x, const float * y, const int32_t * ids, float * dst, const int64_t ncols_x, const int64_t nrows_x, const int64_t ncols_dst, @@ -605,7 +645,7 @@ void mul_mat_f_cuda( int64_t nwarps_best = 1; int64_t niter_best = (ncols_x + warp_size*2 - 1) / (warp_size*2); - int64_t max_block_size = 256; + int64_t max_block_size = mmf_get_max_block_size(cc); for (int64_t nwarps = 2; nwarps <= max_block_size/warp_size; nwarps++) { const int64_t niter = (ncols_x + nwarps*warp_size*2 - 1) / (nwarps*warp_size*2); if (niter < niter_best) { @@ -614,10 +654,9 @@ void mul_mat_f_cuda( } } - constexpr int rows_per_block = MMF_ROWS_PER_BLOCK; - const int nbytes_shared_iter = nwarps_best * (volta_mma_available(cc) ? tile_A_32::I : tile_A_16::I) * (warp_size + 4) * 4; - const int nbytes_cols_per_block_pad = amd_wmma_available(cc) ? tile_B_16::I : tile_B_8::I; - const int nbytes_shared_combine = GGML_PAD(cols_per_block, nbytes_cols_per_block_pad) * (nwarps_best*rows_per_block + 4) * 4; + const int nbytes_shared_iter = nwarps_best * (volta_mma_available(cc) ? tile_A_32::I : tile_A_16::I) * (warp_size + mmf_get_padding(cc)) * 4; + const int nbytes_cols_per_block_pad = (amd_wmma_available(cc) || amd_mfma_available(cc)) ? tile_B_16::I : tile_B_8::I; + const int nbytes_shared_combine = GGML_PAD(cols_per_block, nbytes_cols_per_block_pad) * (nwarps_best*rows_per_block + mmf_get_padding(cc)) * 4; const int nbytes_shared = std::max(nbytes_shared_iter, nbytes_shared_combine); const int nbytes_slotmap = ids ? GGML_PAD(cols_per_block, 16) * sizeof(int) : 0; const int nbytes_shared_total = nbytes_shared + nbytes_slotmap; @@ -628,56 +667,56 @@ void mul_mat_f_cuda( switch (nwarps_best) { case 1: { - mul_mat_f_switch_ids( + mul_mat_f_switch_ids( x, y, ids, dst, ncols_x, ncols_dst, nchannels_dst, stride_row, stride_col_y, stride_col_dst, stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst, sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream, ids_data); } break; case 2: { - mul_mat_f_switch_ids( + mul_mat_f_switch_ids( x, y, ids, dst, ncols_x, ncols_dst, nchannels_dst, stride_row, stride_col_y, stride_col_dst, stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst, sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream, ids_data); } break; case 3: { - mul_mat_f_switch_ids( + mul_mat_f_switch_ids( x, y, ids, dst, ncols_x, ncols_dst, nchannels_dst, stride_row, stride_col_y, stride_col_dst, stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst, sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream, ids_data); } break; case 4: { - mul_mat_f_switch_ids( + mul_mat_f_switch_ids( x, y, ids, dst, ncols_x, ncols_dst, nchannels_dst, stride_row, stride_col_y, stride_col_dst, stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst, sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream, ids_data); } break; case 5: { - mul_mat_f_switch_ids( + mul_mat_f_switch_ids( x, y, ids, dst, ncols_x, ncols_dst, nchannels_dst, stride_row, stride_col_y, stride_col_dst, stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst, sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream, ids_data); } break; case 6: { - mul_mat_f_switch_ids( + mul_mat_f_switch_ids( x, y, ids, dst, ncols_x, ncols_dst, nchannels_dst, stride_row, stride_col_y, stride_col_dst, stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst, sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream, ids_data); } break; case 7: { - mul_mat_f_switch_ids( + mul_mat_f_switch_ids( x, y, ids, dst, ncols_x, ncols_dst, nchannels_dst, stride_row, stride_col_y, stride_col_dst, stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst, sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream, ids_data); } break; case 8: { - mul_mat_f_switch_ids( + mul_mat_f_switch_ids( x, y, ids, dst, ncols_x, ncols_dst, nchannels_dst, stride_row, stride_col_y, stride_col_dst, stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst, sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream, @@ -691,7 +730,7 @@ void mul_mat_f_cuda( GGML_UNUSED_VARS(nchannels_y); } -template +template static void mul_mat_f_switch_cols_per_block( const T * x, const float * y, const int32_t * ids, float * dst, const int64_t ncols_x, const int64_t nrows_x, const int64_t ncols_dst, @@ -708,82 +747,82 @@ static void mul_mat_f_switch_cols_per_block( switch (ncols_case) { case 1: { - mul_mat_f_cuda(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst, + mul_mat_f_cuda(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst, stride_col_id, stride_row_id, nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data); } break; case 2: { - mul_mat_f_cuda(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst, + mul_mat_f_cuda(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst, stride_col_id, stride_row_id, nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data); } break; case 3: { - mul_mat_f_cuda(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst, + mul_mat_f_cuda(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst, stride_col_id, stride_row_id, nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data); } break; case 4: { - mul_mat_f_cuda(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst, + mul_mat_f_cuda(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst, stride_col_id, stride_row_id, nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data); } break; case 5: { - mul_mat_f_cuda(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst, + mul_mat_f_cuda(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst, stride_col_id, stride_row_id, nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data); } break; case 6: { - mul_mat_f_cuda(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst, + mul_mat_f_cuda(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst, stride_col_id, stride_row_id, nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data); } break; case 7: { - mul_mat_f_cuda(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst, + mul_mat_f_cuda(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst, stride_col_id, stride_row_id, nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data); } break; case 8: { - mul_mat_f_cuda(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst, + mul_mat_f_cuda(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst, stride_col_id, stride_row_id, nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data); } break; case 9: { - mul_mat_f_cuda(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst, + mul_mat_f_cuda(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst, stride_col_id, stride_row_id, nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data); } break; case 10: { - mul_mat_f_cuda(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst, + mul_mat_f_cuda(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst, stride_col_id, stride_row_id, nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data); } break; case 11: { - mul_mat_f_cuda(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst, + mul_mat_f_cuda(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst, stride_col_id, stride_row_id, nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data); } break; case 12: { - mul_mat_f_cuda(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst, + mul_mat_f_cuda(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst, stride_col_id, stride_row_id, nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data); } break; case 13: { - mul_mat_f_cuda(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst, + mul_mat_f_cuda(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst, stride_col_id, stride_row_id, nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data); } break; case 14: { - mul_mat_f_cuda(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst, + mul_mat_f_cuda(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst, stride_col_id, stride_row_id, nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data); } break; case 15: { - mul_mat_f_cuda(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst, + mul_mat_f_cuda(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst, stride_col_id, stride_row_id, nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data); } break; case 16: { - mul_mat_f_cuda(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst, + mul_mat_f_cuda(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst, stride_col_id, stride_row_id, nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data); } break; @@ -793,8 +832,36 @@ static void mul_mat_f_switch_cols_per_block( } } -#define DECL_MMF_CASE_HELPER(T, ncols_dst) \ - template void mul_mat_f_cuda( \ +template +static void mul_mat_f_switch_rows_per_block( + const int rows_per_block, const T * x, const float * y, const int32_t * ids, float * dst, + const int64_t ncols_x, const int64_t nrows_x, const int64_t ncols_dst, + const int64_t stride_row, const int64_t stride_col_y, const int64_t stride_col_dst, + const int64_t stride_col_id, const int stride_row_id, + const int64_t nchannels_x, const int64_t nchannels_y, const int64_t nchannels_dst, + const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst, const int64_t nsamples_x, + const int64_t nsamples_dst, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst, + cudaStream_t stream, const mmf_ids_data * ids_data) { + switch (rows_per_block) { + case MMF_ROWS_PER_BLOCK: { + mul_mat_f_switch_cols_per_block( + x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst, + stride_col_id, stride_row_id, nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst, + nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data); + } break; + case MMF_ROWS_PER_BLOCK_CDNA: { + mul_mat_f_switch_cols_per_block( + x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst, + stride_col_id, stride_row_id, nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst, + nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data); + } break; + default: + GGML_ABORT("unsupported rows_per_block: %i", rows_per_block); + } +} + +#define DECL_MMF_CASE_HELPER(T, nrows_dst, ncols_dst) \ + template void mul_mat_f_cuda( \ const T * x, const float * y, const int32_t * ids, float * dst, \ const int64_t ncols_x, const int64_t nrows_x, int64_t ncols_dst_total, const int64_t stride_row, const int64_t stride_col_y, const int64_t stride_col_dst, \ const int64_t stride_col_id, const int64_t stride_row_id, \ @@ -803,16 +870,22 @@ static void mul_mat_f_switch_cols_per_block( const int64_t nsamples_dst, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst, \ cudaStream_t stream, const mmf_ids_data * ids_data); -#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) +#if !defined(GGML_USE_MUSA) #define DECL_MMF_CASE_EXTERN(ncols_dst) \ - extern DECL_MMF_CASE_HELPER(float, ncols_dst) \ - extern DECL_MMF_CASE_HELPER(half2, ncols_dst) \ - extern DECL_MMF_CASE_HELPER(nv_bfloat162, ncols_dst) + extern DECL_MMF_CASE_HELPER(float, MMF_ROWS_PER_BLOCK, ncols_dst) \ + extern DECL_MMF_CASE_HELPER(half2, MMF_ROWS_PER_BLOCK, ncols_dst) \ + extern DECL_MMF_CASE_HELPER(nv_bfloat162, MMF_ROWS_PER_BLOCK, ncols_dst) \ + extern DECL_MMF_CASE_HELPER(float, MMF_ROWS_PER_BLOCK_CDNA, ncols_dst) \ + extern DECL_MMF_CASE_HELPER(half2, MMF_ROWS_PER_BLOCK_CDNA, ncols_dst) \ + extern DECL_MMF_CASE_HELPER(nv_bfloat162, MMF_ROWS_PER_BLOCK_CDNA, ncols_dst) #define DECL_MMF_CASE(ncols_dst) \ - DECL_MMF_CASE_HELPER(float, ncols_dst) \ - DECL_MMF_CASE_HELPER(half2, ncols_dst) \ - DECL_MMF_CASE_HELPER(nv_bfloat162, ncols_dst) + DECL_MMF_CASE_HELPER(float, MMF_ROWS_PER_BLOCK, ncols_dst) \ + DECL_MMF_CASE_HELPER(half2, MMF_ROWS_PER_BLOCK, ncols_dst) \ + DECL_MMF_CASE_HELPER(nv_bfloat162, MMF_ROWS_PER_BLOCK, ncols_dst) \ + DECL_MMF_CASE_HELPER(float, MMF_ROWS_PER_BLOCK_CDNA, ncols_dst) \ + DECL_MMF_CASE_HELPER(half2, MMF_ROWS_PER_BLOCK_CDNA, ncols_dst) \ + DECL_MMF_CASE_HELPER(nv_bfloat162, MMF_ROWS_PER_BLOCK_CDNA, ncols_dst) DECL_MMF_CASE_EXTERN(1); DECL_MMF_CASE_EXTERN(2); diff --git a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_32.cu b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_32.cu new file mode 100644 index 0000000000..1f554d81e5 --- /dev/null +++ b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_32.cu @@ -0,0 +1,5 @@ +// This file has been autogenerated by generate_cu_files.py, do not edit manually. + +#include "../fattn-mma-f16.cuh" + +DECL_FATTN_MMA_F16_CASE(576, 512, 1, 32); diff --git a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_32.cu b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_32.cu new file mode 100644 index 0000000000..264751d65e --- /dev/null +++ b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_32.cu @@ -0,0 +1,5 @@ +// This file has been autogenerated by generate_cu_files.py, do not edit manually. + +#include "../fattn-mma-f16.cuh" + +DECL_FATTN_MMA_F16_CASE(576, 512, 2, 32); diff --git a/ggml/src/ggml-cuda/template-instances/generate_cu_files.py b/ggml/src/ggml-cuda/template-instances/generate_cu_files.py index 10be71ab57..e382df1ae2 100755 --- a/ggml/src/ggml-cuda/template-instances/generate_cu_files.py +++ b/ggml/src/ggml-cuda/template-instances/generate_cu_files.py @@ -71,7 +71,7 @@ for type_k in TYPES_KV: f.write(SOURCE_FATTN_VEC.format(type_k=type_k, type_v=type_v)) for ncols in [8, 16, 32, 64]: - for ncols2 in [1, 2, 4, 8, 16]: + for ncols2 in [1, 2, 4, 8, 16, 32]: if ncols2 > ncols: continue ncols1 = ncols // ncols2 @@ -83,9 +83,9 @@ for ncols in [8, 16, 32, 64]: continue if head_size_kq == 72: continue - if head_size_kq != 576 and ncols2 == 16: + if head_size_kq != 576 and ncols2 in (16, 32): continue - if head_size_kq == 576 and ncols2 not in (4, 16): + if head_size_kq == 576 and ncols2 not in (4, 16, 32): continue head_size_v = head_size_kq if head_size_kq != 576 else 512 f.write(SOURCE_FATTN_MMA_CASE.format(ncols1=ncols1, ncols2=ncols2, head_size_kq=head_size_kq, head_size_v=head_size_v)) diff --git a/ggml/src/ggml-cuda/topk-moe.cu b/ggml/src/ggml-cuda/topk-moe.cu index 48e569efa0..08a88990dd 100644 --- a/ggml/src/ggml-cuda/topk-moe.cu +++ b/ggml/src/ggml-cuda/topk-moe.cu @@ -5,6 +5,13 @@ #include #include +// Kernel config struct - passed by value to CUDA kernel +struct topk_moe_config { + bool use_sigmoid; + bool with_norm; + bool delayed_softmax; +}; + // Warp-local softmax used for both the pre-top-k logits and the post-top-k delayed path. template __device__ void softmax_warp_inplace(float (&vals)[experts_per_thread], const int limit, const int lane) { @@ -50,6 +57,16 @@ __device__ void softmax_warp_inplace(float (&vals)[experts_per_thread], const in } } +template +__device__ void sigmoid_warp_inplace(float (&vals)[experts_per_thread], const int limit, const int lane) { +#pragma unroll + for (int i = 0; i < experts_per_thread; i++) { + const int idx = lane + i * WARP_SIZE; + const bool active = !use_limit || (idx < limit); + vals[i] = active ? 1.f / (1.f + expf(-vals[i])) : -INFINITY; + } +} + /* This kernel does the following: 1. optionally softmax over the logits per token [n_experts, n_tokens] @@ -59,13 +76,16 @@ __device__ void softmax_warp_inplace(float (&vals)[experts_per_thread], const in It is intended as fusion of softmax->top-k->get_rows pipeline for MoE models */ -template -__launch_bounds__(4 * WARP_SIZE, 1) __global__ void topk_moe_cuda(const float * logits, - float * weights, - int32_t * ids, - const int n_rows, - const int n_expert_used, - const float clamp_val) { +template +__launch_bounds__(4 * WARP_SIZE, 1) __global__ void topk_moe_cuda(const float * logits, + float * weights, + int32_t * ids, + float * bias, + const int n_rows, + const int n_expert_used, + const float clamp_val, + const float scale_val, + const topk_moe_config config) { const int row = blockIdx.x * blockDim.y + threadIdx.y; if (row >= n_rows) { return; @@ -79,14 +99,41 @@ __launch_bounds__(4 * WARP_SIZE, 1) __global__ void topk_moe_cuda(const float * float wt[experts_per_thread]; + // Initialize all slots to -INFINITY +#pragma unroll + for (int i = 0; i < experts_per_thread; i++) { + wt[i] = -INFINITY; + } + #pragma unroll for (int i = 0; i < n_experts; i += WARP_SIZE) { const int expert = i + threadIdx.x; wt[i / WARP_SIZE] = (n_experts % WARP_SIZE == 0 || expert < n_experts) ? logits[expert] : -INFINITY; } - if constexpr (!delayed_softmax) { - softmax_warp_inplace(wt, n_experts, threadIdx.x); + if (!config.delayed_softmax) { + if (config.use_sigmoid) { + sigmoid_warp_inplace(wt, n_experts, threadIdx.x); + } else { + softmax_warp_inplace(wt, n_experts, threadIdx.x); + } + } + + // selection_wt is only needed when bias is present (selection uses wt + bias) + // when no bias, we use wt directly for both selection and weight values + float selection_wt[has_bias ? experts_per_thread : 1]; + + if constexpr (has_bias) { +#pragma unroll + for (int i = 0; i < experts_per_thread; i++) { + selection_wt[i] = -INFINITY; + } +#pragma unroll + for (int i = 0; i < n_experts; i += WARP_SIZE) { + const int expert = i + threadIdx.x; + selection_wt[i / WARP_SIZE] = + (n_experts % WARP_SIZE == 0 || expert < n_experts) ? wt[i / WARP_SIZE] + bias[expert] : -INFINITY; + } } //at this point, each thread holds either a portion of the softmax distribution @@ -106,22 +153,56 @@ __launch_bounds__(4 * WARP_SIZE, 1) __global__ void topk_moe_cuda(const float * float max_val = wt[0]; int max_expert = threadIdx.x; -#pragma unroll - for (int i = 1; i < experts_per_thread; i++) { - const int expert = threadIdx.x + i * WARP_SIZE; - if ((n_experts % WARP_SIZE == 0 || expert < n_experts) && wt[i] > max_val) { - max_val = wt[i]; - max_expert = expert; - } - } + if constexpr (has_bias) { + float max_val_s = selection_wt[0]; #pragma unroll - for (int mask = WARP_SIZE / 2; mask > 0; mask /= 2) { - const float val = __shfl_xor_sync(0xFFFFFFFF, max_val, mask, WARP_SIZE); - const int expert = __shfl_xor_sync(0xFFFFFFFF, max_expert, mask, WARP_SIZE); - if (val > max_val || (val == max_val && expert < max_expert)) { - max_val = val; - max_expert = expert; + for (int i = 1; i < experts_per_thread; i++) { + const int expert = threadIdx.x + i * WARP_SIZE; + if ((n_experts % WARP_SIZE == 0 || expert < n_experts) && selection_wt[i] > max_val_s) { + max_val = wt[i]; + max_val_s = selection_wt[i]; + max_expert = expert; + } + } + +#pragma unroll + for (int mask = WARP_SIZE / 2; mask > 0; mask /= 2) { + const float val = __shfl_xor_sync(0xFFFFFFFF, max_val, mask, WARP_SIZE); + const float val_s = __shfl_xor_sync(0xFFFFFFFF, max_val_s, mask, WARP_SIZE); + const int expert = __shfl_xor_sync(0xFFFFFFFF, max_expert, mask, WARP_SIZE); + if (val_s > max_val_s || (val_s == max_val_s && expert < max_expert)) { + max_val = val; + max_val_s = val_s; + max_expert = expert; + } + } + + if ((max_expert & (WARP_SIZE - 1)) == threadIdx.x) { + selection_wt[max_expert / WARP_SIZE] = -INFINITY; + } + } else { +#pragma unroll + for (int i = 1; i < experts_per_thread; i++) { + const int expert = threadIdx.x + i * WARP_SIZE; + if ((n_experts % WARP_SIZE == 0 || expert < n_experts) && wt[i] > max_val) { + max_val = wt[i]; + max_expert = expert; + } + } + +#pragma unroll + for (int mask = WARP_SIZE / 2; mask > 0; mask /= 2) { + const float val = __shfl_xor_sync(0xFFFFFFFF, max_val, mask, WARP_SIZE); + const int expert = __shfl_xor_sync(0xFFFFFFFF, max_expert, mask, WARP_SIZE); + if (val > max_val || (val == max_val && expert < max_expert)) { + max_val = val; + max_expert = expert; + } + } + + if ((max_expert & (WARP_SIZE - 1)) == threadIdx.x) { + wt[max_expert / WARP_SIZE] = -INFINITY; } } @@ -130,16 +211,14 @@ __launch_bounds__(4 * WARP_SIZE, 1) __global__ void topk_moe_cuda(const float * } if ((max_expert & (WARP_SIZE - 1)) == threadIdx.x) { - wt[max_expert / WARP_SIZE] = -INFINITY; - ids[k] = max_expert; - if constexpr (with_norm) { + if (config.with_norm) { wt_sum += max_val; } } } - if constexpr (with_norm) { + if (config.with_norm) { wt_sum = warp_reduce_sum(wt_sum); wt_sum = max(wt_sum, clamp_val); const float inv_sum = 1.0f / wt_sum; @@ -149,7 +228,7 @@ __launch_bounds__(4 * WARP_SIZE, 1) __global__ void topk_moe_cuda(const float * } } - if constexpr (delayed_softmax) { + if (config.delayed_softmax) { softmax_warp_inplace(output_weights, n_expert_used, threadIdx.x); } @@ -157,25 +236,25 @@ __launch_bounds__(4 * WARP_SIZE, 1) __global__ void topk_moe_cuda(const float * for (int i = 0; i < experts_per_thread; i++) { const int idx = i * WARP_SIZE + threadIdx.x; if (idx < n_expert_used) { - weights[idx] = output_weights[i]; + weights[idx] = output_weights[i] * scale_val; } } - - if (!with_norm) { - GGML_UNUSED(clamp_val); - } } -template +template static void launch_topk_moe_cuda(ggml_backend_cuda_context & ctx, const float * logits, float * weights, int32_t * ids, + float * bias, const int n_rows, const int n_expert, const int n_expert_used, - const float clamp_val) { - static_assert(!(with_norm && delayed_softmax), "delayed softmax is not supported with weight normalization"); + const float clamp_val, + const float scale_val, + const topk_moe_config config) { + GGML_ASSERT(!(config.with_norm && config.delayed_softmax) && + "delayed softmax is not supported with weight normalization"); const int rows_per_block = 4; dim3 grid_dims((n_rows + rows_per_block - 1) / rows_per_block, 1, 1); dim3 block_dims(WARP_SIZE, rows_per_block, 1); @@ -183,44 +262,48 @@ static void launch_topk_moe_cuda(ggml_backend_cuda_context & ctx, switch (n_expert) { case 1: - topk_moe_cuda<1, with_norm, delayed_softmax> - <<>>(logits, weights, ids, n_rows, n_expert_used, clamp_val); + topk_moe_cuda<1, has_bias><<>>(logits, weights, ids, bias, n_rows, n_expert_used, + clamp_val, scale_val, config); break; case 2: - topk_moe_cuda<2, with_norm, delayed_softmax> - <<>>(logits, weights, ids, n_rows, n_expert_used, clamp_val); + topk_moe_cuda<2, has_bias><<>>(logits, weights, ids, bias, n_rows, n_expert_used, + clamp_val, scale_val, config); break; case 4: - topk_moe_cuda<4, with_norm, delayed_softmax> - <<>>(logits, weights, ids, n_rows, n_expert_used, clamp_val); + topk_moe_cuda<4, has_bias><<>>(logits, weights, ids, bias, n_rows, n_expert_used, + clamp_val, scale_val, config); break; case 8: - topk_moe_cuda<8, with_norm, delayed_softmax> - <<>>(logits, weights, ids, n_rows, n_expert_used, clamp_val); + topk_moe_cuda<8, has_bias><<>>(logits, weights, ids, bias, n_rows, n_expert_used, + clamp_val, scale_val, config); break; case 16: - topk_moe_cuda<16, with_norm, delayed_softmax> - <<>>(logits, weights, ids, n_rows, n_expert_used, clamp_val); + topk_moe_cuda<16, has_bias><<>>(logits, weights, ids, bias, n_rows, n_expert_used, + clamp_val, scale_val, config); break; case 32: - topk_moe_cuda<32, with_norm, delayed_softmax> - <<>>(logits, weights, ids, n_rows, n_expert_used, clamp_val); + topk_moe_cuda<32, has_bias><<>>(logits, weights, ids, bias, n_rows, n_expert_used, + clamp_val, scale_val, config); break; case 64: - topk_moe_cuda<64, with_norm, delayed_softmax> - <<>>(logits, weights, ids, n_rows, n_expert_used, clamp_val); + topk_moe_cuda<64, has_bias><<>>(logits, weights, ids, bias, n_rows, n_expert_used, + clamp_val, scale_val, config); break; case 128: - topk_moe_cuda<128, with_norm, delayed_softmax> - <<>>(logits, weights, ids, n_rows, n_expert_used, clamp_val); + topk_moe_cuda<128, has_bias><<>>(logits, weights, ids, bias, n_rows, n_expert_used, + clamp_val, scale_val, config); break; case 256: - topk_moe_cuda<256, with_norm, delayed_softmax> - <<>>(logits, weights, ids, n_rows, n_expert_used, clamp_val); + topk_moe_cuda<256, has_bias><<>>(logits, weights, ids, bias, n_rows, n_expert_used, + clamp_val, scale_val, config); break; case 512: - topk_moe_cuda<512, with_norm, delayed_softmax> - <<>>(logits, weights, ids, n_rows, n_expert_used, clamp_val); + topk_moe_cuda<512, has_bias><<>>(logits, weights, ids, bias, n_rows, n_expert_used, + clamp_val, scale_val, config); + break; + case 576: + topk_moe_cuda<576, has_bias><<>>(logits, weights, ids, bias, n_rows, n_expert_used, + clamp_val, scale_val, config); break; default: GGML_ASSERT(false && "fatal error"); @@ -228,13 +311,14 @@ static void launch_topk_moe_cuda(ggml_backend_cuda_context & ctx, } } -void ggml_cuda_op_topk_moe(ggml_backend_cuda_context & ctx, - const ggml_tensor * logits, - ggml_tensor * weights, - ggml_tensor * ids, - const bool with_norm, - const bool delayed_softmax, - ggml_tensor * clamp) { +void ggml_cuda_op_topk_moe(ggml_backend_cuda_context & ctx, + const ggml_tensor * logits, + ggml_tensor * weights, + ggml_tensor * ids, + const ggml_tensor * clamp, + const ggml_tensor * scale, + const ggml_tensor * bias, + const ggml_cuda_topk_moe_args & args) { GGML_ASSERT(logits->type == GGML_TYPE_F32); GGML_ASSERT(weights->type == GGML_TYPE_F32); GGML_ASSERT(ids->type == GGML_TYPE_I32); @@ -245,107 +329,75 @@ void ggml_cuda_op_topk_moe(ggml_backend_cuda_context & ctx, const float * logits_d = (const float *) logits->data; float * weights_d = (float *) weights->data; int32_t * ids_d = (int32_t *) ids->data; + float * bias_d = bias ? (float *) bias->data : nullptr; + + float scale_val = scale ? ggml_get_op_params_f32(scale, 0) : 1.0f; GGML_ASSERT(ids->nb[1] / ggml_type_size(ids->type) == (size_t) n_experts); const int n_expert_used = weights->ne[1]; + const bool with_norm = clamp != nullptr; + float clamp_val = -INFINITY; - if (with_norm) { - if (clamp) { - clamp_val = ggml_get_op_params_f32(clamp, 0); - } - launch_topk_moe_cuda(ctx, logits_d, weights_d, ids_d, n_rows, n_experts, n_expert_used, clamp_val); + if (clamp) { + clamp_val = ggml_get_op_params_f32(clamp, 0); + } + + topk_moe_config config; + config.use_sigmoid = args.sigmoid; + config.with_norm = with_norm; + config.delayed_softmax = args.delayed_softmax; + + if (bias) { + launch_topk_moe_cuda(ctx, logits_d, weights_d, ids_d, bias_d, n_rows, n_experts, n_expert_used, clamp_val, + scale_val, config); } else { - GGML_ASSERT(clamp == nullptr); - if (delayed_softmax) { - launch_topk_moe_cuda(ctx, logits_d, weights_d, ids_d, n_rows, n_experts, n_expert_used, - clamp_val); - } else { - launch_topk_moe_cuda(ctx, logits_d, weights_d, ids_d, n_rows, n_experts, n_expert_used, - clamp_val); - } + launch_topk_moe_cuda(ctx, logits_d, weights_d, ids_d, bias_d, n_rows, n_experts, n_expert_used, clamp_val, + scale_val, config); } } -bool ggml_cuda_should_use_topk_moe(const ggml_tensor * softmax, +bool ggml_cuda_should_use_topk_moe(const ggml_tensor * gating_op, const ggml_tensor * weights, - const ggml_tensor * get_rows, - const ggml_tensor * argsort, - const ggml_tensor * clamp, - int n_expert) { - ggml_tensor * probs = get_rows->src[0]; - if (probs->op != GGML_OP_RESHAPE) { - return false; - } - probs = probs->src[0]; - ggml_tensor * selection_probs = argsort->src[0]; - - if (probs != selection_probs) { + const ggml_tensor * logits, + const ggml_tensor * ids) { + const int n_expert = ids->nb[1] / ids->nb[0]; + if (((n_expert & (n_expert - 1)) != 0 || n_expert > 512) && n_expert != 576) { return false; } - float scale = 1.0f; - float max_bias = 0.0f; - - memcpy(&scale, (const float *) softmax->op_params + 0, sizeof(float)); - memcpy(&max_bias, (const float *) softmax->op_params + 1, sizeof(float)); - - if (!ggml_is_contiguous(softmax->src[0]) || !ggml_is_contiguous(weights)) { + if (!ggml_is_contiguous(weights) || !ggml_is_contiguous(logits)) { return false; } - if (scale != 1.0f || max_bias != 0.0f) { - return false; - } + if (gating_op->op == GGML_OP_SOFT_MAX) { + const ggml_tensor * softmax = gating_op; + float scale = 1.0f; + float max_bias = 0.0f; - // don't fuse when masks or sinks are present - if (softmax->src[1] || softmax->src[2]) { - return false; - } + memcpy(&scale, (const float *) softmax->op_params + 0, sizeof(float)); + memcpy(&max_bias, (const float *) softmax->op_params + 1, sizeof(float)); - // n_expert must be a power of 2 - if ((n_expert & (n_expert - 1)) != 0 || n_expert > 512) { - return false; - } - - if (clamp) { - if (clamp->op != GGML_OP_CLAMP) { + if (!ggml_is_contiguous(softmax->src[0])) { return false; } - float max_val = ggml_get_op_params_f32(clamp, 1); - if (max_val != INFINITY) { + if (scale != 1.0f || max_bias != 0.0f) { + return false; + } + + // don't fuse when masks or sinks are present + if (softmax->src[1] || softmax->src[2]) { + return false; + } + } else if (gating_op->op == GGML_OP_UNARY) { + ggml_unary_op op = ggml_get_unary_op(gating_op); + + if (op != GGML_UNARY_OP_SIGMOID) { return false; } } - return true; } - -std::initializer_list ggml_cuda_topk_moe_ops(bool norm, bool delayed_softmax) { - static std::initializer_list norm_ops = { GGML_OP_SOFT_MAX, GGML_OP_RESHAPE, GGML_OP_ARGSORT, - GGML_OP_VIEW, GGML_OP_GET_ROWS, GGML_OP_RESHAPE, - GGML_OP_SUM_ROWS, GGML_OP_CLAMP, GGML_OP_DIV, - GGML_OP_RESHAPE }; - - static std::initializer_list no_norm_ops = { GGML_OP_SOFT_MAX, GGML_OP_RESHAPE, GGML_OP_ARGSORT, - GGML_OP_VIEW, GGML_OP_GET_ROWS }; - - static std::initializer_list delayed_softmax_ops = { GGML_OP_ARGSORT, GGML_OP_VIEW, - GGML_OP_GET_ROWS, GGML_OP_RESHAPE, - GGML_OP_SOFT_MAX, GGML_OP_RESHAPE }; - - GGML_ASSERT(!norm || !delayed_softmax); - - if (delayed_softmax) { - return delayed_softmax_ops; - } - - if (norm) { - return norm_ops; - } - - return no_norm_ops; -} diff --git a/ggml/src/ggml-cuda/topk-moe.cuh b/ggml/src/ggml-cuda/topk-moe.cuh index 6b6c13c587..243dc2f1c4 100644 --- a/ggml/src/ggml-cuda/topk-moe.cuh +++ b/ggml/src/ggml-cuda/topk-moe.cuh @@ -3,19 +3,25 @@ #include -void ggml_cuda_op_topk_moe(ggml_backend_cuda_context & ctx, - const ggml_tensor * logits, - ggml_tensor * weights, - ggml_tensor * ids, - const bool with_norm, - const bool delayed_softmax = false, - ggml_tensor * weight_clamp = nullptr); +struct ggml_cuda_topk_moe_args { + bool sigmoid{}; + bool softmax{}; + bool delayed_softmax{}; + bool prob_bias{}; + bool norm{}; + bool scale{}; +}; -bool ggml_cuda_should_use_topk_moe(const ggml_tensor * softmax, +void ggml_cuda_op_topk_moe(ggml_backend_cuda_context & ctx, + const ggml_tensor * logits, + ggml_tensor * weights, + ggml_tensor * ids, + const ggml_tensor * clamp, + const ggml_tensor * scale, + const ggml_tensor * bias, + const ggml_cuda_topk_moe_args & args); + +bool ggml_cuda_should_use_topk_moe(const ggml_tensor * gating_op, const ggml_tensor * weights, - const ggml_tensor * get_rows, - const ggml_tensor * argsort, - const ggml_tensor * clamp, - int n_expert); - -std::initializer_list ggml_cuda_topk_moe_ops(bool with_norm, bool delayed_softmax = false); + const ggml_tensor * logits, + const ggml_tensor * ids); diff --git a/ggml/src/ggml-hip/CMakeLists.txt b/ggml/src/ggml-hip/CMakeLists.txt index 23b6889919..80037d2436 100644 --- a/ggml/src/ggml-hip/CMakeLists.txt +++ b/ggml/src/ggml-hip/CMakeLists.txt @@ -62,6 +62,8 @@ file(GLOB SRCS "../ggml-cuda/template-instances/fattn-mma*.cu") list(APPEND GGML_SOURCES_ROCM ${SRCS}) file(GLOB SRCS "../ggml-cuda/template-instances/mmq*.cu") list(APPEND GGML_SOURCES_ROCM ${SRCS}) +file(GLOB SRCS "../ggml-cuda/template-instances/mmf*.cu") +list(APPEND GGML_SOURCES_ROCM ${SRCS}) if (GGML_CUDA_FA_ALL_QUANTS) file(GLOB SRCS "../ggml-cuda/template-instances/fattn-vec*.cu") diff --git a/ggml/src/ggml-metal/ggml-metal-device.m b/ggml/src/ggml-metal/ggml-metal-device.m index eb4e2c209c..7f9c384c34 100644 --- a/ggml/src/ggml-metal/ggml-metal-device.m +++ b/ggml/src/ggml-metal/ggml-metal-device.m @@ -785,8 +785,12 @@ ggml_metal_device_t ggml_metal_device_init(void) { dev->props.op_offload_min_batch_size = getenv("GGML_OP_OFFLOAD_MIN_BATCH") ? atoi(getenv("GGML_OP_OFFLOAD_MIN_BATCH")) : 32; dev->props.max_buffer_size = dev->mtl_device.maxBufferLength; - dev->props.max_working_set_size = dev->mtl_device.recommendedMaxWorkingSetSize; dev->props.max_theadgroup_memory_size = dev->mtl_device.maxThreadgroupMemoryLength; + if (@available(macOS 10.12, iOS 16.0, *)) { + dev->props.max_working_set_size = dev->mtl_device.recommendedMaxWorkingSetSize; + } else { + dev->props.max_working_set_size = dev->mtl_device.maxBufferLength; + } strncpy(dev->props.name, [[dev->mtl_device name] UTF8String], sizeof(dev->props.name) - 1); diff --git a/ggml/src/ggml-opencl/CMakeLists.txt b/ggml/src/ggml-opencl/CMakeLists.txt index 79039c30e1..0259474b6e 100644 --- a/ggml/src/ggml-opencl/CMakeLists.txt +++ b/ggml/src/ggml-opencl/CMakeLists.txt @@ -85,7 +85,8 @@ set(GGML_OPENCL_KERNELS mul_mv_q4_0_f32_8x_flat mul_mv_q4_0_f32_1d_8x_flat mul_mv_q4_0_f32_1d_16x_flat - mul_mv_q6_k + mul_mv_q6_k_f32 + mul_mv_q6_k_f32_flat mul_mv_q8_0_f32 mul_mv_q8_0_f32_flat mul_mv_mxfp4_f32 diff --git a/ggml/src/ggml-opencl/ggml-opencl.cpp b/ggml/src/ggml-opencl/ggml-opencl.cpp index 27b2761ef1..678e40965a 100644 --- a/ggml/src/ggml-opencl/ggml-opencl.cpp +++ b/ggml/src/ggml-opencl/ggml-opencl.cpp @@ -533,8 +533,10 @@ struct ggml_backend_opencl_context { cl_kernel kernel_mul_mat_q4_0_f32_8x_flat; cl_kernel kernel_convert_block_q4_0_noshuffle; cl_kernel kernel_restore_block_q4_0_noshuffle; + cl_kernel kernel_convert_block_q6_K, kernel_restore_block_q6_K; cl_kernel kernel_mul_mat_q4_0_f32_1d_8x_flat, kernel_mul_mat_q4_0_f32_1d_16x_flat; cl_kernel kernel_mul_mv_q6_K_f32; + cl_kernel kernel_mul_mv_q6_K_f32_flat; cl_kernel kernel_mul_mv_mxfp4_f32, kernel_mul_mv_mxfp4_f32_flat; cl_kernel kernel_mul_mv_q8_0_f32, kernel_mul_mv_q8_0_f32_flat; cl_kernel kernel_solve_tri_f32; @@ -892,6 +894,8 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve CL_CHECK((backend_ctx->kernel_restore_block_mxfp4 = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_mxfp4", &err), err)); CL_CHECK((backend_ctx->kernel_convert_block_q8_0 = clCreateKernel(backend_ctx->program_cvt, "kernel_convert_block_q8_0", &err), err)); CL_CHECK((backend_ctx->kernel_restore_block_q8_0 = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_q8_0", &err), err)); + CL_CHECK((backend_ctx->kernel_convert_block_q6_K = clCreateKernel(backend_ctx->program_cvt, "kernel_convert_block_q6_K", &err), err)); + CL_CHECK((backend_ctx->kernel_restore_block_q6_K = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_q6_K", &err), err)); GGML_LOG_CONT("."); } @@ -1114,14 +1118,14 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve GGML_LOG_CONT("."); } - // mul_mv_q6_k + // mul_mv_q6_k_f32 { #ifdef GGML_OPENCL_EMBED_KERNELS const std::string kernel_src { - #include "mul_mv_q6_k.cl.h" + #include "mul_mv_q6_k_f32.cl.h" }; #else - const std::string kernel_src = read_file("mul_mv_q6_k.cl"); + const std::string kernel_src = read_file("mul_mv_q6_k_f32.cl"); #endif backend_ctx->program_mul_mv_q6_K = build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts); @@ -1130,6 +1134,23 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve GGML_LOG_CONT("."); } + // mul_mv_q6_k_f32_flat + { +#ifdef GGML_OPENCL_EMBED_KERNELS + const std::string kernel_src { + #include "mul_mv_q6_k_f32_flat.cl.h" + }; +#else + const std::string kernel_src = read_file("mul_mv_q6_k_f32_flat.cl"); +#endif + cl_program prog = + build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts); + + CL_CHECK((backend_ctx->kernel_mul_mv_q6_K_f32_flat = clCreateKernel(prog, "kernel_mul_mv_q6_K_f32_flat", &err), err)); + CL_CHECK(clReleaseProgram(prog)); + GGML_LOG_CONT("."); + } + // mul_mv_q8_0_f32 { #ifdef GGML_OPENCL_EMBED_KERNELS @@ -2919,6 +2940,50 @@ struct ggml_tensor_extra_cl_q8_0 { } }; +struct ggml_tensor_extra_cl_q6_K { + // Lower 4 bits of quantized weights. + cl_mem ql = nullptr; + // Upper 2 bits of quantized weights. + cl_mem qh = nullptr; + // Scales for each block. + cl_mem s = nullptr; + // Scales for each super block. + cl_mem d = nullptr; + + size_t size_ql = 0; + size_t size_qh = 0; + size_t size_s = 0; + size_t size_d = 0; + + ~ggml_tensor_extra_cl_q6_K() { + reset(); + } + + void reset() { + if (ql != nullptr) { + CL_CHECK(clReleaseMemObject(ql)); + ql = nullptr; + } + if (qh != nullptr) { + CL_CHECK(clReleaseMemObject(qh)); + qh = nullptr; + } + if (s != nullptr) { + CL_CHECK(clReleaseMemObject(s)); + s = nullptr; + } + if (d != nullptr) { + CL_CHECK(clReleaseMemObject(d)); + d = nullptr; + } + + size_ql = 0; + size_qh = 0; + size_s = 0; + size_d = 0; + } +}; + //------------------------------------------------------------------------------ // Backend API //------------------------------------------------------------------------------ @@ -3465,6 +3530,12 @@ struct ggml_backend_opencl_buffer_context { for (ggml_tensor_extra_cl_q8_0 * e : temp_tensor_extras_q8_0_in_use) { delete e; } + for (ggml_tensor_extra_cl_q6_K * e : temp_tensor_extras_q6_K) { + delete e; + } + for (ggml_tensor_extra_cl_q6_K * e : temp_tensor_extras_q6_K_in_use) { + delete e; + } } ggml_tensor_extra_cl * ggml_opencl_alloc_temp_tensor_extra() { @@ -3527,6 +3598,21 @@ struct ggml_backend_opencl_buffer_context { return extra; } + ggml_tensor_extra_cl_q6_K * ggml_opencl_alloc_temp_tensor_extra_q6_K() { + ggml_tensor_extra_cl_q6_K * extra; + if (temp_tensor_extras_q6_K.empty()) { + extra = new ggml_tensor_extra_cl_q6_K(); + } else { + extra = temp_tensor_extras_q6_K.back(); + temp_tensor_extras_q6_K.pop_back(); + } + + temp_tensor_extras_q6_K_in_use.push_back(extra); + + extra->reset(); + return extra; + } + void reset() { for (ggml_tensor_extra_cl * e : temp_tensor_extras_in_use) { temp_tensor_extras.push_back(e); @@ -3547,6 +3633,11 @@ struct ggml_backend_opencl_buffer_context { temp_tensor_extras_q8_0.push_back(e); } temp_tensor_extras_q8_0_in_use.clear(); + + for (ggml_tensor_extra_cl_q6_K * e : temp_tensor_extras_q6_K_in_use) { + temp_tensor_extras_q6_K.push_back(e); + } + temp_tensor_extras_q6_K_in_use.clear(); } // Pools for extras. Available extras are in `temp_tensor_extras`. Extras @@ -3562,6 +3653,8 @@ struct ggml_backend_opencl_buffer_context { std::vector temp_tensor_extras_mxfp4_in_use; std::vector temp_tensor_extras_q8_0; std::vector temp_tensor_extras_q8_0_in_use; + std::vector temp_tensor_extras_q6_K; + std::vector temp_tensor_extras_q6_K_in_use; // The buffer_context is initially created by ggml_backend_buft_alloc_buffer // before any tensor is initialized (at the beginning of alloc_tensor_range). @@ -4068,6 +4161,92 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer, return; } + if (tensor->type == GGML_TYPE_Q6_K) { + ggml_tensor_extra_cl * extra_orig = (ggml_tensor_extra_cl *)tensor->extra; + GGML_ASSERT(extra_orig && "Tesnors in OpenCL backend should have been allocated and initialized"); + + // Allocate the new extra and create aliases from the original. + ggml_backend_opencl_buffer_context * ctx = (ggml_backend_opencl_buffer_context *) buffer->context; + ggml_tensor_extra_cl_q6_K * extra = ctx->ggml_opencl_alloc_temp_tensor_extra_q6_K(); + + size_t size_ql = ggml_nelements(tensor)/ggml_blck_size(tensor->type)*ggml_blck_size(tensor->type)/2; + size_t size_qh = ggml_nelements(tensor)/ggml_blck_size(tensor->type)*ggml_blck_size(tensor->type)/4; + size_t size_s = ggml_nelements(tensor)/ggml_blck_size(tensor->type)*ggml_blck_size(tensor->type)/16; + size_t size_d = ggml_nelements(tensor)/ggml_blck_size(tensor->type)*sizeof(ggml_fp16_t); + GGML_ASSERT(size_ql + size_qh + size_s + size_d == ggml_nbytes(tensor) && + "Incorrect tensor size"); + + cl_int err; + cl_mem data_device = clCreateBuffer(context, CL_MEM_READ_WRITE, + ggml_nbytes(tensor), NULL, &err); + CL_CHECK(err); + CL_CHECK(clEnqueueWriteBuffer( + queue, data_device, CL_TRUE, 0, + ggml_nbytes(tensor), data, 0, NULL, NULL)); + + cl_buffer_region region; + + // Subbuffer for ql + region.origin = align_to(extra_orig->offset + tensor->view_offs + offset, backend_ctx->alignment); + region.size = size_ql; + extra->ql = clCreateSubBuffer( + extra_orig->data_device, CL_MEM_READ_WRITE, + CL_BUFFER_CREATE_TYPE_REGION, ®ion, &err); + CL_CHECK(err); + auto previous_origin = region.origin; + + // Subbuffer for qh + region.origin = align_to(previous_origin + size_ql, backend_ctx->alignment); + region.size = size_qh; + extra->qh = clCreateSubBuffer( + extra_orig->data_device, CL_MEM_READ_WRITE, + CL_BUFFER_CREATE_TYPE_REGION, ®ion, &err); + CL_CHECK(err); + previous_origin = region.origin; + + // Subbuffer for scales + region.origin = align_to(previous_origin + size_qh, backend_ctx->alignment); + region.size = size_s; + extra->s = clCreateSubBuffer( + extra_orig->data_device, CL_MEM_READ_WRITE, + CL_BUFFER_CREATE_TYPE_REGION, ®ion, &err); + CL_CHECK(err); + previous_origin = region.origin; + + // Create subbuffer for d. + region.origin = align_to(previous_origin + size_s, backend_ctx->alignment); + region.size = size_d; + extra->d = clCreateSubBuffer( + extra_orig->data_device, CL_MEM_READ_WRITE, + CL_BUFFER_CREATE_TYPE_REGION, ®ion, &err); + CL_CHECK(err); + previous_origin = region.origin; + + // Flatten the weights + cl_kernel kernel = backend_ctx->kernel_convert_block_q6_K; + + CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &data_device)); + CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &extra->ql)); + CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &extra->qh)); + CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem), &extra->s)); + CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem), &extra->d)); + + size_t global_work_size[] = {(size_t)ggml_nelements(tensor)/ggml_blck_size(tensor->type), 1, 1}; + size_t local_work_size[] = {64, 1, 1}; + + cl_event evt; + CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt)); + CL_CHECK(clWaitForEvents(1, &evt)); + CL_CHECK(clReleaseMemObject(data_device)); + + extra->size_ql = size_ql; + extra->size_qh = size_qh; + extra->size_s = size_s; + extra->size_d = size_d; + + tensor->extra = extra; + return; + } #endif // GGML_OPENCL_SOA_Q ggml_tensor_extra_cl * extra = (ggml_tensor_extra_cl *) tensor->extra; @@ -4277,6 +4456,34 @@ static void ggml_backend_opencl_buffer_get_tensor(ggml_backend_buffer_t buffer, size_t global_work_size[] = {(size_t)ggml_nelements(tensor)/ggml_blck_size(tensor->type), 1, 1}; size_t local_work_size[] = {1, 1, 1}; + cl_event evt; + CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, + global_work_size, local_work_size, 0, NULL, &evt)); + CL_CHECK(clWaitForEvents(1, &evt)); + CL_CHECK(clEnqueueReadBuffer( + queue, data_device, CL_TRUE, offset, + size, data, 0, NULL, NULL)); + CL_CHECK(clReleaseMemObject(data_device)); + return; + } + if (tensor->type == GGML_TYPE_Q6_K) { + ggml_tensor_extra_cl_q6_K * extra = (ggml_tensor_extra_cl_q6_K *)tensor->extra; + + cl_int err; + cl_mem data_device = clCreateBuffer(context, CL_MEM_READ_WRITE, + ggml_nbytes(tensor), NULL, &err); + CL_CHECK(err); + + cl_kernel kernel = backend_ctx->kernel_restore_block_q6_K; + CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &extra->ql)); + CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &extra->qh)); + CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &extra->s)); + CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem), &extra->d)); + CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem), &data_device)); + + size_t global_work_size[] = {(size_t)ggml_nelements(tensor)/ggml_blck_size(tensor->type), 1, 1}; + size_t local_work_size[] = {1, 1, 1}; + cl_event evt; CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt)); @@ -7765,6 +7972,7 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co ggml_tensor_extra_cl_q4_0 * extra0_q4_0 = (ggml_tensor_extra_cl_q4_0 *)src0->extra; ggml_tensor_extra_cl_mxfp4 * extra0_mxfp4 = (ggml_tensor_extra_cl_mxfp4 *)src0->extra; ggml_tensor_extra_cl_q8_0 * extra0_q8_0 = (ggml_tensor_extra_cl_q8_0 *)src0->extra; + ggml_tensor_extra_cl_q6_K * extra0_q6_K = (ggml_tensor_extra_cl_q6_K *)src0->extra; #endif const int ne00 = src0 ? src0->ne[0] : 0; @@ -8648,14 +8856,49 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co case GGML_TYPE_Q4_K: case GGML_TYPE_Q5_K: case GGML_TYPE_Q6_K: +#ifdef GGML_OPENCL_SOA_Q + kernel = backend_ctx->kernel_mul_mv_q6_K_f32_flat; + + if (backend_ctx->gpu_family == INTEL) { + nth0 = 16; + nth1 = 2; + ndst = 4; + } else if (backend_ctx->gpu_family == ADRENO) { + nth0 = 64; + nth1 = 2; + ndst = 4; + } else { + GGML_ASSERT(false && "TODO: Unknown GPU"); + } + + CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &extra0_q6_K->ql)); + CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &extra0_q6_K->qh)); + CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &extra0_q6_K->s)); + CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem), &extra0_q6_K->d)); + CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem), &extra1->data_device)); + CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong), &offset1)); + CL_CHECK(clSetKernelArg(kernel, 6, sizeof(cl_mem), &extrad->data_device)); + CL_CHECK(clSetKernelArg(kernel, 7, sizeof(cl_ulong), &offsetd)); + CL_CHECK(clSetKernelArg(kernel, 8, sizeof(int), &ne00)); + CL_CHECK(clSetKernelArg(kernel, 9, sizeof(int), &ne01)); + CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int), &ne02)); + CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int), &ne10)); + CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int), &ne12)); + CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int), &ne0)); + CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int), &ne1)); + CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int), &r2)); + CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int), &r3)); +#else kernel = backend_ctx->kernel_mul_mv_q6_K_f32; if (backend_ctx->gpu_family == INTEL) { - nth0 = 2; - nth1 = 16; + nth0 = 16; + nth1 = 2; + ndst = 1; } else if (backend_ctx->gpu_family == ADRENO) { - nth0 = 2; - nth1 = 64; + nth0 = 64; + nth1 = 2; + ndst = 1; } else { GGML_ASSERT(false && "TODO: Unknown GPU"); } @@ -8675,6 +8918,7 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int), &ne1)); CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int), &r2)); CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int), &r3)); +#endif // GGML_OPENCL_SOA_Q break; case GGML_TYPE_MXFP4: { #ifdef GGML_OPENCL_SOA_Q @@ -8777,7 +9021,7 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co } else if (src0t == GGML_TYPE_Q5_K) { GGML_ASSERT(false && "not implemented"); } else if (src0t == GGML_TYPE_Q6_K) { - size_t global_work_size[] = {(size_t)(ne01+1)/2*nth0, (size_t)ne11*nth1, (size_t)ne12*ne13}; + size_t global_work_size[] = {(size_t)(ne01+ndst*nth1-1)/(ndst*nth1)*nth0, (size_t)ne11*nth1, (size_t)ne12*ne13}; size_t local_work_size[] = {(size_t)nth0, (size_t)nth1, 1}; backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst); diff --git a/ggml/src/ggml-opencl/kernels/cvt.cl b/ggml/src/ggml-opencl/kernels/cvt.cl index 513a4d3e28..adf576a839 100644 --- a/ggml/src/ggml-opencl/kernels/cvt.cl +++ b/ggml/src/ggml-opencl/kernels/cvt.cl @@ -46,6 +46,16 @@ struct block_q4_0 uint8_t qs[QK4_0 / 2]; }; +//------------------------------------------------------------------------------ +// block_q6_K +//------------------------------------------------------------------------------ +struct block_q6_K { + uint8_t ql[QK_K/2]; // quants, lower 4 bits + uint8_t qh[QK_K/4]; // quants, upper 2 bits + int8_t scales[QK_K/16]; // scales, quantized with 8 bits + half d; // super-block scale +}; + //------------------------------------------------------------------------------ // kernel_convert_block_q4_0 // Convert the block_q4_0 format to 2 separate arrays (AOS -> SOA). @@ -263,3 +273,63 @@ kernel void kernel_restore_block_q8_0( b->qs[i] = q[i]; } } + +//------------------------------------------------------------------------------ +// kernel_convert_block_q6_K +// Convert the block_q6_K format to 3 separate arrays (AOS -> SOA). +// This kernel does not deshuffle the bits. +// Each thread processes a super block. +//------------------------------------------------------------------------------ +kernel void kernel_convert_block_q6_K( + global struct block_q6_K * src0, + global uchar * dst_ql, + global uchar * dst_qh, + global char * dst_s, + global half * dst_d +) { + global struct block_q6_K * b = (global struct block_q6_K *) src0 + get_global_id(0); + global uchar * ql = (global uchar *) dst_ql + QK_K/2*get_global_id(0); + global uchar * qh = (global uchar *) dst_qh + QK_K/4*get_global_id(0); + global char * s = (global char *) dst_s + QK_K/16*get_global_id(0); + global half * d = (global half *) dst_d + get_global_id(0); + + *d = b->d; + + for (int i = 0; i < QK_K/2; ++i) { + ql[i] = b->ql[i]; + } + for (int i = 0; i < QK_K/4; ++i) { + qh[i] = b->qh[i]; + } + for (int i = 0; i < QK_K/16; ++i) { + s[i] = b->scales[i]; + } +} + +// Restore block_q6_K from flattened arrays. +// Each thread processes a super block. +kernel void kernel_restore_block_q6_K( + global uchar * dst_ql, + global uchar * dst_qh, + global char * dst_s, + global half * dst_d, + global struct block_q6_K * dst +) { + global struct block_q6_K * b = (global struct block_q6_K *) dst + get_global_id(0); + global uchar * ql = (global uchar *) dst_ql + QK_K/2*get_global_id(0); + global uchar * qh = (global uchar *) dst_qh + QK_K/4*get_global_id(0); + global char * s = (global char *) dst_s + QK_K/16*get_global_id(0); + global half * d = (global half *) dst_d + get_global_id(0); + + b->d = *d; + + for (int i = 0; i < QK_K/2; ++i) { + b->ql[i] = ql[i]; + } + for (int i = 0; i < QK_K/4; ++i) { + b->qh[i] = qh[i]; + } + for (int i = 0; i < QK_K/16; ++i) { + b->scales[i] = s[i]; + } +} diff --git a/ggml/src/ggml-opencl/kernels/mul_mv_q6_k.cl b/ggml/src/ggml-opencl/kernels/mul_mv_q6_k_f32.cl similarity index 100% rename from ggml/src/ggml-opencl/kernels/mul_mv_q6_k.cl rename to ggml/src/ggml-opencl/kernels/mul_mv_q6_k_f32.cl diff --git a/ggml/src/ggml-opencl/kernels/mul_mv_q6_k_f32_flat.cl b/ggml/src/ggml-opencl/kernels/mul_mv_q6_k_f32_flat.cl new file mode 100644 index 0000000000..86fe09c6dd --- /dev/null +++ b/ggml/src/ggml-opencl/kernels/mul_mv_q6_k_f32_flat.cl @@ -0,0 +1,194 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +#ifdef cl_intel_subgroups +#pragma OPENCL EXTENSION cl_intel_subgroups : enable +#else +#pragma OPENCL EXTENSION cl_khr_subgroups : enable +#endif + +#ifdef cl_intel_required_subgroup_size +#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable +#define INTEL_GPU 1 +#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16))) +#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32))) +#elif defined(cl_qcom_reqd_sub_group_size) +#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable +#define ADRENO_GPU 1 +#define REQD_SUBGROUP_SIZE_64 __attribute__((qcom_reqd_sub_group_size("half"))) +#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full"))) +#endif + +//------------------------------------------------------------------------------ +// kernel_mul_mv_q6_K_f32_flat +//------------------------------------------------------------------------------ +#define Q6_K_MASK1 0x03 +#define Q6_K_MASK2 0x0C +#define Q6_K_MASK3 0x30 +#define Q6_K_MASK4 0xC0 + +#define QK_K 256 + +inline float block_q_6_K_dot_y_flat( + global uchar * blk_ql, + global uchar * blk_qh, + global char * blk_scales, + global half * blk_d, + global float * yy, + int ib, + int ip, + int is, + int l0 +) { + int y_offset = 128*ip + l0; + int q_offset_l = 64*ip + l0; + int q_offset_h = 32*ip + l0; + + global uchar * q1 = blk_ql + ib*128 + q_offset_l; + global uchar * q2 = q1 + QK_K/8; + global uchar * qh = blk_qh + ib*64 + q_offset_h; + global char * sc = blk_scales + ib*16 + is; + + global float * y = yy + ib * QK_K + y_offset; + + float dall = blk_d[ib]; + + float sumf = 0; + float4 sums = {0.f, 0.f, 0.f, 0.f}; + + sums.s0 += y[0+ 0] * ((float)((q1[0] & 0xF) | ((qh[0] & Q6_K_MASK1) << 4)) - 32.f); + sums.s1 += y[0+32] * ((float)((q2[0] & 0xF) | ((qh[0] & Q6_K_MASK2) << 2)) - 32.f); + sums.s2 += y[0+64] * ((float)((q1[0] >> 4) | ((qh[0] & Q6_K_MASK3) << 0)) - 32.f); + sums.s3 += y[0+96] * ((float)((q2[0] >> 4) | ((qh[0] & Q6_K_MASK4) >> 2)) - 32.f); + + sums.s0 += y[1+ 0] * ((float)((q1[1] & 0xF) | ((qh[1] & Q6_K_MASK1) << 4)) - 32.f); + sums.s1 += y[1+32] * ((float)((q2[1] & 0xF) | ((qh[1] & Q6_K_MASK2) << 2)) - 32.f); + sums.s2 += y[1+64] * ((float)((q1[1] >> 4) | ((qh[1] & Q6_K_MASK3) << 0)) - 32.f); + sums.s3 += y[1+96] * ((float)((q2[1] >> 4) | ((qh[1] & Q6_K_MASK4) >> 2)) - 32.f); + + sums.s0 += y[2+ 0] * ((float)((q1[2] & 0xF) | ((qh[2] & Q6_K_MASK1) << 4)) - 32.f); + sums.s1 += y[2+32] * ((float)((q2[2] & 0xF) | ((qh[2] & Q6_K_MASK2) << 2)) - 32.f); + sums.s2 += y[2+64] * ((float)((q1[2] >> 4) | ((qh[2] & Q6_K_MASK3) << 0)) - 32.f); + sums.s3 += y[2+96] * ((float)((q2[2] >> 4) | ((qh[2] & Q6_K_MASK4) >> 2)) - 32.f); + + sums.s0 += y[3+ 0] * ((float)((q1[3] & 0xF) | ((qh[3] & Q6_K_MASK1) << 4)) - 32.f); + sums.s1 += y[3+32] * ((float)((q2[3] & 0xF) | ((qh[3] & Q6_K_MASK2) << 2)) - 32.f); + sums.s2 += y[3+64] * ((float)((q1[3] >> 4) | ((qh[3] & Q6_K_MASK3) << 0)) - 32.f); + sums.s3 += y[3+96] * ((float)((q2[3] >> 4) | ((qh[3] & Q6_K_MASK4) >> 2)) - 32.f); + + sumf += dall * (sums.s0 * sc[0] + sums.s1 * sc[2] + sums.s2 * sc[4] + sums.s3 * sc[6]); + + return sumf; +} + +#undef N_DST +#undef N_SIMDGROUP +#undef N_SIMDWIDTH + +#ifdef INTEL_GPU +#define N_DST 4 +#define N_SIMDGROUP 2 +#define N_SIMDWIDTH 16 +#elif defined (ADRENO_GPU) +#define N_DST 4 +#define N_SIMDGROUP 2 +#define N_SIMDWIDTH 64 +#endif + +#define BLOCK_STRIDE (N_SIMDWIDTH/16) // number of blocks each subgroup processes + +#ifdef INTEL_GPU +REQD_SUBGROUP_SIZE_16 +#elif defined (ADRENO_GPU) +REQD_SUBGROUP_SIZE_64 +#endif +kernel void kernel_mul_mv_q6_K_f32_flat( + global uchar * src0_ql, + global uchar * src0_qh, + global char * src0_s, + global half * src0_d, + global float * src1, + ulong offset1, + global float * dst, + ulong offsetd, + int ne00, + int ne01, + int ne02, + int ne10, + int ne12, + int ne0, + int ne1, + int r2, + int r3 +) { + src1 = (global float*)((global char*)src1 + offset1); + dst = (global float*)((global char*)dst + offsetd); + + int nb = ne00/QK_K; + + int r0 = get_group_id(0); + int r1 = get_group_id(1); + int im = get_group_id(2); + + int i12 = im%ne12; + int i13 = im/ne12; + + int first_row = (N_SIMDGROUP * r0 + get_sub_group_id()) * N_DST; + + ulong offset_src0 = first_row*nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02); + ulong offset_src0_ql = offset_src0 * 128; + ulong offset_src0_qh = offset_src0 * 64; + ulong offset_src0_s = offset_src0 * 16; + ulong offset_src0_d = offset_src0; + + global uchar * blk_ql = (global uchar *) src0_ql + offset_src0_ql; + global uchar * blk_qh = (global uchar *) src0_qh + offset_src0_qh; + global char * blk_scales = (global char *) src0_s + offset_src0_s; + global half * blk_d = (global half *) src0_d + offset_src0_d; + global float * yy = (global float *) src1 + r1*ne10 + im*ne00*ne1; + + int tid = get_sub_group_local_id()/BLOCK_STRIDE; // first block_stride groups have tid=0 + int ix = get_sub_group_local_id()%BLOCK_STRIDE; // first block is 0..block_stride-1 + int ip = tid/8; // first or second half of (super) block (0 or 1) + int il = tid%8; // each half has 8 parts, one per scale + int n = 4; // 4 scales at a time (and 4 sums) + int l0 = n*il; // offset into half-block, 0..28 + int is = 8*ip + l0/16; // 0, 1, 8, 9 + + float4 sumf = 0; + + for (int ib = ix; ib < nb; ib += BLOCK_STRIDE) { + if (first_row + 0 < ne01) { + sumf.s0 += block_q_6_K_dot_y_flat(blk_ql + 0*nb*128, blk_qh + 0*nb*64, blk_scales + 0*nb*16, blk_d + 0*nb, yy, ib, ip, is, l0); + } + if (first_row + 1 < ne01) { + sumf.s1 += block_q_6_K_dot_y_flat(blk_ql + 1*nb*128, blk_qh + 1*nb*64, blk_scales + 1*nb*16, blk_d + 1*nb, yy, ib, ip, is, l0); + } + if (first_row + 2 < ne01) { + sumf.s2 += block_q_6_K_dot_y_flat(blk_ql + 2*nb*128, blk_qh + 2*nb*64, blk_scales + 2*nb*16, blk_d + 2*nb, yy, ib, ip, is, l0); + } + if (first_row + 3 < ne01) { + sumf.s3 += block_q_6_K_dot_y_flat(blk_ql + 3*nb*128, blk_qh + 3*nb*64, blk_scales + 3*nb*16, blk_d + 3*nb, yy, ib, ip, is, l0); + } + } + + float4 tot = (float4)( + sub_group_reduce_add(sumf.s0), + sub_group_reduce_add(sumf.s1), + sub_group_reduce_add(sumf.s2), + sub_group_reduce_add(sumf.s3) + ); + if (get_sub_group_local_id() == 0) { + if (first_row + 0 < ne01) { + dst[r1*ne0 + im*ne0*ne1 + first_row + 0] = tot.s0; + } + if (first_row + 1 < ne01) { + dst[r1*ne0 + im*ne0*ne1 + first_row + 1] = tot.s1; + } + if (first_row + 2 < ne01) { + dst[r1*ne0 + im*ne0*ne1 + first_row + 2] = tot.s2; + } + if (first_row + 3 < ne01) { + dst[r1*ne0 + im*ne0*ne1 + first_row + 3] = tot.s3; + } + } +} diff --git a/ggml/src/ggml-sycl/dpct/helper.hpp b/ggml/src/ggml-sycl/dpct/helper.hpp index 30ec1e8daf..8ae8098717 100644 --- a/ggml/src/ggml-sycl/dpct/helper.hpp +++ b/ggml/src/ggml-sycl/dpct/helper.hpp @@ -15,7 +15,6 @@ #include #include -#include #include #ifdef GGML_SYCL_USE_INTEL_ONEMKL diff --git a/ggml/src/ggml-sycl/ggml-sycl.cpp b/ggml/src/ggml-sycl/ggml-sycl.cpp index ce2f0d41c9..3a4c092af5 100644 --- a/ggml/src/ggml-sycl/ggml-sycl.cpp +++ b/ggml/src/ggml-sycl/ggml-sycl.cpp @@ -4606,14 +4606,12 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g return (op->type == GGML_TYPE_F32 && op->src[0]->type == GGML_TYPE_F32) && (op->type == op->src[0]->type); #endif case GGML_OP_NORM: - return true; case GGML_OP_L2_NORM: case GGML_OP_GROUP_NORM: - return ggml_is_contiguous(op->src[0]); case GGML_OP_RMS_NORM: - return ((op->src[0]->ne[0] % WARP_SIZE) == 0); + return true; case GGML_OP_RMS_NORM_BACK: - return ((op->src[0]->ne[0] % WARP_SIZE) == 0); + return ggml_is_contiguous(op->src[0]); case GGML_OP_SCALE: return true; case GGML_OP_CONT: diff --git a/ggml/src/ggml-sycl/norm.cpp b/ggml/src/ggml-sycl/norm.cpp index 823d3a4828..00702b5d09 100644 --- a/ggml/src/ggml-sycl/norm.cpp +++ b/ggml/src/ggml-sycl/norm.cpp @@ -251,7 +251,6 @@ static void norm_f32_sycl(const float * x, float * dst, const int ncols, const i const float eps, queue_ptr stream, int device) { const sycl::range<3> global_dims(nsamples, nchannels, nrows); - GGML_ASSERT(ncols % WARP_SIZE == 0); if (ncols < 1024) { const sycl::range<3> block_dims(1, 1, WARP_SIZE); stream->submit([&](sycl::handler& cgh) { @@ -334,7 +333,6 @@ static void group_norm_f32_sycl(const float* x, float* dst, static void rms_norm_f32_sycl(const float* x, float* dst, const int ncols, const int nrows, const int nchannels, const int nsamples, const int64_t stride_row, const int64_t stride_channel, const int64_t stride_sample, const float eps, queue_ptr stream, int device) { - GGML_ASSERT(ncols % WARP_SIZE == 0); // printf("%s ncols=%d, nrows=%d, WARP_SIZE=%d\n", __func__, ncols, nrows, WARP_SIZE); const sycl::range<3> global_dims(nsamples, nchannels, nrows); @@ -374,7 +372,6 @@ static void rms_norm_f32_sycl(const float* x, float* dst, const int ncols, const static void l2_norm_f32_sycl(const float* x, float* dst, const int ncols, const int nrows, const float eps, queue_ptr stream, int device) { - GGML_ASSERT(ncols % WARP_SIZE == 0); // printf("%s ncols=%d, nrows=%d, WARP_SIZE=%d\n", __func__, ncols, nrows, WARP_SIZE); if (ncols < 1024) { const sycl::range<3> block_dims(1, 1, WARP_SIZE); diff --git a/ggml/src/ggml-virtgpu/CMakeLists.txt b/ggml/src/ggml-virtgpu/CMakeLists.txt new file mode 100644 index 0000000000..e6b020beb5 --- /dev/null +++ b/ggml/src/ggml-virtgpu/CMakeLists.txt @@ -0,0 +1,70 @@ +cmake_minimum_required(VERSION 3.19) +cmake_policy(SET CMP0114 NEW) + +include(ExternalProject) + +message(STATUS "Including the VirtGPU/Virglrenderer API Remoting") + +# Download venus_hw.h from virglrenderer repository +ExternalProject_Add( + venus_hw_header + URL https://gitlab.freedesktop.org/virgl/virglrenderer/-/raw/virglrenderer-1.2.0/src/venus_hw.h + DOWNLOAD_NO_EXTRACT YES + DOWNLOAD_DIR ${CMAKE_CURRENT_SOURCE_DIR}/include + DOWNLOAD_NAME venus_hw.h + CONFIGURE_COMMAND "" + BUILD_COMMAND "" + INSTALL_COMMAND "" + LOG_DOWNLOAD ON +) + +if (NOT GGML_VIRTGPU_BACKEND STREQUAL "ONLY") + message(STATUS "Enable the VirtGPU/Virglrenderer API Remoting frontend library") + + find_package(PkgConfig REQUIRED) + pkg_check_modules(DRM REQUIRED libdrm) + if (NOT GGML_BACKEND_DL) + # cannot simply use USE_VIRTGPU, as in the 'else()' case the + # frontend isn't compiled + target_compile_definitions(ggml PUBLIC "GGML_USE_VIRTGPU_FRONTEND") + endif() + + ggml_add_backend_library(ggml-virtgpu + ggml-backend-buffer.cpp + ggml-backend.cpp + ggml-backend-device.cpp + ggml-backend-reg.cpp + ggml-backend-buffer-type.cpp + virtgpu-apir.h + virtgpu-forward.gen.h + virtgpu.cpp + virtgpu-shm.cpp + virtgpu-utils.cpp + virtgpu-forward-device.cpp + virtgpu-forward-buffer-type.cpp + virtgpu-forward-buffer.cpp + virtgpu-forward-backend.cpp + virtgpu-forward-impl.h + apir_cs_ggml-rpc-front.cpp + ../../include/ggml-virtgpu.h) + + target_include_directories(ggml-virtgpu PUBLIC /usr/include/libdrm/) + + target_link_libraries(ggml-virtgpu PUBLIC ${DRM_LIBRARIES}) + target_include_directories(ggml-virtgpu PUBLIC ${DRM_INCLUDE_DIRS}) + target_compile_options(ggml-virtgpu PUBLIC ${DRM_CFLAGS_OTHER}) + + target_include_directories(ggml-virtgpu PUBLIC ./include) + target_include_directories(ggml-virtgpu PRIVATE ${CMAKE_CURRENT_BINARY_DIR}) + + # Ensure venus_hw.h is downloaded before building ggml-virtgpu + add_dependencies(ggml-virtgpu venus_hw_header) + + target_compile_options(ggml-virtgpu PRIVATE -std=c++20) +else() + message(STATUS "Not building the VirtGPU/Virglrenderer API Remoting frontend library") +endif() + +if (NOT GGML_VIRTGPU_BACKEND STREQUAL "OFF") + add_subdirectory("backend") +endif() diff --git a/ggml/src/ggml-virtgpu/apir_cs_ggml-rpc-front.cpp b/ggml/src/ggml-virtgpu/apir_cs_ggml-rpc-front.cpp new file mode 100644 index 0000000000..f60ae3556c --- /dev/null +++ b/ggml/src/ggml-virtgpu/apir_cs_ggml-rpc-front.cpp @@ -0,0 +1,87 @@ +#include "backend/shared/apir_cs_rpc.h" +#include "ggml-backend-impl.h" +#include "ggml-impl.h" +#include "ggml-remoting.h" + +#include +#include +#include +#include + +apir_rpc_tensor apir_serialize_tensor(const ggml_tensor * tensor) { + apir_rpc_tensor result; + result.id = reinterpret_cast(tensor); + result.type = tensor->type; + if (tensor->buffer) { + ggml_backend_buffer_t buffer = tensor->buffer; + + result.buffer = BUFFER_TO_HOST_HANDLE(buffer); + } else { + result.buffer = 0; + } + for (uint32_t i = 0; i < GGML_MAX_DIMS; i++) { + result.ne[i] = tensor->ne[i]; + result.nb[i] = tensor->nb[i]; + } + result.op = tensor->op; + for (uint32_t i = 0; i < GGML_MAX_OP_PARAMS / sizeof(int32_t); i++) { + result.op_params[i] = tensor->op_params[i]; + } + result.flags = tensor->flags; + for (uint32_t i = 0; i < GGML_MAX_SRC; i++) { + result.src[i] = reinterpret_cast(tensor->src[i]); + } + result.view_src = reinterpret_cast(tensor->view_src); + result.view_offs = tensor->view_offs; + result.data = reinterpret_cast(tensor->data); + if (tensor->data) { + if (!tensor->buffer) { + GGML_ABORT("tensor has data but not buffer"); + } + // tensor->data is serialized as an offset to the buffer base address + result.data -= reinterpret_cast(BUFFER_TO_GGML_CONTEXT(tensor->buffer)->base); + } + snprintf(result.name, GGML_MAX_NAME, "%s", tensor->name); + return result; +} + +void apir_add_tensor(ggml_tensor * tensor, + std::vector & tensors, + std::unordered_set & visited) { + if (tensor == nullptr) { + return; + } + if (visited.find(tensor) != visited.end()) { + return; + } + visited.insert(tensor); + for (int i = 0; i < GGML_MAX_SRC; i++) { + apir_add_tensor(tensor->src[i], tensors, visited); + } + apir_add_tensor(tensor->view_src, tensors, visited); + tensors.push_back(apir_serialize_tensor(tensor)); +} + +void apir_serialize_graph(const ggml_cgraph * cgraph, std::vector & output) { + uint32_t n_nodes = cgraph->n_nodes; + std::vector tensors; + std::unordered_set visited; + for (uint32_t i = 0; i < n_nodes; i++) { + apir_add_tensor(cgraph->nodes[i], tensors, visited); + } + // serialization format: + // | n_nodes (4 bytes) | nodes (n_nodes * sizeof(uint64_t) | n_tensors (4 bytes) | tensors (n_tensors * sizeof(apir_rpc_tensor)) | + uint32_t n_tensors = tensors.size(); + int output_size = + sizeof(uint32_t) + n_nodes * sizeof(uint64_t) + sizeof(uint32_t) + n_tensors * sizeof(apir_rpc_tensor); + output.resize(output_size, 0); + memcpy(output.data(), &n_nodes, sizeof(n_nodes)); + for (uint32_t i = 0; i < n_nodes; i++) { + memcpy(output.data() + sizeof(n_nodes) + i * sizeof(uint64_t), &cgraph->nodes[i], sizeof(uint64_t)); + } + uint32_t * out_ntensors = (uint32_t *) (output.data() + sizeof(n_nodes) + n_nodes * sizeof(uint64_t)); + *out_ntensors = n_tensors; + apir_rpc_tensor * out_tensors = + (apir_rpc_tensor *) (output.data() + sizeof(n_nodes) + n_nodes * sizeof(uint64_t) + sizeof(uint32_t)); + memcpy(out_tensors, tensors.data(), n_tensors * sizeof(apir_rpc_tensor)); +} diff --git a/ggml/src/ggml-virtgpu/backend/CMakeLists.txt b/ggml/src/ggml-virtgpu/backend/CMakeLists.txt new file mode 100644 index 0000000000..0b49c403b9 --- /dev/null +++ b/ggml/src/ggml-virtgpu/backend/CMakeLists.txt @@ -0,0 +1,21 @@ +cmake_minimum_required(VERSION 3.19) +cmake_policy(SET CMP0114 NEW) + +message(STATUS "Enable the VirtGPU/Virglrenderer backend library") + +ggml_add_backend_library(ggml-virtgpu-backend + backend.cpp + backend-dispatched.cpp + backend-dispatched-backend.cpp + backend-dispatched-device.cpp + backend-dispatched-buffer.cpp + backend-dispatched-buffer-type.cpp + shared/api_remoting.h + shared/apir_backend.h + shared/apir_cs.h + apir_cs_ggml-rpc-back.cpp) + +target_compile_options(ggml-virtgpu-backend PRIVATE -std=c++20) + +# Add include directory for ggml-backend-impl.h and other core headers +target_include_directories(ggml-virtgpu-backend PRIVATE ../..) diff --git a/ggml/src/ggml-virtgpu/backend/apir_cs_ggml-rpc-back.cpp b/ggml/src/ggml-virtgpu/backend/apir_cs_ggml-rpc-back.cpp new file mode 100644 index 0000000000..60a8a93bfb --- /dev/null +++ b/ggml/src/ggml-virtgpu/backend/apir_cs_ggml-rpc-back.cpp @@ -0,0 +1,115 @@ +#include "ggml-backend-impl.h" +#include "ggml-impl.h" +#include "shared/apir_cs_rpc.h" + +#include +#include +#include +#include + +std::unordered_set backend_buffers; + +void apir_track_backend_buffer(ggml_backend_buffer_t buffer) { + backend_buffers.insert(buffer); +} + +bool apir_untrack_backend_buffer(ggml_backend_buffer_t buffer) { + auto it = backend_buffers.find(buffer); + if (it == backend_buffers.end()) { + return false; + } + + backend_buffers.erase(it); + return true; +} + +std::unordered_set apir_get_track_backend_buffers() { + return backend_buffers; +} + +ggml_tensor * apir_deserialize_tensor(ggml_context * ctx, const apir_rpc_tensor * tensor) { + ggml_tensor * result = + ggml_new_tensor_4d(ctx, (ggml_type) tensor->type, tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3]); + for (uint32_t i = 0; i < GGML_MAX_DIMS; i++) { + result->nb[i] = tensor->nb[i]; + } + result->buffer = reinterpret_cast(tensor->buffer); + if (result->buffer && backend_buffers.find(result->buffer) == backend_buffers.end()) { + printf("WARNING: HOST BUFFER NOT FOUND | %p\n", (void *) result->buffer); + result->buffer = nullptr; + } + + uint64_t tensor_data = tensor->data; + if (result->buffer) { + // require that the tensor data does not go beyond the buffer end + uint64_t tensor_size = (uint64_t) ggml_nbytes(result); + uint64_t buffer_start = (uint64_t) ggml_backend_buffer_get_base(result->buffer); + uint64_t buffer_size = (uint64_t) ggml_backend_buffer_get_size(result->buffer); + + // tensor->data is serialized as an offset to the buffer base address + tensor_data += buffer_start; + + GGML_ASSERT(tensor_data + tensor_size >= tensor_data); // check for overflow + GGML_ASSERT(tensor_data >= buffer_start && tensor_data + tensor_size <= buffer_start + buffer_size); + } + + result->op = (ggml_op) tensor->op; + for (uint32_t i = 0; i < GGML_MAX_OP_PARAMS / sizeof(int32_t); i++) { + result->op_params[i] = tensor->op_params[i]; + } + result->flags = tensor->flags; + result->data = reinterpret_cast(tensor_data); + ggml_set_name(result, tensor->name); + return result; +} + +ggml_tensor * apir_create_node(uint64_t id, + ggml_context * ctx, + const std::unordered_map & tensor_ptrs, + std::unordered_map & tensor_map) { + if (id == 0) { + return nullptr; + } + if (tensor_map.find(id) != tensor_map.end()) { + return tensor_map[id]; + } + const apir_rpc_tensor * tensor = tensor_ptrs.at(id); + ggml_tensor * result = apir_deserialize_tensor(ctx, tensor); + if (result == nullptr) { + return nullptr; + } + tensor_map[id] = result; + for (int i = 0; i < GGML_MAX_SRC; i++) { + result->src[i] = apir_create_node(tensor->src[i], ctx, tensor_ptrs, tensor_map); + } + result->view_src = apir_create_node(tensor->view_src, ctx, tensor_ptrs, tensor_map); + result->view_offs = tensor->view_offs; + return result; +} + +ggml_cgraph * apir_deserialize_graph(uint32_t n_nodes, + uint32_t n_tensors, + const apir_rpc_tensor * tensors, + const uint64_t * nodes) { + size_t buf_size = ggml_tensor_overhead() * (n_nodes + n_tensors) + ggml_graph_overhead_custom(n_nodes, false); + ggml_init_params params = { + /*.mem_size =*/buf_size, + /*.mem_buffer =*/NULL, + /*.no_alloc =*/true, + }; + ggml_context * ctx = ggml_init(params); + ggml_cgraph * graph = ggml_new_graph_custom(ctx, n_nodes, false); + graph->n_nodes = n_nodes; + std::unordered_map tensor_ptrs; + for (uint32_t i = 0; i < n_tensors; i++) { + tensor_ptrs[tensors[i].id] = &tensors[i]; + } + std::unordered_map tensor_map; + for (uint32_t i = 0; i < n_nodes; i++) { + int64_t id; + memcpy(&id, &nodes[i], sizeof(id)); + graph->nodes[i] = apir_create_node(id, ctx, tensor_ptrs, tensor_map); + } + + return graph; +} diff --git a/ggml/src/ggml-virtgpu/backend/backend-convert.h b/ggml/src/ggml-virtgpu/backend/backend-convert.h new file mode 100644 index 0000000000..1978d21f7e --- /dev/null +++ b/ggml/src/ggml-virtgpu/backend/backend-convert.h @@ -0,0 +1,13 @@ +#include "shared/apir_backend.h" + +#define BUFFER_TO_HOST_HANDLE(name) ggml_buffer_to_apir_handle(name) + +static inline apir_buffer_host_handle_t ggml_buffer_to_apir_handle(ggml_backend_buffer_t buffer) { + // in the backend, the buffer handle is the buffer pointer + return (apir_buffer_host_handle_t) buffer; +} + +static inline apir_buffer_type_host_handle_t ggml_buffer_type_to_apir_handle(ggml_backend_buffer_type_t buft) { + // in the backend, the buffer handle is the buffer pointer + return (apir_buffer_type_host_handle_t) buft; +} diff --git a/ggml/src/ggml-virtgpu/backend/backend-dispatched-backend.cpp b/ggml/src/ggml-virtgpu/backend/backend-dispatched-backend.cpp new file mode 100644 index 0000000000..77b4ee71e1 --- /dev/null +++ b/ggml/src/ggml-virtgpu/backend/backend-dispatched-backend.cpp @@ -0,0 +1,65 @@ +#include "backend-dispatched.h" +#include "backend-virgl-apir.h" +#include "ggml-backend-impl.h" +#include "ggml-backend.h" +#include "ggml-impl.h" +#include "shared/apir_backend.h" + +#include + +uint32_t backend_backend_graph_compute(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) { + GGML_UNUSED(ctx); + GGML_UNUSED(enc); + + static bool async_backend_initialized = false; + static bool async_backend; + + if (!async_backend_initialized) { + ggml_backend_dev_props props; + + dev->iface.get_props(dev, &props); + async_backend = props.caps.async; + async_backend_initialized = true; + } + + uint32_t shmem_res_id; + apir_decode_virtgpu_shmem_res_id(dec, &shmem_res_id); + + const void * shmem_data = ctx->iface->get_shmem_ptr(ctx->ctx_id, shmem_res_id); + if (!shmem_data) { + GGML_LOG_ERROR("Couldn't get the shmem addr from virgl\n"); + apir_decoder_set_fatal(dec); + return 1; + } + size_t cgraph_size; + apir_decode_size_t(dec, &cgraph_size); + + apir_decoder secondary_dec = apir_new_decoder((const char *) shmem_data, cgraph_size); + + ggml_cgraph * cgraph = apir_decode_ggml_cgraph(&secondary_dec, cgraph_size); + + ggml_status status; +#if APIR_BACKEND_CHECK_SUPPORTS_OP == 1 + for (int idx = 0; idx < cgraph->n_nodes; idx++) { + ggml_tensor * op = ggml_graph_node(cgraph, idx); + if (dev->iface.supports_op(dev, op)) { + continue; + } + GGML_LOG_ERROR("Graph node %d (%s) not supported by the backend\n", idx, ggml_op_desc(op)); + + status = GGML_STATUS_ABORTED; + apir_encode_ggml_status(enc, &status); + + return 0; + } +#endif + status = bck->iface.graph_compute(bck, cgraph); + + if (async_backend) { + bck->iface.synchronize(bck); + } + + apir_encode_ggml_status(enc, &status); + + return 0; +} diff --git a/ggml/src/ggml-virtgpu/backend/backend-dispatched-buffer-type.cpp b/ggml/src/ggml-virtgpu/backend/backend-dispatched-buffer-type.cpp new file mode 100644 index 0000000000..8ea1bb4fb4 --- /dev/null +++ b/ggml/src/ggml-virtgpu/backend/backend-dispatched-buffer-type.cpp @@ -0,0 +1,89 @@ +#include "backend-dispatched.h" +#include "backend-virgl-apir.h" +#include "ggml-backend-impl.h" +#include "ggml-backend.h" +#include "ggml-impl.h" + +#include + +uint32_t backend_buffer_type_get_name(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) { + GGML_UNUSED(ctx); + ggml_backend_buffer_type_t buft; + buft = apir_decode_ggml_buffer_type(dec); + + const char * string = buft->iface.get_name(buft); + + const size_t string_size = strlen(string) + 1; + apir_encode_array_size(enc, string_size); + apir_encode_char_array(enc, string, string_size); + + return 0; +} + +uint32_t backend_buffer_type_get_alignment(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) { + GGML_UNUSED(ctx); + ggml_backend_buffer_type_t buft; + buft = apir_decode_ggml_buffer_type(dec); + + size_t value = buft->iface.get_alignment(buft); + apir_encode_size_t(enc, &value); + + return 0; +} + +uint32_t backend_buffer_type_get_max_size(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) { + GGML_UNUSED(ctx); + ggml_backend_buffer_type_t buft; + buft = apir_decode_ggml_buffer_type(dec); + + size_t value = buft->iface.get_max_size(buft); + apir_encode_size_t(enc, &value); + + return 0; +} + +uint32_t backend_buffer_type_is_host(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) { + GGML_UNUSED(ctx); + ggml_backend_buffer_type_t buft; + buft = apir_decode_ggml_buffer_type(dec); + + bool is_host = buft->iface.is_host(buft); + apir_encode_bool_t(enc, &is_host); + + return 0; +} + +uint32_t backend_buffer_type_alloc_buffer(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) { + GGML_UNUSED(ctx); + ggml_backend_buffer_type_t buft; + buft = apir_decode_ggml_buffer_type(dec); + + size_t size; + apir_decode_size_t(dec, &size); + + ggml_backend_buffer_t buffer; + + buffer = buft->iface.alloc_buffer(buft, size); + + apir_encode_ggml_buffer(enc, buffer); + + if (buffer) { + apir_track_backend_buffer(buffer); + } + + return 0; +} + +uint32_t backend_buffer_type_get_alloc_size(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) { + GGML_UNUSED(ctx); + ggml_backend_buffer_type_t buft; + buft = apir_decode_ggml_buffer_type(dec); + + const ggml_tensor * op = apir_decode_ggml_tensor_inplace(dec); + + size_t value = buft->iface.get_alloc_size(buft, op); + + apir_encode_size_t(enc, &value); + + return 0; +} diff --git a/ggml/src/ggml-virtgpu/backend/backend-dispatched-buffer.cpp b/ggml/src/ggml-virtgpu/backend/backend-dispatched-buffer.cpp new file mode 100644 index 0000000000..cf81888e98 --- /dev/null +++ b/ggml/src/ggml-virtgpu/backend/backend-dispatched-buffer.cpp @@ -0,0 +1,131 @@ +#include "backend-dispatched.h" +#include "backend-virgl-apir.h" +#include "ggml-backend-impl.h" +#include "ggml-backend.h" +#include "ggml-impl.h" + +#include + +uint32_t backend_buffer_get_base(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) { + GGML_UNUSED(ctx); + ggml_backend_buffer_t buffer; + buffer = apir_decode_ggml_buffer(dec); + + uintptr_t base = (uintptr_t) buffer->iface.get_base(buffer); + apir_encode_uintptr_t(enc, &base); + + return 0; +} + +uint32_t backend_buffer_set_tensor(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) { + GGML_UNUSED(ctx); + GGML_UNUSED(enc); + + ggml_backend_buffer_t buffer; + buffer = apir_decode_ggml_buffer(dec); + + ggml_tensor * tensor; + // safe to remove the const qualifier here + tensor = (ggml_tensor *) (uintptr_t) apir_decode_ggml_tensor(dec); + + uint32_t shmem_res_id; + apir_decode_virtgpu_shmem_res_id(dec, &shmem_res_id); + + size_t offset; + apir_decode_size_t(dec, &offset); + + size_t size; + apir_decode_size_t(dec, &size); + + void * shmem_data = ctx->iface->get_shmem_ptr(ctx->ctx_id, shmem_res_id); + + if (!shmem_data) { + GGML_LOG_ERROR("Couldn't get the shmem addr from virgl\n"); + return 1; + } + + buffer->iface.set_tensor(buffer, tensor, shmem_data, offset, size); + + return 0; +} + +uint32_t backend_buffer_get_tensor(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) { + GGML_UNUSED(ctx); + GGML_UNUSED(enc); + + ggml_backend_buffer_t buffer; + buffer = apir_decode_ggml_buffer(dec); + + const ggml_tensor * tensor; + // safe to remove the const qualifier here + tensor = apir_decode_ggml_tensor(dec); + + uint32_t shmem_res_id; + apir_decode_virtgpu_shmem_res_id(dec, &shmem_res_id); + + size_t offset; + apir_decode_size_t(dec, &offset); + + size_t size; + apir_decode_size_t(dec, &size); + + void * shmem_data = ctx->iface->get_shmem_ptr(ctx->ctx_id, shmem_res_id); + if (!shmem_data) { + GGML_LOG_ERROR("Couldn't get the shmem addr from virgl\n"); + return 1; + } + + buffer->iface.get_tensor(buffer, tensor, shmem_data, offset, size); + + return 0; +} + +uint32_t backend_buffer_cpy_tensor(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) { + GGML_UNUSED(ctx); + + ggml_backend_buffer_t buffer; + buffer = apir_decode_ggml_buffer(dec); + + const ggml_tensor * src; + // safe to remove the const qualifier here + src = apir_decode_ggml_tensor(dec); + ggml_tensor * dst = (ggml_tensor *) (uintptr_t) apir_decode_ggml_tensor(dec); + + bool ret = buffer->iface.cpy_tensor(buffer, src, (ggml_tensor *) dst); + + apir_encode_bool_t(enc, &ret); + + return 0; +} + +uint32_t backend_buffer_clear(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) { + GGML_UNUSED(ctx); + GGML_UNUSED(enc); + + ggml_backend_buffer_t buffer; + buffer = apir_decode_ggml_buffer(dec); + + uint8_t value; + apir_decode_uint8_t(dec, &value); + + buffer->iface.clear(buffer, value); + + return 0; +} + +uint32_t backend_buffer_free_buffer(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) { + GGML_UNUSED(ctx); + GGML_UNUSED(enc); + + ggml_backend_buffer_t buffer; + buffer = apir_decode_ggml_buffer(dec); + + if (!apir_untrack_backend_buffer(buffer)) { + GGML_LOG_WARN("%s: unknown buffer %p\n", __func__, (void *) buffer); + return 1; + } + + buffer->iface.free_buffer(buffer); + + return 0; +} diff --git a/ggml/src/ggml-virtgpu/backend/backend-dispatched-device.cpp b/ggml/src/ggml-virtgpu/backend/backend-dispatched-device.cpp new file mode 100644 index 0000000000..497f737a88 --- /dev/null +++ b/ggml/src/ggml-virtgpu/backend/backend-dispatched-device.cpp @@ -0,0 +1,148 @@ +#include "backend-dispatched.h" +#include "backend-virgl-apir.h" +#include "ggml-backend-impl.h" +#include "ggml-backend.h" +#include "ggml-impl.h" + +#include + +uint32_t backend_device_get_device_count(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) { + GGML_UNUSED(ctx); + GGML_UNUSED(ctx); + GGML_UNUSED(dec); + + int32_t dev_count = reg->iface.get_device_count(reg); + apir_encode_int32_t(enc, &dev_count); + + return 0; +} + +uint32_t backend_device_get_count(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) { + GGML_UNUSED(ctx); + GGML_UNUSED(ctx); + GGML_UNUSED(dec); + + int32_t dev_count = reg->iface.get_device_count(reg); + apir_encode_int32_t(enc, &dev_count); + + return 0; +} + +uint32_t backend_device_get_name(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) { + GGML_UNUSED(ctx); + GGML_UNUSED(dec); + + const char * string = dev->iface.get_name(dev); + + const size_t string_size = strlen(string) + 1; + apir_encode_array_size(enc, string_size); + apir_encode_char_array(enc, string, string_size); + + return 0; +} + +uint32_t backend_device_get_description(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) { + GGML_UNUSED(ctx); + GGML_UNUSED(dec); + + const char * string = dev->iface.get_description(dev); + + const size_t string_size = strlen(string) + 1; + apir_encode_array_size(enc, string_size); + apir_encode_char_array(enc, string, string_size); + + return 0; +} + +uint32_t backend_device_get_type(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) { + GGML_UNUSED(ctx); + GGML_UNUSED(dec); + + uint32_t type = dev->iface.get_type(dev); + apir_encode_uint32_t(enc, &type); + + return 0; +} + +uint32_t backend_device_get_memory(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) { + GGML_UNUSED(ctx); + GGML_UNUSED(dec); + + size_t free, total; + dev->iface.get_memory(dev, &free, &total); + + apir_encode_size_t(enc, &free); + apir_encode_size_t(enc, &total); + + return 0; +} + +uint32_t backend_device_supports_op(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) { + GGML_UNUSED(ctx); + + const ggml_tensor * op = apir_decode_ggml_tensor_inplace(dec); + + bool supports_op = dev->iface.supports_op(dev, op); + + apir_encode_bool_t(enc, &supports_op); + + return 0; +} + +uint32_t backend_device_get_buffer_type(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) { + GGML_UNUSED(ctx); + GGML_UNUSED(dec); + + ggml_backend_buffer_type_t bufft = dev->iface.get_buffer_type(dev); + + apir_encode_ggml_buffer_type(enc, bufft); + + return 0; +} + +uint32_t backend_device_get_props(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) { + GGML_UNUSED(ctx); + GGML_UNUSED(dec); + + ggml_backend_dev_props props; + dev->iface.get_props(dev, &props); + + apir_encode_bool_t(enc, &props.caps.async); + apir_encode_bool_t(enc, &props.caps.host_buffer); + apir_encode_bool_t(enc, &props.caps.buffer_from_host_ptr); + apir_encode_bool_t(enc, &props.caps.events); + + return 0; +} + +uint32_t backend_device_buffer_from_ptr(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) { + GGML_UNUSED(ctx); + GGML_UNUSED(dec); + + uint32_t shmem_res_id; + apir_decode_virtgpu_shmem_res_id(dec, &shmem_res_id); + + void * shmem_ptr = ctx->iface->get_shmem_ptr(ctx->ctx_id, shmem_res_id); + if (!shmem_ptr) { + GGML_LOG_ERROR("Couldn't get the shmem addr from virgl\n"); + apir_decoder_set_fatal(dec); + return 1; + } + + size_t size; + apir_decode_size_t(dec, &size); + size_t max_tensor_size; + apir_decode_size_t(dec, &max_tensor_size); + + ggml_backend_buffer_t buffer; + buffer = dev->iface.buffer_from_host_ptr(dev, shmem_ptr, size, max_tensor_size); + + apir_encode_ggml_buffer(enc, buffer); + apir_encode_ggml_buffer_type(enc, buffer->buft); + + if (buffer) { + apir_track_backend_buffer(buffer); + } + + return 0; +} diff --git a/ggml/src/ggml-virtgpu/backend/backend-dispatched.cpp b/ggml/src/ggml-virtgpu/backend/backend-dispatched.cpp new file mode 100644 index 0000000000..51d445725f --- /dev/null +++ b/ggml/src/ggml-virtgpu/backend/backend-dispatched.cpp @@ -0,0 +1,46 @@ +#include "backend-dispatched.h" +#include "backend-virgl-apir.h" + +#include "ggml-backend-impl.h" +#include "ggml-backend.h" +#include "ggml-impl.h" + +#include + +ggml_backend_reg_t reg = NULL; +ggml_backend_dev_t dev = NULL; +ggml_backend_t bck = NULL; + +uint64_t timer_start = 0; +uint64_t timer_total = 0; +uint64_t timer_count = 0; + +uint32_t backend_dispatch_initialize(void * ggml_backend_reg_fct_p) { + if (reg != NULL) { + GGML_LOG_WARN("%s: already initialized\n", __func__); + return APIR_BACKEND_INITIALIZE_ALREADY_INITED; + } + ggml_backend_reg_t (*ggml_backend_reg_fct)(void) = (ggml_backend_reg_t (*)()) ggml_backend_reg_fct_p; + + reg = ggml_backend_reg_fct(); + if (reg == NULL) { + GGML_LOG_ERROR("%s: backend registration failed\n", __func__); + return APIR_BACKEND_INITIALIZE_BACKEND_REG_FAILED; + } + + if (!reg->iface.get_device_count(reg)) { + GGML_LOG_ERROR("%s: backend initialization failed: no device found\n", __func__); + return APIR_BACKEND_INITIALIZE_NO_DEVICE; + } + + dev = reg->iface.get_device(reg, 0); + + if (!dev) { + GGML_LOG_ERROR("%s: backend initialization failed: no device received\n", __func__); + return APIR_BACKEND_INITIALIZE_NO_DEVICE; + } + + bck = dev->iface.init_backend(dev, NULL); + + return APIR_BACKEND_INITIALIZE_SUCCESS; +} diff --git a/ggml/src/ggml-virtgpu/backend/backend-dispatched.gen.h b/ggml/src/ggml-virtgpu/backend/backend-dispatched.gen.h new file mode 100644 index 0000000000..b81fd5039b --- /dev/null +++ b/ggml/src/ggml-virtgpu/backend/backend-dispatched.gen.h @@ -0,0 +1,130 @@ +#pragma once + +/* device */ +uint32_t backend_device_get_device_count(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx); +uint32_t backend_device_get_count(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx); +uint32_t backend_device_get_name(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx); +uint32_t backend_device_get_description(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx); +uint32_t backend_device_get_type(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx); +uint32_t backend_device_get_memory(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx); +uint32_t backend_device_supports_op(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx); +uint32_t backend_device_get_buffer_type(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx); +uint32_t backend_device_get_props(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx); +uint32_t backend_device_buffer_from_ptr(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx); + +/* buffer-type */ +uint32_t backend_buffer_type_get_name(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx); +uint32_t backend_buffer_type_get_alignment(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx); +uint32_t backend_buffer_type_get_max_size(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx); +uint32_t backend_buffer_type_is_host(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx); +uint32_t backend_buffer_type_alloc_buffer(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx); +uint32_t backend_buffer_type_get_alloc_size(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx); + +/* buffer */ +uint32_t backend_buffer_get_base(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx); +uint32_t backend_buffer_set_tensor(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx); +uint32_t backend_buffer_get_tensor(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx); +uint32_t backend_buffer_cpy_tensor(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx); +uint32_t backend_buffer_clear(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx); +uint32_t backend_buffer_free_buffer(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx); + +/* backend */ +uint32_t backend_backend_graph_compute(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx); + +static inline const char * backend_dispatch_command_name(ApirBackendCommandType type) { + switch (type) { + /* device */ + case APIR_COMMAND_TYPE_DEVICE_GET_DEVICE_COUNT: + return "backend_device_get_device_count"; + case APIR_COMMAND_TYPE_DEVICE_GET_COUNT: + return "backend_device_get_count"; + case APIR_COMMAND_TYPE_DEVICE_GET_NAME: + return "backend_device_get_name"; + case APIR_COMMAND_TYPE_DEVICE_GET_DESCRIPTION: + return "backend_device_get_description"; + case APIR_COMMAND_TYPE_DEVICE_GET_TYPE: + return "backend_device_get_type"; + case APIR_COMMAND_TYPE_DEVICE_GET_MEMORY: + return "backend_device_get_memory"; + case APIR_COMMAND_TYPE_DEVICE_SUPPORTS_OP: + return "backend_device_supports_op"; + case APIR_COMMAND_TYPE_DEVICE_GET_BUFFER_TYPE: + return "backend_device_get_buffer_type"; + case APIR_COMMAND_TYPE_DEVICE_GET_PROPS: + return "backend_device_get_props"; + case APIR_COMMAND_TYPE_DEVICE_BUFFER_FROM_PTR: + return "backend_device_buffer_from_ptr"; + /* buffer-type */ + case APIR_COMMAND_TYPE_BUFFER_TYPE_GET_NAME: + return "backend_buffer_type_get_name"; + case APIR_COMMAND_TYPE_BUFFER_TYPE_GET_ALIGNMENT: + return "backend_buffer_type_get_alignment"; + case APIR_COMMAND_TYPE_BUFFER_TYPE_GET_MAX_SIZE: + return "backend_buffer_type_get_max_size"; + case APIR_COMMAND_TYPE_BUFFER_TYPE_IS_HOST: + return "backend_buffer_type_is_host"; + case APIR_COMMAND_TYPE_BUFFER_TYPE_ALLOC_BUFFER: + return "backend_buffer_type_alloc_buffer"; + case APIR_COMMAND_TYPE_BUFFER_TYPE_GET_ALLOC_SIZE: + return "backend_buffer_type_get_alloc_size"; + /* buffer */ + case APIR_COMMAND_TYPE_BUFFER_GET_BASE: + return "backend_buffer_get_base"; + case APIR_COMMAND_TYPE_BUFFER_SET_TENSOR: + return "backend_buffer_set_tensor"; + case APIR_COMMAND_TYPE_BUFFER_GET_TENSOR: + return "backend_buffer_get_tensor"; + case APIR_COMMAND_TYPE_BUFFER_CPY_TENSOR: + return "backend_buffer_cpy_tensor"; + case APIR_COMMAND_TYPE_BUFFER_CLEAR: + return "backend_buffer_clear"; + case APIR_COMMAND_TYPE_BUFFER_FREE_BUFFER: + return "backend_buffer_free_buffer"; + /* backend */ + case APIR_COMMAND_TYPE_BACKEND_GRAPH_COMPUTE: + return "backend_backend_graph_compute"; + + default: + return "unknown"; + } +} + +extern "C" { +static const backend_dispatch_t apir_backend_dispatch_table[APIR_BACKEND_DISPATCH_TABLE_COUNT] = { + + /* device */ + + /* APIR_COMMAND_TYPE_DEVICE_GET_DEVICE_COUNT = */ backend_device_get_device_count, + /* APIR_COMMAND_TYPE_DEVICE_GET_COUNT = */ backend_device_get_count, + /* APIR_COMMAND_TYPE_DEVICE_GET_NAME = */ backend_device_get_name, + /* APIR_COMMAND_TYPE_DEVICE_GET_DESCRIPTION = */ backend_device_get_description, + /* APIR_COMMAND_TYPE_DEVICE_GET_TYPE = */ backend_device_get_type, + /* APIR_COMMAND_TYPE_DEVICE_GET_MEMORY = */ backend_device_get_memory, + /* APIR_COMMAND_TYPE_DEVICE_SUPPORTS_OP = */ backend_device_supports_op, + /* APIR_COMMAND_TYPE_DEVICE_GET_BUFFER_TYPE = */ backend_device_get_buffer_type, + /* APIR_COMMAND_TYPE_DEVICE_GET_PROPS = */ backend_device_get_props, + /* APIR_COMMAND_TYPE_DEVICE_BUFFER_FROM_PTR = */ backend_device_buffer_from_ptr, + + /* buffer-type */ + + /* APIR_COMMAND_TYPE_BUFFER_TYPE_GET_NAME = */ backend_buffer_type_get_name, + /* APIR_COMMAND_TYPE_BUFFER_TYPE_GET_ALIGNMENT = */ backend_buffer_type_get_alignment, + /* APIR_COMMAND_TYPE_BUFFER_TYPE_GET_MAX_SIZE = */ backend_buffer_type_get_max_size, + /* APIR_COMMAND_TYPE_BUFFER_TYPE_IS_HOST = */ backend_buffer_type_is_host, + /* APIR_COMMAND_TYPE_BUFFER_TYPE_ALLOC_BUFFER = */ backend_buffer_type_alloc_buffer, + /* APIR_COMMAND_TYPE_BUFFER_TYPE_GET_ALLOC_SIZE = */ backend_buffer_type_get_alloc_size, + + /* buffer */ + + /* APIR_COMMAND_TYPE_BUFFER_GET_BASE = */ backend_buffer_get_base, + /* APIR_COMMAND_TYPE_BUFFER_SET_TENSOR = */ backend_buffer_set_tensor, + /* APIR_COMMAND_TYPE_BUFFER_GET_TENSOR = */ backend_buffer_get_tensor, + /* APIR_COMMAND_TYPE_BUFFER_CPY_TENSOR = */ backend_buffer_cpy_tensor, + /* APIR_COMMAND_TYPE_BUFFER_CLEAR = */ backend_buffer_clear, + /* APIR_COMMAND_TYPE_BUFFER_FREE_BUFFER = */ backend_buffer_free_buffer, + + /* backend */ + + /* APIR_COMMAND_TYPE_BACKEND_GRAPH_COMPUTE = */ backend_backend_graph_compute, +}; +} diff --git a/ggml/src/ggml-virtgpu/backend/backend-dispatched.h b/ggml/src/ggml-virtgpu/backend/backend-dispatched.h new file mode 100644 index 0000000000..6ccbecf078 --- /dev/null +++ b/ggml/src/ggml-virtgpu/backend/backend-dispatched.h @@ -0,0 +1,23 @@ +#pragma once + +#include +#include + +#include + +#include "backend-convert.h" +#include "backend-virgl-apir.h" +#include "shared/apir_backend.h" +#include "shared/apir_cs.h" +#include "shared/apir_cs_ggml.h" + +struct virgl_apir_context { + uint32_t ctx_id; + virgl_apir_callbacks * iface; +}; + +typedef uint32_t (*backend_dispatch_t)(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx); + +#include "backend-dispatched.gen.h" + +uint32_t backend_dispatch_initialize(void * ggml_backend_reg_fct_p); diff --git a/ggml/src/ggml-virtgpu/backend/backend-virgl-apir.h b/ggml/src/ggml-virtgpu/backend/backend-virgl-apir.h new file mode 100644 index 0000000000..44b347f853 --- /dev/null +++ b/ggml/src/ggml-virtgpu/backend/backend-virgl-apir.h @@ -0,0 +1,32 @@ +#pragma once + +#include "ggml-backend-impl.h" +#include "ggml-backend.h" +#include "ggml-impl.h" +#include "shared/api_remoting.h" + +#include +#include +#include + +extern ggml_backend_reg_t reg; +extern ggml_backend_dev_t dev; +extern ggml_backend_t bck; + +struct virgl_apir_callbacks { + const char * (*get_config)(uint32_t virgl_ctx_id, const char * key); + void * (*get_shmem_ptr)(uint32_t virgl_ctx_id, uint32_t res_id); +}; + +extern "C" { +ApirLoadLibraryReturnCode apir_backend_initialize(uint32_t virgl_ctx_id, struct virgl_apir_callbacks *virgl_cbs); +void apir_backend_deinit(uint32_t virgl_ctx_id); +uint32_t apir_backend_dispatcher(uint32_t virgl_ctx_id, + virgl_apir_callbacks * virgl_cbs, + uint32_t cmd_type, + char * dec_cur, + const char * dec_end, + char * enc_cur, + const char * enc_end, + char ** enc_cur_after); +} diff --git a/ggml/src/ggml-virtgpu/backend/backend.cpp b/ggml/src/ggml-virtgpu/backend/backend.cpp new file mode 100644 index 0000000000..95d602ed60 --- /dev/null +++ b/ggml/src/ggml-virtgpu/backend/backend.cpp @@ -0,0 +1,148 @@ +#include "backend-dispatched.h" +#include "backend-virgl-apir.h" + +#include "shared/api_remoting.h" +#include "shared/apir_backend.h" +#include "shared/apir_cs.h" + +#include +#include + +#include + +#define APIR_LLAMA_CPP_GGML_LIBRARY_PATH_ENV "APIR_LLAMA_CPP_GGML_LIBRARY_PATH" +#define APIR_LLAMA_CPP_GGML_LIBRARY_REG_ENV "APIR_LLAMA_CPP_GGML_LIBRARY_REG" +#define APIR_LLAMA_CPP_LOG_TO_FILE_ENV "APIR_LLAMA_CPP_LOG_TO_FILE" + +#define GGML_DEFAULT_BACKEND_REG "ggml_backend_init" + +static void * backend_library_handle = NULL; +static FILE * apir_logfile = NULL; + +static void log_to_file_callback(enum ggml_log_level level, const char * text, void * user_data) { + FILE * logfile = (FILE *)user_data; + fprintf(logfile, "[%d] %s", level, text); + fflush(logfile); +} + +extern "C" { +void apir_backend_deinit(uint32_t virgl_ctx_id) { + GGML_UNUSED(virgl_ctx_id); + + auto buffers = apir_get_track_backend_buffers(); + for (const auto & buffer : buffers) { + apir_untrack_backend_buffer(buffer); + buffer->iface.free_buffer(buffer); + } + + if (dev) { + size_t free, total; + dev->iface.get_memory(dev, &free, &total); + GGML_LOG_INFO("%s: free memory: %ld MB\n", __func__, (size_t) free / 1024 / 1024); + } + + if (backend_library_handle) { + GGML_LOG_INFO("%s: The GGML backend library was loaded. Unloading it.\n", __func__); + dlclose(backend_library_handle); + backend_library_handle = NULL; + } + + if (apir_logfile) { + fclose(apir_logfile); + apir_logfile = NULL; + } +} + +#define APIR_GGML_LIBRARY_PATH_KEY "ggml.library.path" +#define APIR_GGML_LIBRARY_REG_KEY "ggml.library.reg" + +ApirLoadLibraryReturnCode apir_backend_initialize(uint32_t virgl_ctx_id, struct virgl_apir_callbacks *virgl_cbs) { + const char * dlsym_error; + + const char * apir_log_to_file = getenv(APIR_LLAMA_CPP_LOG_TO_FILE_ENV); + if (apir_log_to_file) { + apir_logfile = fopen(apir_log_to_file, "w"); + if (apir_logfile) { + ggml_log_set(log_to_file_callback, apir_logfile); + } else { + GGML_LOG_INFO("Could not open the log file at '%s'\n", apir_log_to_file); + } + } + + const char * library_name = virgl_cbs->get_config(virgl_ctx_id, APIR_GGML_LIBRARY_PATH_KEY); + const char * virgl_library_reg = virgl_cbs->get_config(virgl_ctx_id, APIR_GGML_LIBRARY_REG_KEY); + const char * library_reg = virgl_library_reg ? virgl_library_reg : GGML_DEFAULT_BACKEND_REG; + + if (!library_name) { + GGML_LOG_ERROR("cannot open the GGML library: env var '%s' not defined\n", APIR_LLAMA_CPP_GGML_LIBRARY_PATH_ENV); + + return APIR_LOAD_LIBRARY_ENV_VAR_MISSING; + } + + backend_library_handle = dlopen(library_name, RTLD_LAZY); + + if (!backend_library_handle) { + GGML_LOG_ERROR("cannot open the GGML library: %s\n", dlerror()); + + return APIR_LOAD_LIBRARY_CANNOT_OPEN; + } + + if (!library_reg) { + GGML_LOG_ERROR("cannot register the GGML library: env var '%s' not defined\n", APIR_LLAMA_CPP_GGML_LIBRARY_REG_ENV); + + return APIR_LOAD_LIBRARY_ENV_VAR_MISSING; + } + + void * ggml_backend_reg_fct = dlsym(backend_library_handle, library_reg); + dlsym_error = dlerror(); + if (dlsym_error) { + GGML_LOG_ERROR("cannot find the GGML backend registration symbol '%s' (from %s): %s\n", library_reg, + APIR_LLAMA_CPP_GGML_LIBRARY_REG_ENV, dlsym_error); + + return APIR_LOAD_LIBRARY_SYMBOL_MISSING; + } + + uint32_t ret = backend_dispatch_initialize(ggml_backend_reg_fct); + + return (ApirLoadLibraryReturnCode) (APIR_LOAD_LIBRARY_INIT_BASE_INDEX + ret); +} + +uint32_t apir_backend_dispatcher(uint32_t virgl_ctx_id, + virgl_apir_callbacks * virgl_cbs, + uint32_t cmd_type, + char * dec_cur, + const char * dec_end, + char * enc_cur, + const char * enc_end, + char ** enc_cur_after) { + apir_encoder enc = { + .cur = enc_cur, + .start = enc_cur, + .end = enc_end, + .fatal = false, + }; + + apir_decoder dec = { + .cur = dec_cur, + .end = dec_end, + .fatal = false, + }; + + virgl_apir_context ctx = { + .ctx_id = virgl_ctx_id, + .iface = virgl_cbs, + }; + + if (cmd_type >= APIR_BACKEND_DISPATCH_TABLE_COUNT) { + GGML_LOG_ERROR("Received an invalid dispatch index (%d >= %d)\n", cmd_type, APIR_BACKEND_DISPATCH_TABLE_COUNT); + return APIR_BACKEND_FORWARD_INDEX_INVALID; + } + + backend_dispatch_t forward_fct = apir_backend_dispatch_table[cmd_type]; + uint32_t ret = forward_fct(&enc, &dec, &ctx); + + *enc_cur_after = enc.cur; + + return ret; +} +} diff --git a/ggml/src/ggml-virtgpu/backend/shared/api_remoting.h b/ggml/src/ggml-virtgpu/backend/shared/api_remoting.h new file mode 100644 index 0000000000..f19a5d12d1 --- /dev/null +++ b/ggml/src/ggml-virtgpu/backend/shared/api_remoting.h @@ -0,0 +1,90 @@ +#pragma once + +/* the rest of this file must match virglrenderer/src/apir-protocol.h */ + +#include + +#include + +#define APIR_PROTOCOL_MAJOR 0 +#define APIR_PROTOCOL_MINOR 1 + +#define APIR_HANDSHAKE_MAGIC 0xab1e + +enum ApirCommandType { + APIR_COMMAND_TYPE_HANDSHAKE = 0, + APIR_COMMAND_TYPE_LOADLIBRARY = 1, + APIR_COMMAND_TYPE_FORWARD = 2, + + APIR_COMMAND_TYPE_LENGTH = 3, +}; + +typedef uint64_t ApirCommandFlags; + +enum ApirLoadLibraryReturnCode { + APIR_LOAD_LIBRARY_SUCCESS = 0, + APIR_LOAD_LIBRARY_HYPERCALL_INITIALIZATION_ERROR = 1, + APIR_LOAD_LIBRARY_ALREADY_LOADED = 2, + APIR_LOAD_LIBRARY_ENV_VAR_MISSING = 3, + APIR_LOAD_LIBRARY_CANNOT_OPEN = 4, + APIR_LOAD_LIBRARY_SYMBOL_MISSING = 5, + APIR_LOAD_LIBRARY_INIT_BASE_INDEX = 6, // anything above this is a APIR backend library initialization return code +}; + +enum ApirForwardReturnCode { + APIR_FORWARD_SUCCESS = 0, + APIR_FORWARD_NO_DISPATCH_FCT = 1, + APIR_FORWARD_TIMEOUT = 2, + + APIR_FORWARD_BASE_INDEX = 3, // anything above this is a APIR backend library forward return code +} ; + +__attribute__((unused)) static inline const char * apir_command_name(ApirCommandType type) { + switch (type) { + case APIR_COMMAND_TYPE_HANDSHAKE: + return "HandShake"; + case APIR_COMMAND_TYPE_LOADLIBRARY: + return "LoadLibrary"; + case APIR_COMMAND_TYPE_FORWARD: + return "Forward"; + default: + return "unknown"; + } +} + +__attribute__((unused)) static const char * apir_load_library_error(ApirLoadLibraryReturnCode code) { +#define APIR_LOAD_LIBRARY_ERROR(code_name) \ + do { \ + if (code == code_name) \ + return #code_name; \ + } while (0) + + APIR_LOAD_LIBRARY_ERROR(APIR_LOAD_LIBRARY_SUCCESS); + APIR_LOAD_LIBRARY_ERROR(APIR_LOAD_LIBRARY_HYPERCALL_INITIALIZATION_ERROR); + APIR_LOAD_LIBRARY_ERROR(APIR_LOAD_LIBRARY_ALREADY_LOADED); + APIR_LOAD_LIBRARY_ERROR(APIR_LOAD_LIBRARY_ENV_VAR_MISSING); + APIR_LOAD_LIBRARY_ERROR(APIR_LOAD_LIBRARY_CANNOT_OPEN); + APIR_LOAD_LIBRARY_ERROR(APIR_LOAD_LIBRARY_SYMBOL_MISSING); + APIR_LOAD_LIBRARY_ERROR(APIR_LOAD_LIBRARY_INIT_BASE_INDEX); + + return "Unknown APIR_COMMAND_TYPE_LoadLibrary error"; + +#undef APIR_LOAD_LIBRARY_ERROR +} + +__attribute__((unused)) static const char * apir_forward_error(ApirForwardReturnCode code) { +#define APIR_FORWARD_ERROR(code_name) \ + do { \ + if (code == code_name) \ + return #code_name; \ + } while (0) + + APIR_FORWARD_ERROR(APIR_FORWARD_SUCCESS); + APIR_FORWARD_ERROR(APIR_FORWARD_NO_DISPATCH_FCT); + APIR_FORWARD_ERROR(APIR_FORWARD_TIMEOUT); + APIR_FORWARD_ERROR(APIR_FORWARD_BASE_INDEX); + + return "Unknown APIR_COMMAND_TYPE_FORWARD error"; + +#undef APIR_FORWARD_ERROR +} diff --git a/ggml/src/ggml-virtgpu/backend/shared/apir_backend.gen.h b/ggml/src/ggml-virtgpu/backend/shared/apir_backend.gen.h new file mode 100644 index 0000000000..d214b6f2a9 --- /dev/null +++ b/ggml/src/ggml-virtgpu/backend/shared/apir_backend.gen.h @@ -0,0 +1,36 @@ +typedef enum ApirBackendCommandType { + + /* device */ + APIR_COMMAND_TYPE_DEVICE_GET_DEVICE_COUNT = 0, + APIR_COMMAND_TYPE_DEVICE_GET_COUNT = 1, + APIR_COMMAND_TYPE_DEVICE_GET_NAME = 2, + APIR_COMMAND_TYPE_DEVICE_GET_DESCRIPTION = 3, + APIR_COMMAND_TYPE_DEVICE_GET_TYPE = 4, + APIR_COMMAND_TYPE_DEVICE_GET_MEMORY = 5, + APIR_COMMAND_TYPE_DEVICE_SUPPORTS_OP = 6, + APIR_COMMAND_TYPE_DEVICE_GET_BUFFER_TYPE = 7, + APIR_COMMAND_TYPE_DEVICE_GET_PROPS = 8, + APIR_COMMAND_TYPE_DEVICE_BUFFER_FROM_PTR = 9, + + /* buffer-type */ + APIR_COMMAND_TYPE_BUFFER_TYPE_GET_NAME = 10, + APIR_COMMAND_TYPE_BUFFER_TYPE_GET_ALIGNMENT = 11, + APIR_COMMAND_TYPE_BUFFER_TYPE_GET_MAX_SIZE = 12, + APIR_COMMAND_TYPE_BUFFER_TYPE_IS_HOST = 13, + APIR_COMMAND_TYPE_BUFFER_TYPE_ALLOC_BUFFER = 14, + APIR_COMMAND_TYPE_BUFFER_TYPE_GET_ALLOC_SIZE = 15, + + /* buffer */ + APIR_COMMAND_TYPE_BUFFER_GET_BASE = 16, + APIR_COMMAND_TYPE_BUFFER_SET_TENSOR = 17, + APIR_COMMAND_TYPE_BUFFER_GET_TENSOR = 18, + APIR_COMMAND_TYPE_BUFFER_CPY_TENSOR = 19, + APIR_COMMAND_TYPE_BUFFER_CLEAR = 20, + APIR_COMMAND_TYPE_BUFFER_FREE_BUFFER = 21, + + /* backend */ + APIR_COMMAND_TYPE_BACKEND_GRAPH_COMPUTE = 22, + + // last command_type index + 1 + APIR_BACKEND_DISPATCH_TABLE_COUNT = 23, +} ApirBackendCommandType; diff --git a/ggml/src/ggml-virtgpu/backend/shared/apir_backend.h b/ggml/src/ggml-virtgpu/backend/shared/apir_backend.h new file mode 100644 index 0000000000..f3efa52c72 --- /dev/null +++ b/ggml/src/ggml-virtgpu/backend/shared/apir_backend.h @@ -0,0 +1,46 @@ +#pragma once + +#include "apir_backend.gen.h" + +#include // for uintptr_t +#include // for timespec, clock_gettime + +#define APIR_BACKEND_INITIALIZE_SUCCESS 0 +#define APIR_BACKEND_INITIALIZE_CANNOT_OPEN_BACKEND_LIBRARY 1 +#define APIR_BACKEND_INITIALIZE_CANNOT_OPEN_GGML_LIBRARY 2 +#define APIR_BACKEND_INITIALIZE_MISSING_BACKEND_SYMBOLS 3 +#define APIR_BACKEND_INITIALIZE_MISSING_GGML_SYMBOLS 4 +#define APIR_BACKEND_INITIALIZE_BACKEND_FAILED 5 +#define APIR_BACKEND_INITIALIZE_BACKEND_REG_FAILED 6 +#define APIR_BACKEND_INITIALIZE_ALREADY_INITED 7 +#define APIR_BACKEND_INITIALIZE_NO_DEVICE 8 + + +// new entries here need to be added to the apir_backend_initialize_error function below + +#define APIR_BACKEND_FORWARD_INDEX_INVALID 6 + +// 0 is fast, 1 avoids the backend to crash if an unsupported tensor is received +#define APIR_BACKEND_CHECK_SUPPORTS_OP 0 + +typedef uintptr_t apir_buffer_type_host_handle_t; +typedef uintptr_t apir_buffer_host_handle_t; + +static const char * apir_backend_initialize_error(int code) { +#define APIR_BACKEND_INITIALIZE_ERROR(code_name) \ + do { \ + if (code == code_name) \ + return #code_name; \ + } while (0) + + APIR_BACKEND_INITIALIZE_ERROR(APIR_BACKEND_INITIALIZE_SUCCESS); + APIR_BACKEND_INITIALIZE_ERROR(APIR_BACKEND_INITIALIZE_CANNOT_OPEN_BACKEND_LIBRARY); + APIR_BACKEND_INITIALIZE_ERROR(APIR_BACKEND_INITIALIZE_CANNOT_OPEN_GGML_LIBRARY); + APIR_BACKEND_INITIALIZE_ERROR(APIR_BACKEND_INITIALIZE_MISSING_BACKEND_SYMBOLS); + APIR_BACKEND_INITIALIZE_ERROR(APIR_BACKEND_INITIALIZE_MISSING_GGML_SYMBOLS); + APIR_BACKEND_INITIALIZE_ERROR(APIR_BACKEND_INITIALIZE_BACKEND_FAILED); + + return "Unknown APIR_BACKEND_INITIALIZE error:/"; + +#undef APIR_BACKEND_INITIALIZE_ERROR +} diff --git a/ggml/src/ggml-virtgpu/backend/shared/apir_cs.h b/ggml/src/ggml-virtgpu/backend/shared/apir_cs.h new file mode 100644 index 0000000000..27a61091ff --- /dev/null +++ b/ggml/src/ggml-virtgpu/backend/shared/apir_cs.h @@ -0,0 +1,383 @@ +#pragma once + +#include "ggml-impl.h" + +#include +#include + +#define likely(x) __builtin_expect(!!(x), 1) +#define unlikely(x) __builtin_expect(!!(x), 0) + +struct apir_encoder { + char * cur; + const char * start; + const char * end; + bool fatal; + +}; + +struct apir_decoder { + const char * cur; + const char * end; + bool fatal; +}; + +/* + * new encoder and decoder + */ + +static apir_decoder apir_new_decoder(const char * ptr, size_t size) { + apir_decoder dec = { + .cur = ptr, + .end = ptr + size, + .fatal = false, + }; + + return dec; +} + +static apir_encoder apir_new_encoder(char * ptr, size_t size) { + apir_encoder enc = { + .cur = ptr, + .start = ptr, + .end = ptr + size, + .fatal = false, + }; + + return enc; +} + +/* + * fatal flag handling + */ + +static inline void apir_encoder_reset_fatal(apir_encoder * enc) { + enc->fatal = false; +} + +static inline void apir_encoder_set_fatal(apir_encoder * enc) { + enc->fatal = true; +} + +static inline bool apir_encoder_get_fatal(const apir_encoder * enc) { + return enc->fatal; +} + +static inline void apir_decoder_reset_fatal(apir_decoder * dec) { + dec->fatal = false; +} + +static inline void apir_decoder_set_fatal(apir_decoder * dec) { + dec->fatal = true; +} + +static inline bool apir_decoder_get_fatal(const apir_decoder * dec) { + return dec->fatal; +} + +/* + * encode peek + */ + +static inline bool apir_decoder_peek_internal(apir_decoder * dec, + size_t size, + void * val, + size_t val_size) { + assert(val_size <= size); + + if (unlikely(size > (size_t) (dec->end - dec->cur))) { + GGML_LOG_ERROR("reading too much from the decoder ...\n"); + apir_decoder_set_fatal(dec); + memset(val, 0, val_size); + return false; + } + + /* we should not rely on the compiler to optimize away memcpy... */ + memcpy(val, dec->cur, val_size); + return true; +} + +static inline void apir_decoder_peek(apir_decoder * dec, size_t size, void * val, size_t val_size) { + apir_decoder_peek_internal(dec, size, val, val_size); +} + +static inline const void * apir_decoder_use_inplace(apir_decoder * dec, size_t size) { + if (unlikely(size > (size_t) (dec->end - dec->cur))) { + GGML_LOG_ERROR("reading too much from the decoder ...\n"); + apir_decoder_set_fatal(dec); + return NULL; + } + const void * addr = dec->cur; + dec->cur += size; + + return addr; +} + +/* + * read/write + */ + +static inline void apir_decoder_read(apir_decoder * dec, size_t size, void * val, size_t val_size) { + if (apir_decoder_peek_internal(dec, size, val, val_size)) { + dec->cur += size; + } +} + +static inline char * apir_encoder_write(apir_encoder * enc, size_t size, const void * val, size_t val_size) { + assert(val_size <= size); + assert(size <= ((size_t) (enc->end - enc->cur))); + + char * write_addr = enc->cur; + /* we should not rely on the compiler to optimize away memcpy... */ + memcpy(write_addr, val, val_size); + enc->cur += size; + + return write_addr; +} + +/* + * encode/decode + */ + +static inline void apir_decode(apir_decoder * dec, size_t size, void * data, size_t data_size) { + assert(size % 4 == 0); + apir_decoder_read(dec, size, data, data_size); +} + +static inline void apir_encode(apir_encoder * enc, size_t size, const void * data, size_t data_size) { + assert(size % 4 == 0); + apir_encoder_write(enc, size, data, data_size); +} + +/* + * typed encode/decode + */ + +/* uint8_t */ + +static inline void apir_encode_uint8_t(apir_encoder * enc, const uint8_t * val) { + apir_encode(enc, sizeof(int), val, sizeof(*val)); +} + +static inline void apir_decode_uint8_t(apir_decoder * dec, uint8_t * val) { + apir_decode(dec, sizeof(int), val, sizeof(*val)); +} + +/* uint64_t */ + +static inline void apir_encode_uint64_t(apir_encoder * enc, const uint64_t * val) { + apir_encode(enc, 8, val, sizeof(*val)); +} + +static inline void apir_decode_uint64_t(apir_decoder * dec, uint64_t * val) { + apir_decode(dec, 8, val, sizeof(*val)); +} + +static inline void apir_encode_uint64_t_array(apir_encoder * enc, const uint64_t * val, uint32_t count) { + const size_t size = sizeof(*val) * count; + assert(size >= count); + apir_encode(enc, size, val, size); +} + +static inline void apir_decode_uint64_t_array(apir_decoder * dec, uint64_t * val, uint32_t count) { + const size_t size = sizeof(*val) * count; + assert(size >= count); + apir_decode(dec, size, val, size); +} + +static inline const uint64_t * apir_decode_uint64_t_array_inplace(apir_decoder * dec, uint32_t count) { + return (uint64_t *) (uintptr_t) apir_decoder_use_inplace(dec, count * sizeof(uint64_t)); +} + +/* int32_t */ + +static inline void apir_encode_int32_t(apir_encoder * enc, const int32_t * val) { + apir_encode(enc, 4, val, sizeof(*val)); +} + +static inline void apir_decode_int32_t(apir_decoder * dec, int32_t * val) { + apir_decode(dec, 4, val, sizeof(*val)); +} + +static inline void apir_encode_int32_t_array(apir_encoder * enc, const int32_t * val, uint32_t count) { + const size_t size = sizeof(*val) * count; + assert(size >= count); + apir_encode(enc, size, val, size); +} + +static inline void apir_decode_int32_t_array(apir_decoder * dec, int32_t * val, uint32_t count) { + const size_t size = sizeof(*val) * count; + assert(size >= count); + apir_decode(dec, size, val, size); +} + +/* array size (uint64_t) */ + +static inline void apir_encode_array_size(apir_encoder * enc, uint64_t size) { + apir_encode_uint64_t(enc, &size); +} + +static inline uint64_t apir_decode_array_size(apir_decoder * dec, uint64_t expected_size) { + uint64_t size; + apir_decode_uint64_t(dec, &size); + if (size != expected_size) { + GGML_LOG_ERROR("Couldn't decode array from the decoder\n"); + apir_decoder_set_fatal(dec); + size = 0; + } + return size; +} + +static inline uint64_t apir_decode_array_size_unchecked(apir_decoder * dec) { + uint64_t size; + apir_decode_uint64_t(dec, &size); + return size; +} + +/* non-array pointer */ + +static inline bool apir_encode_simple_pointer(apir_encoder * enc, const void * val) { + apir_encode_array_size(enc, val ? 1 : 0); + return val; +} + +static inline bool apir_decode_simple_pointer(apir_decoder * dec) { + return apir_decode_array_size_unchecked(dec); +} + +/* uint32_t */ + +static inline void apir_encode_uint32_t(apir_encoder * enc, const uint32_t * val) { + apir_encode(enc, 4, val, sizeof(*val)); +} + +static inline void apir_decode_uint32_t(apir_decoder * dec, uint32_t * val) { + apir_decode(dec, 4, val, sizeof(*val)); +} + +static inline void apir_encode_uint32_t_array(apir_encoder * enc, const uint32_t * val, uint32_t count) { + const size_t size = sizeof(*val) * count; + assert(size >= count); + apir_encode(enc, size, val, size); +} + +static inline void apir_decode_uint32_t_array(apir_decoder * dec, uint32_t * val, uint32_t count) { + const size_t size = sizeof(*val) * count; + assert(size >= count); + apir_decode(dec, size, val, size); +} + +/* size_t */ + +static inline void apir_encode_size_t(apir_encoder * enc, const size_t * val) { + const uint64_t tmp = *val; + apir_encode_uint64_t(enc, &tmp); +} + +static inline void apir_decode_size_t(apir_decoder * dec, size_t * val) { + uint64_t tmp; + apir_decode_uint64_t(dec, &tmp); + *val = tmp; +} + +static inline void apir_encode_size_t_array(apir_encoder * enc, const size_t * val, uint32_t count) { + if (sizeof(size_t) == sizeof(uint64_t)) { + apir_encode_uint64_t_array(enc, (const uint64_t *) val, count); + } else { + for (uint32_t i = 0; i < count; i++) { + apir_encode_size_t(enc, &val[i]); + } + } +} + +static inline void apir_decode_size_t_array(apir_decoder * dec, size_t * val, uint32_t count) { + if (sizeof(size_t) == sizeof(uint64_t)) { + apir_decode_uint64_t_array(dec, (uint64_t *) val, count); + } else { + for (uint32_t i = 0; i < count; i++) { + apir_decode_size_t(dec, &val[i]); + } + } +} + +/* opaque blob */ + +static inline void apir_encode_blob_array(apir_encoder * enc, const void * val, size_t size) { + apir_encode(enc, (size + 3) & ~3, val, size); +} + +static inline void apir_decode_blob_array(apir_decoder * dec, void * val, size_t size) { + apir_decode(dec, (size + 3) & ~3, val, size); +} + +/* string */ + +static inline void apir_encode_char_array(apir_encoder * enc, const char * val, size_t size) { + assert(size && strlen(val) < size); + apir_encode_blob_array(enc, val, size); +} + +static inline void apir_decode_char_array(apir_decoder * dec, char * val, size_t size) { + apir_decode_blob_array(dec, val, size); + if (size) { + val[size - 1] = '\0'; + } else { + GGML_LOG_ERROR("Couldn't decode the blog array\n"); + apir_decoder_set_fatal(dec); + } +} + +/* (temp) buffer allocation */ + +static inline void * apir_decoder_alloc_array(size_t size, size_t count) { + size_t alloc_size; + if (unlikely(__builtin_mul_overflow(size, count, &alloc_size))) { + GGML_LOG_ERROR("overflow in array allocation of %zu * %zu bytes\n", size, count); + return NULL; + } + + return malloc(alloc_size); +} + +/* bool */ + +static inline void apir_encode_bool_t(apir_encoder * enc, const bool * val) { + apir_encode(enc, sizeof(int), val, sizeof(bool)); +} + +static inline void apir_decode_bool_t(apir_decoder * dec, bool * val) { + apir_decode(dec, sizeof(int), val, sizeof(bool)); +} + +/* apir_buffer_type_host_handle_t */ + +static inline void apir_encode_apir_buffer_type_host_handle_t(apir_encoder * enc, + const apir_buffer_type_host_handle_t * val) { + apir_encode(enc, sizeof(apir_buffer_type_host_handle_t), val, sizeof(apir_buffer_type_host_handle_t)); +} + +static inline void apir_decode_apir_buffer_type_host_handle_t(apir_decoder * dec, + apir_buffer_type_host_handle_t * val) { + apir_decode(dec, sizeof(apir_buffer_type_host_handle_t), val, sizeof(apir_buffer_type_host_handle_t)); +} + +/* apir_buffer_host_handle_t */ + +static inline void apir_encode_apir_buffer_host_handle_t(apir_encoder * enc, + const apir_buffer_host_handle_t * val) { + apir_encode(enc, sizeof(apir_buffer_host_handle_t), val, sizeof(apir_buffer_host_handle_t)); +} + +static inline void apir_decode_apir_buffer_host_handle_t(apir_decoder * dec, apir_buffer_host_handle_t * val) { + apir_decode(dec, sizeof(apir_buffer_host_handle_t), val, sizeof(apir_buffer_host_handle_t)); +} + +/* uintptr_t */ + +static inline void apir_encode_uintptr_t(apir_encoder * enc, const uintptr_t * val) { + apir_encode(enc, sizeof(*val), val, sizeof(*val)); +} + +static inline void apir_decode_uintptr_t(apir_decoder * dec, uintptr_t * val) { + apir_decode(dec, sizeof(*val), val, sizeof(*val)); +} diff --git a/ggml/src/ggml-virtgpu/backend/shared/apir_cs_ggml.h b/ggml/src/ggml-virtgpu/backend/shared/apir_cs_ggml.h new file mode 100644 index 0000000000..070c3b25fb --- /dev/null +++ b/ggml/src/ggml-virtgpu/backend/shared/apir_cs_ggml.h @@ -0,0 +1,211 @@ +#include "ggml-impl.h" +#include "apir_cs.h" +#include "apir_cs_rpc.h" + +// ggml_buffer_to_apir_host_handle(ggml_backend_buffer_t buffer); + +static inline void apir_encode_ggml_buffer_host_handle(apir_encoder * enc, + const apir_buffer_host_handle_t * handle); + +static inline ggml_backend_buffer_t apir_decode_ggml_buffer(apir_decoder * dec); + +/* apir_rpc_tensor */ + +static inline void apir_encode_rcp_tensor(apir_encoder * enc, const apir_rpc_tensor * apir_rpc_tensor) { + size_t apir_rpc_tensor_size = sizeof(*apir_rpc_tensor); + apir_encode(enc, apir_rpc_tensor_size, apir_rpc_tensor, apir_rpc_tensor_size); +} + +static inline apir_rpc_tensor * apir_decode_apir_rpc_tensor_inplace(apir_decoder * dec) { + size_t apir_rpc_tensor_size = sizeof(apir_rpc_tensor); + + return (apir_rpc_tensor *) (uintptr_t) apir_decoder_use_inplace(dec, apir_rpc_tensor_size); +} + +static inline apir_rpc_tensor * apir_decode_apir_rpc_tensor_array_inplace(apir_decoder * dec, + uint32_t n_tensors) { + size_t apir_rpc_tensor_size = sizeof(apir_rpc_tensor) * n_tensors; + + return (apir_rpc_tensor *) (uintptr_t) apir_decoder_use_inplace(dec, apir_rpc_tensor_size); +} + +/* ggml_tensor */ + +static inline void apir_encode_ggml_tensor(apir_encoder * enc, const ggml_tensor * tensor) { + apir_rpc_tensor serialized = apir_serialize_tensor(tensor); + + apir_encode_rcp_tensor(enc, &serialized); +} + +static inline const ggml_tensor * apir_decode_ggml_tensor(apir_decoder * dec) { + const apir_rpc_tensor * apir_rpc_tensor = apir_decode_apir_rpc_tensor_inplace(dec); + ggml_init_params params{ + /*.mem_size =*/ ggml_tensor_overhead(), + /*.mem_buffer =*/ NULL, + /*.no_alloc =*/ true, + }; + ggml_context * ctx = ggml_init(params); + + const ggml_tensor * tensor = apir_deserialize_tensor(ctx, apir_rpc_tensor); + + return tensor; +} + +/* *** ggml_backend_buffer_type_t *** */ + +// ggml_backend_buffer_type_t is a POINTER (to a struct). +// Only the host pointer is shared between the host and guest. +// The guest stores it in `buft->context`. +// The host simply writes the pointer address in the buffer variable. + +static inline void apir_encode_ggml_buffer_type(apir_encoder * enc, ggml_backend_buffer_type_t buft) { + apir_buffer_type_host_handle_t handle = ggml_buffer_type_to_apir_handle(buft); + apir_encoder_write(enc, sizeof(handle), &handle, sizeof(handle)); +} + +static inline ggml_backend_buffer_type_t apir_decode_ggml_buffer_type(apir_decoder * dec) { + apir_buffer_type_host_handle_t handle; + + apir_decoder_read(dec, sizeof(handle), &handle, sizeof(handle)); + + return (ggml_backend_buffer_type_t) handle; +} + +static inline apir_buffer_type_host_handle_t apir_decode_apir_buffer_type_host_handle(apir_decoder * dec) { + apir_buffer_type_host_handle_t handle; + + apir_decoder_read(dec, sizeof(handle), &handle, sizeof(handle)); + + return handle; +} + +/* *** ggml_backend_type_t *** */ + +// ggml_backend_buffer_t is a POINTER. +// same logic as for ggml_backend_buffer_type_t + +static inline void apir_encode_ggml_buffer(apir_encoder * enc, const ggml_backend_buffer_t buffer) { + apir_buffer_host_handle_t handle = BUFFER_TO_HOST_HANDLE(buffer); + apir_encoder_write(enc, sizeof(handle), &handle, sizeof(handle)); +} + +static inline ggml_backend_buffer_t apir_decode_ggml_buffer(apir_decoder * dec) { + ggml_backend_buffer_t buffer; + size_t buffer_ptr_size = sizeof(buffer); + + apir_decoder_read(dec, buffer_ptr_size, &buffer, buffer_ptr_size); + + return buffer; +} + +/* enum ggml_status */ + +static inline void apir_encode_ggml_status(apir_encoder * enc, const ggml_status * status) { + apir_encoder_write(enc, sizeof(*status), status, sizeof(*status)); +} + +static inline void apir_decode_ggml_status(apir_decoder * dec, ggml_status * status) { + apir_decoder_read(dec, sizeof(*status), status, sizeof(*status)); +} + +/* virtgpu_shmem */ + +static inline void apir_encode_virtgpu_shmem_res_id(apir_encoder * enc, uint32_t shmem_res_id) { + apir_encode_uint32_t(enc, &shmem_res_id); +} + +static inline void apir_decode_virtgpu_shmem_res_id(apir_decoder * dec, uint32_t * shmem_res_id) { + apir_decode_uint32_t(dec, shmem_res_id); +} + +/* ggml_cgraph */ + +static inline size_t apir_serialize_ggml_cgraph(ggml_cgraph * cgraph, std::vector & cgraph_data) { + apir_serialize_graph(cgraph, cgraph_data); + + return cgraph_data.size(); +} + +static inline void apir_encode_cgraph_data(apir_encoder * enc, std::vector & cgraph_data) { + size_t cgraph_size = cgraph_data.size(); + + apir_encode(enc, cgraph_size, cgraph_data.data(), cgraph_size); +} + +static inline ggml_cgraph * apir_decode_ggml_cgraph(apir_decoder * dec, size_t cgraph_size) { + GGML_UNUSED(cgraph_size); + + uint32_t n_nodes; + apir_decode_uint32_t(dec, &n_nodes); + const uint64_t * nodes = apir_decode_uint64_t_array_inplace(dec, n_nodes); + + uint32_t n_tensors; + apir_decode_uint32_t(dec, &n_tensors); + const apir_rpc_tensor * tensors = apir_decode_apir_rpc_tensor_array_inplace(dec, n_tensors); + + return apir_deserialize_graph(n_nodes, n_tensors, tensors, nodes); +} + +static inline void apir_encode_ggml_buffer_handle(apir_encoder * enc, const apir_buffer_host_handle_t * handle) { + apir_encoder_write(enc, sizeof(*handle), &handle, sizeof(*handle)); +} + +static inline void apir_encode_ggml_tensor_inline(apir_encoder * enc, const ggml_tensor * tensor) { + size_t tensor_size = sizeof(*tensor); + + if (tensor->extra) { + GGML_ABORT("Cannot pass tensors with extra"); + } + + if (tensor->src[0] && tensor->buffer) { + static int first = 1; + if (first) { + GGML_LOG_WARN("Cannot pass tensors with src and buffer\n"); + first = 0; + } + } + + apir_encoder_write(enc, tensor_size, tensor, tensor_size); + + // tensor->data is a pointer inside the device buffer. No need to touch it + // tensor->buffer is a pointer to a buffer. Encoding the buffer handle in sequence. + // (could also make a copy of the tensor, and update locally.) + + if (tensor->buffer) { + apir_buffer_host_handle_t buffer_handle = ggml_buffer_to_apir_handle(tensor->buffer); + apir_encode_ggml_buffer_handle(enc, &buffer_handle); + } + + if (tensor->view_src) { + apir_encoder_write(enc, tensor_size, tensor->view_src, tensor_size); + } + + for (int i = 0; tensor->src[i]; i++) { + const ggml_tensor * tensor_src = tensor->src[i]; + apir_encoder_write(enc, tensor_size, tensor_src, tensor_size); + } +} + +static inline const ggml_tensor * apir_decode_ggml_tensor_inplace(apir_decoder * dec) { + // it safe to remove the `const` qualifier here, we *do* want to + // modify the shared memory data to fix the `src` pointers. + ggml_tensor * tensor = (ggml_tensor *) (uintptr_t) apir_decoder_use_inplace(dec, sizeof(ggml_tensor)); + + // tensor->data is a pointer inside the device buffer. No need to touch it + // tensor->buffer is a pointer to a buffer. Decode the buffer handle encoded in sequence. + if (tensor->buffer) { + tensor->buffer = apir_decode_ggml_buffer(dec); + } + + if (tensor->view_src) { + ggml_tensor * tensor_view_src = (ggml_tensor *) (uintptr_t) apir_decoder_use_inplace(dec, sizeof(ggml_tensor)); + tensor->view_src = tensor_view_src; + } + + for (int i = 0; tensor->src[i]; i++) { + ggml_tensor * tensor_src = (ggml_tensor *) (uintptr_t) apir_decoder_use_inplace(dec, sizeof(ggml_tensor)); + tensor->src[i] = tensor_src; // overwrite op->src[i] pointer with the actual location of the src tensor + } + + return tensor; +} diff --git a/ggml/src/ggml-virtgpu/backend/shared/apir_cs_rpc.h b/ggml/src/ggml-virtgpu/backend/shared/apir_cs_rpc.h new file mode 100644 index 0000000000..f681798952 --- /dev/null +++ b/ggml/src/ggml-virtgpu/backend/shared/apir_cs_rpc.h @@ -0,0 +1,54 @@ +#include "ggml.h" +#include "ggml-backend-impl.h" + +#include +#include +#include +#include + +// ggml_tensor is serialized into apir_rpc_tensor +struct apir_rpc_tensor { + uint64_t id; + uint32_t type; + uint64_t buffer; + uint32_t ne[GGML_MAX_DIMS]; + uint32_t nb[GGML_MAX_DIMS]; + uint32_t op; + int32_t op_params[GGML_MAX_OP_PARAMS / sizeof(int32_t)]; + int32_t flags; + uint64_t src[GGML_MAX_SRC]; + uint64_t view_src; + uint64_t view_offs; + uint64_t data; + char name[GGML_MAX_NAME]; + + char padding[4]; +}; + +/* frontend */ + +apir_rpc_tensor apir_serialize_tensor(const ggml_tensor * tensor); + +void apir_serialize_graph(const ggml_cgraph * cgraph, std::vector & output); + +/* backend */ + +void apir_track_backend_buffer(ggml_backend_buffer_t buffer); +bool apir_untrack_backend_buffer(ggml_backend_buffer_t buffer); +std::unordered_set apir_get_track_backend_buffers(); + +void apir_add_tensor(ggml_tensor * tensor, + std::vector & tensors, + std::unordered_set & visited); + +ggml_tensor * apir_deserialize_tensor(ggml_context * ctx, const apir_rpc_tensor * tensor); + +ggml_tensor * apir_create_node(uint64_t id, + ggml_context * ctx, + const std::unordered_map & tensor_ptrs, + std::unordered_map & tensor_map); + +ggml_cgraph * apir_deserialize_graph(uint32_t n_nodes, + uint32_t n_tensors, + const apir_rpc_tensor * tensors, + const uint64_t * nodes); diff --git a/ggml/src/ggml-virtgpu/ggml-backend-buffer-type.cpp b/ggml/src/ggml-virtgpu/ggml-backend-buffer-type.cpp new file mode 100644 index 0000000000..7f650659b8 --- /dev/null +++ b/ggml/src/ggml-virtgpu/ggml-backend-buffer-type.cpp @@ -0,0 +1,98 @@ +#include "ggml-remoting.h" + +static ggml_backend_buffer_t ggml_backend_remoting_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, + size_t size) { + virtgpu * gpu = BUFT_TO_GPU(buft); + + ggml_backend_remoting_buffer_context * context = (ggml_backend_remoting_buffer_context *) malloc(sizeof(*context)); + if (!context) { + GGML_ABORT("Couldn't allocate the buffer context ..."); + } + + context->gpu = gpu; + + bool async__unused, host_buffer__unused, events__unused; + bool buffer_from_host_ptr; + apir_device_get_props(gpu, &async__unused, &host_buffer__unused, &buffer_from_host_ptr, &events__unused); + + if (buffer_from_host_ptr) { + context->apir_context = apir_device_buffer_from_ptr(gpu, size, size); + context->base = context->apir_context.shmem.mmap_ptr; + context->is_from_ptr = true; + } else { + context->apir_context = apir_buffer_type_alloc_buffer(gpu, buft, size); + context->is_from_ptr = false; + context->base = NULL; + } + + ggml_backend_buffer_t buffer = + ggml_backend_buffer_init(buft, ggml_backend_remoting_buffer_interface, (void *) context, size); + + return buffer; +} + +static const char * ggml_backend_remoting_buffer_type_get_name(ggml_backend_buffer_type_t buft) { + virtgpu * gpu = BUFT_TO_GPU(buft); + + return apir_buffer_type_get_name(gpu, buft); +} + +static size_t ggml_backend_remoting_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) { + virtgpu * gpu = BUFT_TO_GPU(buft); + + static size_t align = 0; + + if (align == 0) { + align = apir_buffer_type_get_alignment(gpu, buft); + } + + return align; +} + +static size_t ggml_backend_remoting_buffer_type_get_max_size(ggml_backend_buffer_type_t buft) { + virtgpu * gpu = BUFT_TO_GPU(buft); + + static size_t max_size = 0; + if (max_size == 0) { + max_size = apir_buffer_type_get_max_size(gpu, buft); + } + + return max_size; +} + +static bool ggml_backend_remoting_buffer_type_is_host(ggml_backend_buffer_type_t buft) { + virtgpu * gpu = BUFT_TO_GPU(buft); + + return apir_buffer_type_is_host(gpu, buft); +} + +static size_t ggml_backend_remoting_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, + const ggml_tensor * tensor) { + virtgpu * gpu = BUFT_TO_GPU(buft); + + if (tensor->buffer == NULL + || !tensor->buffer->context + || !buft->device->iface.supports_buft(buft->device, tensor->buffer->buft)) { + return ggml_nbytes(tensor); + } + + return apir_buffer_type_get_alloc_size(gpu, buft, tensor); +} + +const ggml_backend_buffer_type_i ggml_backend_remoting_buffer_type_interface = { + /* .get_name = */ ggml_backend_remoting_buffer_type_get_name, + /* .alloc_buffer = */ ggml_backend_remoting_buffer_type_alloc_buffer, + /* .get_alignment = */ ggml_backend_remoting_buffer_type_get_alignment, + /* .get_max_size = */ ggml_backend_remoting_buffer_type_get_max_size, + /* .get_alloc_size = */ ggml_backend_remoting_buffer_type_get_alloc_size, + /* .is_host = */ NULL, +}; + +const ggml_backend_buffer_type_i ggml_backend_remoting_buffer_from_ptr_type_interface = { + /* .get_name = */ ggml_backend_remoting_buffer_type_get_name, + /* .alloc_buffer = */ NULL, + /* .get_alignment = */ ggml_backend_remoting_buffer_type_get_alignment, + /* .get_max_size = */ ggml_backend_remoting_buffer_type_get_max_size, + /* .get_alloc_size = */ ggml_backend_remoting_buffer_type_get_alloc_size, + /* .is_host = */ NULL, +}; diff --git a/ggml/src/ggml-virtgpu/ggml-backend-buffer.cpp b/ggml/src/ggml-virtgpu/ggml-backend-buffer.cpp new file mode 100644 index 0000000000..6b95362dd8 --- /dev/null +++ b/ggml/src/ggml-virtgpu/ggml-backend-buffer.cpp @@ -0,0 +1,119 @@ +#include "ggml-remoting.h" + +#define BUFFER_TO_GPU(name) ((ggml_backend_remoting_buffer_context *) (name)->context)->gpu + +static void * ggml_backend_remoting_buffer_get_base(ggml_backend_buffer_t buffer) { + ggml_backend_remoting_buffer_context * context = (ggml_backend_remoting_buffer_context *) buffer->context; + if (context->base) { + return context->base; + } + + context->base = apir_buffer_get_base(BUFFER_TO_GPU(buffer), BUFFER_TO_APIR_CONTEXT(buffer)); + + return context->base; +} + +static void ggml_backend_remoting_buffer_set_tensor(ggml_backend_buffer_t buffer, + ggml_tensor * tensor, + const void * data, + size_t offset, + size_t size) { + virtgpu * gpu = BUFFER_TO_GPU(buffer); + + ggml_backend_remoting_buffer_context * context = BUFFER_TO_GGML_CONTEXT(buffer); + if (context->is_from_ptr) { + memcpy((char *) tensor->data + offset, data, size); + } else { + apir_buffer_set_tensor(gpu, BUFFER_TO_APIR_CONTEXT(buffer), tensor, data, offset, size); + } + + return; +} + +static void ggml_backend_remoting_buffer_get_tensor(ggml_backend_buffer_t buffer, + const ggml_tensor * tensor, + void * data, + size_t offset, + size_t size) { + virtgpu * gpu = BUFFER_TO_GPU(buffer); + ggml_backend_remoting_buffer_context * context = BUFFER_TO_GGML_CONTEXT(buffer); + if (context->is_from_ptr) { + memcpy(data, (const char *) tensor->data + offset, size); + } else { + apir_buffer_get_tensor(gpu, BUFFER_TO_APIR_CONTEXT(buffer), tensor, data, offset, size); + } +} + +static void ggml_backend_remoting_buffer_set_tensor_from_ptr(ggml_backend_buffer_t buffer, + ggml_tensor * tensor, + const void * data, + size_t offset, + size_t size) { + UNUSED(buffer); + + memcpy((char *) tensor->data + offset, data, size); + + return; +} + +static void ggml_backend_remoting_buffer_get_tensor_from_ptr(ggml_backend_buffer_t buffer, + const ggml_tensor * tensor, + void * data, + size_t offset, + size_t size) { + UNUSED(buffer); + + memcpy(data, (const char *) tensor->data + offset, size); +} + +static bool ggml_backend_remoting_buffer_cpy_tensor(ggml_backend_buffer_t buffer, + const ggml_tensor * src, + ggml_tensor * dst) { + virtgpu * gpu = BUFFER_TO_GPU(buffer); + + bool ret = apir_buffer_cpy_tensor(gpu, BUFFER_TO_APIR_CONTEXT(buffer), src, dst); + + return ret; +} + +static void ggml_backend_remoting_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) { + virtgpu * gpu = BUFFER_TO_GPU(buffer); + + apir_buffer_clear(gpu, BUFFER_TO_APIR_CONTEXT(buffer), value); + + return; +} + +static void ggml_backend_remoting_buffer_free_buffer(ggml_backend_buffer_t buffer) { + virtgpu * gpu = BUFFER_TO_GPU(buffer); + + apir_buffer_free_buffer(gpu, BUFFER_TO_APIR_CONTEXT(buffer)); + + ggml_backend_remoting_buffer_context * context = BUFFER_TO_GGML_CONTEXT(buffer); + free(context); + buffer->context = NULL; +} + +const ggml_backend_buffer_i ggml_backend_remoting_buffer_interface = { + /* .free_buffer = */ ggml_backend_remoting_buffer_free_buffer, + /* .get_base = */ ggml_backend_remoting_buffer_get_base, + /* .init_tensor = */ NULL, + /* .memset_tensor = */ NULL, + /* .set_tensor = */ ggml_backend_remoting_buffer_set_tensor, + /* .get_tensor = */ ggml_backend_remoting_buffer_get_tensor, + /* .cpy_tensor = */ ggml_backend_remoting_buffer_cpy_tensor, + /* .clear = */ ggml_backend_remoting_buffer_clear, + /* .reset = */ NULL, +}; + +const ggml_backend_buffer_i ggml_backend_remoting_buffer_from_ptr_interface = { + /* .free_buffer = */ ggml_backend_remoting_buffer_free_buffer, + /* .get_base = */ ggml_backend_remoting_buffer_get_base, + /* .init_tensor = */ NULL, + /* .memset_tensor = */ NULL, + /* .set_tensor = */ ggml_backend_remoting_buffer_set_tensor_from_ptr, + /* .get_tensor = */ ggml_backend_remoting_buffer_get_tensor_from_ptr, + /* .cpy_tensor = */ ggml_backend_remoting_buffer_cpy_tensor, + /* .clear = */ ggml_backend_remoting_buffer_clear, + /* .reset = */ NULL, +}; diff --git a/ggml/src/ggml-virtgpu/ggml-backend-device.cpp b/ggml/src/ggml-virtgpu/ggml-backend-device.cpp new file mode 100644 index 0000000000..579eb99078 --- /dev/null +++ b/ggml/src/ggml-virtgpu/ggml-backend-device.cpp @@ -0,0 +1,144 @@ +#include "ggml-remoting.h" + +static const char * ggml_backend_remoting_device_get_name(ggml_backend_dev_t dev) { + virtgpu * gpu = DEV_TO_GPU(dev); + + return apir_device_get_name(gpu); +} + +static const char * ggml_backend_remoting_device_get_description(ggml_backend_dev_t dev) { + virtgpu * gpu = DEV_TO_GPU(dev); + + return apir_device_get_description(gpu); +} + +static enum ggml_backend_dev_type ggml_backend_remoting_device_get_type(ggml_backend_dev_t dev) { + virtgpu * gpu = DEV_TO_GPU(dev); + + static enum ggml_backend_dev_type type; + static bool has_type = false; + if (!has_type) { + has_type = true; + type = (enum ggml_backend_dev_type) apir_device_get_type(gpu); + } + + return type; +} + +static void ggml_backend_remoting_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) { + virtgpu * gpu = DEV_TO_GPU(dev); + + return apir_device_get_memory(gpu, free, total); +} + +static bool ggml_backend_remoting_device_supports_op(ggml_backend_dev_t dev, const ggml_tensor * op) { +#if USE_ALWAYS_TRUE_SUPPORTS_OP == 1 + /* ggml-rpc cheats it like this */ + /* with the current implementation of serialize_tensor, the src/view aren't properly passed */ + UNUSED(dev); + UNUSED(op); + + return true; +#else + virtgpu * gpu = DEV_TO_GPU(dev); + + return apir_device_supports_op(gpu, op); +#endif +} + +static bool ggml_backend_remoting_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) { + bool supported = buft->device == dev; + + return supported; +} + +static bool ggml_backend_remoting_device_offload_op(ggml_backend_dev_t dev, const ggml_tensor * op) { + UNUSED(dev); + UNUSED(op); + + return false; +} + +static void ggml_backend_remoting_device_get_props(ggml_backend_dev_t dev, ggml_backend_dev_props * props) { + props->name = ggml_backend_remoting_device_get_name(dev); + props->description = ggml_backend_remoting_device_get_description(dev); + props->type = ggml_backend_remoting_device_get_type(dev); + ggml_backend_remoting_device_get_memory(dev, &props->memory_free, &props->memory_total); + + virtgpu * gpu = DEV_TO_GPU(dev); + apir_device_get_props(gpu, &props->caps.async, &props->caps.host_buffer, &props->caps.buffer_from_host_ptr, + &props->caps.events); + + props->caps.buffer_from_host_ptr = false; + props->caps.async = false; + props->caps.events = false; +} + +ggml_backend_buffer_type_t ggml_backend_remoting_device_get_buffer_type(ggml_backend_dev_t dev) { + virtgpu * gpu = DEV_TO_GPU(dev); + + apir_buffer_type_host_handle_t ctx = apir_device_get_buffer_type(gpu); + + static ggml_backend_buffer_type buft{ + /* .iface = */ ggml_backend_remoting_buffer_type_interface, + /* .device = */ dev, + /* .context = */ (void *) ctx, + }; + + return &buft; +} + +static ggml_backend_buffer_type_t ggml_backend_remoting_device_get_buffer_from_ptr_type(ggml_backend_dev_t dev) { + virtgpu * gpu = DEV_TO_GPU(dev); + + apir_buffer_type_host_handle_t ctx = apir_device_get_buffer_type(gpu); + + static ggml_backend_buffer_type buft{ + /* .iface = */ ggml_backend_remoting_buffer_from_ptr_type_interface, + /* .device = */ dev, + /* .context = */ (void *) ctx, + }; + + return &buft; +} + +static ggml_backend_buffer_t ggml_backend_remoting_device_buffer_from_ptr(ggml_backend_dev_t dev, + void * ptr, + size_t size, + size_t max_tensor_size) { + virtgpu * gpu = DEV_TO_GPU(dev); + + ggml_backend_remoting_buffer_context * context = (ggml_backend_remoting_buffer_context *) malloc(sizeof(*context)); + if (!context) { + GGML_ABORT("Couldn't allocate the buffer context ..."); + } + + context->gpu = gpu; + context->apir_context = apir_device_buffer_from_ptr(gpu, size, max_tensor_size); + context->base = ptr; + context->is_from_ptr = true; + + ggml_backend_buffer_t buffer = + ggml_backend_buffer_init(ggml_backend_remoting_device_get_buffer_from_ptr_type(dev), + ggml_backend_remoting_buffer_from_ptr_interface, (void *) context, size); + + return buffer; +} + +const ggml_backend_device_i ggml_backend_remoting_device_interface = { + /* .get_name = */ ggml_backend_remoting_device_get_name, + /* .get_description = */ ggml_backend_remoting_device_get_description, + /* .get_memory = */ ggml_backend_remoting_device_get_memory, + /* .get_type = */ ggml_backend_remoting_device_get_type, + /* .get_props = */ ggml_backend_remoting_device_get_props, + /* .init_backend = */ ggml_backend_remoting_device_init, + /* .get_buffer_type = */ ggml_backend_remoting_device_get_buffer_type, + /* .get_host_buffer_type = */ NULL, + /* .buffer_from_host_ptr = */ ggml_backend_remoting_device_buffer_from_ptr, + /* .supports_op = */ ggml_backend_remoting_device_supports_op, + /* .supports_buft = */ ggml_backend_remoting_device_supports_buft, + /* .offload_op = */ ggml_backend_remoting_device_offload_op, + /* .event_new = */ NULL, + /* .event_free = */ NULL, + /* .event_synchronize = */ NULL, +}; diff --git a/ggml/src/ggml-virtgpu/ggml-backend-reg.cpp b/ggml/src/ggml-virtgpu/ggml-backend-reg.cpp new file mode 100644 index 0000000000..c46cf51c02 --- /dev/null +++ b/ggml/src/ggml-virtgpu/ggml-backend-reg.cpp @@ -0,0 +1,137 @@ +#include "ggml-remoting.h" +#include "ggml-virtgpu.h" + +#include +#include + +static virtgpu * apir_initialize() { + static virtgpu * apir_gpu_instance = NULL; + static bool apir_initialized = false; + + { + static std::mutex mutex; + std::lock_guard lock(mutex); + + if (apir_initialized) { + return apir_gpu_instance; + } + + apir_gpu_instance = create_virtgpu(); + if (!apir_gpu_instance) { + GGML_ABORT("failed to initialize the virtgpu"); + } + + apir_initialized = true; + } + + return apir_gpu_instance; +} + +static int ggml_backend_remoting_get_device_count() { + virtgpu * gpu = apir_initialize(); + if (!gpu) { + GGML_LOG_WARN("apir_initialize failed\n"); + return 0; + } + + return apir_device_get_count(gpu); +} + +static size_t ggml_backend_remoting_reg_get_device_count(ggml_backend_reg_t reg) { + UNUSED(reg); + + return ggml_backend_remoting_get_device_count(); +} + +static std::vector devices; + +ggml_backend_dev_t ggml_backend_remoting_get_device(size_t device) { + GGML_ASSERT(device < devices.size()); + return devices[device]; +} + +static void ggml_backend_remoting_reg_init_devices(ggml_backend_reg_t reg) { + if (devices.size() > 0) { + GGML_LOG_INFO("%s: already initialized\n", __func__); + return; + } + + virtgpu * gpu = apir_initialize(); + if (!gpu) { + GGML_LOG_ERROR("apir_initialize failed\n"); + return; + } + + static bool initialized = false; + + { + static std::mutex mutex; + std::lock_guard lock(mutex); + if (!initialized) { + for (int i = 0; i < ggml_backend_remoting_get_device_count(); i++) { + ggml_backend_remoting_device_context * ctx = new ggml_backend_remoting_device_context; + char desc[256] = "API Remoting device"; + + ctx->device = i; + ctx->name = GGML_REMOTING_FRONTEND_NAME + std::to_string(i); + ctx->description = desc; + ctx->gpu = gpu; + + ggml_backend_dev_t dev = new ggml_backend_device{ + /* .iface = */ ggml_backend_remoting_device_interface, + /* .reg = */ reg, + /* .context = */ ctx, + }; + devices.push_back(dev); + } + initialized = true; + } + } +} + +static ggml_backend_dev_t ggml_backend_remoting_reg_get_device(ggml_backend_reg_t reg, size_t device) { + UNUSED(reg); + + return ggml_backend_remoting_get_device(device); +} + +static const char * ggml_backend_remoting_reg_get_name(ggml_backend_reg_t reg) { + UNUSED(reg); + + return GGML_REMOTING_FRONTEND_NAME; +} + +static const ggml_backend_reg_i ggml_backend_remoting_reg_i = { + /* .get_name = */ ggml_backend_remoting_reg_get_name, + /* .get_device_count = */ ggml_backend_remoting_reg_get_device_count, + /* .get_device = */ ggml_backend_remoting_reg_get_device, + /* .get_proc_address = */ NULL, +}; + +ggml_backend_reg_t ggml_backend_virtgpu_reg() { + virtgpu * gpu = apir_initialize(); + if (!gpu) { + GGML_LOG_ERROR("virtgpu_apir_initialize failed\n"); + return NULL; + } + + static ggml_backend_reg reg = { + /* .api_version = */ GGML_BACKEND_API_VERSION, + /* .iface = */ ggml_backend_remoting_reg_i, + /* .context = */ gpu, + }; + + static bool initialized = false; + if (initialized) { + return ® + } + initialized = true; + + ggml_backend_remoting_reg_init_devices(®); + + GGML_LOG_INFO("%s: initialized\n", __func__); + + return ® +} + +GGML_BACKEND_DL_IMPL(ggml_backend_virtgpu_reg) diff --git a/ggml/src/ggml-virtgpu/ggml-backend.cpp b/ggml/src/ggml-virtgpu/ggml-backend.cpp new file mode 100644 index 0000000000..5cd6c0c060 --- /dev/null +++ b/ggml/src/ggml-virtgpu/ggml-backend.cpp @@ -0,0 +1,69 @@ +#include "ggml-remoting.h" +#include "../../include/ggml-virtgpu.h" + +static const char * ggml_backend_remoting_get_name(ggml_backend_t backend) { + UNUSED(backend); + + return "API Remoting backend"; +} + +static void ggml_backend_remoting_free(ggml_backend_t backend) { + delete backend; +} + +static ggml_status ggml_backend_remoting_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) { + virtgpu * gpu = DEV_TO_GPU(backend->device); + + return apir_backend_graph_compute(gpu, cgraph); +} + +static void ggml_backend_remoting_graph_optimize(ggml_backend_t backend, ggml_cgraph * cgraph) { + virtgpu * gpu = DEV_TO_GPU(backend->device); +#if true + UNUSED(gpu); + UNUSED(cgraph); +#else + // not working yet + + apir_backend_graph_optimize(gpu, cgraph); +#endif +} + +static ggml_backend_i ggml_backend_remoting_interface = { + /* .get_name = */ ggml_backend_remoting_get_name, + /* .free = */ ggml_backend_remoting_free, + /* .set_tensor_async = */ NULL, // ggml_backend_remoting_set_tensor_async, + /* .get_tensor_async = */ NULL, // ggml_backend_remoting_get_tensor_async, + /* .cpy_tensor_async = */ NULL, // ggml_backend_remoting_cpy_tensor_async, + /* .synchronize = */ NULL, // ggml_backend_remoting_synchronize, + /* .graph_plan_create = */ NULL, + /* .graph_plan_free = */ NULL, + /* .graph_plan_update = */ NULL, + /* .graph_plan_compute = */ NULL, + /* .graph_compute = */ ggml_backend_remoting_graph_compute, + /* .event_record = */ NULL, + /* .event_wait = */ NULL, + /* .graph_optimize = */ ggml_backend_remoting_graph_optimize, +}; + +static ggml_guid_t ggml_backend_remoting_guid() { + static ggml_guid guid = { 0xb8, 0xf7, 0x4f, 0x86, 0x14, 0x03, 0x86, 0x02, + 0x91, 0xc8, 0xdd, 0xe9, 0x02, 0x3f, 0xc0, 0x2b }; + + return &guid; +} + +ggml_backend_t ggml_backend_remoting_device_init(ggml_backend_dev_t dev, const char * params) { + UNUSED(params); + + ggml_backend_remoting_device_context * ctx = (ggml_backend_remoting_device_context *) dev->context; + + ggml_backend_t remoting_backend = new ggml_backend{ + /* .guid = */ ggml_backend_remoting_guid(), + /* .interface = */ ggml_backend_remoting_interface, + /* .device = */ ggml_backend_reg_dev_get(ggml_backend_virtgpu_reg(), ctx->device), + /* .context = */ ctx, + }; + + return remoting_backend; +} diff --git a/ggml/src/ggml-virtgpu/ggml-remoting.h b/ggml/src/ggml-virtgpu/ggml-remoting.h new file mode 100644 index 0000000000..36fc6b2a7b --- /dev/null +++ b/ggml/src/ggml-virtgpu/ggml-remoting.h @@ -0,0 +1,68 @@ +#pragma once + +#include "ggml-backend-impl.h" +#include "ggml-backend.h" +#include "ggml-impl.h" +#include "virtgpu.h" + +#include +#include + +// USE_ALWAYS_TRUE_SUPPORTS_OP: 1 is fast, 0 avoid micro-benchmark crashes + +#define USE_ALWAYS_TRUE_SUPPORTS_OP 1 +#define USE_METAL_GUEST_SUPPORTS_OP 0 + +#define DEV_TO_GPU(name) ((ggml_backend_remoting_device_context *) (name)->context)->gpu + +#define BUFFER_TO_GGML_CONTEXT(name) ((ggml_backend_remoting_buffer_context *) (name)->context) + +#define BUFFER_TO_APIR_CONTEXT(name) &((ggml_backend_remoting_buffer_context *) (name)->context)->apir_context + +#define BUFFER_TO_HOST_HANDLE(name) ((ggml_backend_remoting_buffer_context *) (name)->context)->apir_context.host_handle + +#define GET_DEVICE_CONTEXT() (ggml_backend_remoting_device_context *) ggml_backend_remoting_get_device(0)->context + +#define BUFT_TO_GPU(name) ((ggml_backend_remoting_device_context *) (name)->device->context)->gpu + +struct ggml_backend_remoting_device_context { + size_t device; + std::string name; + std::string description; + + std::vector> shared_memory; + + virtgpu * gpu; +}; + +struct ggml_backend_remoting_buffer_context { + apir_buffer_context_t apir_context; + + virtgpu * gpu; + + void * base; + + bool is_from_ptr; +}; + +extern const ggml_backend_buffer_type_i ggml_backend_remoting_buffer_type_interface; +extern const ggml_backend_device_i ggml_backend_remoting_device_interface; +extern const ggml_backend_buffer_i ggml_backend_remoting_buffer_interface; +extern const ggml_backend_buffer_type_i ggml_backend_remoting_buffer_from_ptr_type_interface; +extern const ggml_backend_buffer_i ggml_backend_remoting_buffer_from_ptr_interface; + +ggml_backend_dev_t ggml_backend_remoting_get_device(size_t device); +ggml_backend_t ggml_backend_remoting_device_init(ggml_backend_dev_t dev, const char * params); +ggml_backend_buffer_type_t ggml_backend_remoting_device_get_buffer_type(ggml_backend_dev_t dev); + +static inline apir_buffer_type_host_handle_t ggml_buffer_type_to_apir_handle(ggml_backend_buffer_type_t buft) { + // in the backend, the buffer handle is the buffer pointer + return (apir_buffer_type_host_handle_t) buft->context; +} + +static inline apir_buffer_host_handle_t ggml_buffer_to_apir_handle(ggml_backend_buffer_t buffer) { + if (!buffer->context) { + GGML_ABORT("%s: no context available :/", __func__); + } + return BUFFER_TO_HOST_HANDLE(buffer); +} diff --git a/ggml/src/ggml-virtgpu/ggmlremoting_functions.yaml b/ggml/src/ggml-virtgpu/ggmlremoting_functions.yaml new file mode 100644 index 0000000000..0b7cccfe9c --- /dev/null +++ b/ggml/src/ggml-virtgpu/ggmlremoting_functions.yaml @@ -0,0 +1,168 @@ +# YAML schema for GGML remoting API functions +# This defines the structure for generating the remoting layer code + +# Configuration for the generated files +config: + # Base path for the generated files + base_path: "ggml/src" + + # Header files to update + files: + apir_backend_header: "ggml-virtgpu-apir/backend/shared/apir_backend.gen.h" + backend_dispatched_header: "ggml-virtgpu-apir/backend/backend-dispatched.gen.h" + virtgpu_forward_header: "ggml-virtgpu-apir/virtgpu-forward.gen.h" + +# Simplified function definitions with grouping and metadata combined +functions: + device: + group_description: "device" + functions: + get_device_count: + # No specific metadata - uses default void return and base params + + get_count: + frontend_return: "int" + + get_name: + frontend_return: "const char *" + + get_description: + frontend_return: "const char *" + + get_type: + frontend_return: "uint32_t" + + get_memory: + frontend_return: "void" + frontend_extra_params: + - "size_t *free" + - "size_t *total" + + supports_op: + frontend_return: "bool" + frontend_extra_params: + - "const ggml_tensor *op" + + get_buffer_type: + frontend_return: "apir_buffer_type_host_handle_t" + + get_props: + frontend_return: "void" + frontend_extra_params: + - "bool *async" + - "bool *host_buffer" + - "bool *buffer_from_host_ptr" + - "bool *events" + + buffer_from_ptr: + frontend_return: "apir_buffer_context_t" + frontend_extra_params: + - "size_t size" + - "size_t max_tensor_size" + + buffer_type: + group_description: "buffer-type" + functions: + get_name: + frontend_return: "const char *" + frontend_extra_params: + - "ggml_backend_buffer_type_t buft" + + get_alignment: + frontend_return: "size_t" + frontend_extra_params: + - "ggml_backend_buffer_type_t buft" + + get_max_size: + frontend_return: "size_t" + frontend_extra_params: + - "ggml_backend_buffer_type_t buft" + + is_host: + frontend_return: "bool" + frontend_extra_params: + - "ggml_backend_buffer_type_t buft" + + alloc_buffer: + frontend_return: "apir_buffer_context_t" + frontend_extra_params: + - "ggml_backend_buffer_type_t buffer_buft" + - "size_t size" + + get_alloc_size: + frontend_return: "size_t" + frontend_extra_params: + - "ggml_backend_buffer_type_t buft" + - "const ggml_tensor *op" + + buffer: + group_description: "buffer" + functions: + get_base: + frontend_return: "void *" + frontend_extra_params: + - "apir_buffer_context_t *buffer_context" + + set_tensor: + frontend_return: "void" + frontend_extra_params: + - "apir_buffer_context_t *buffer_context" + - "ggml_tensor *tensor" + - "const void *data" + - "size_t offset" + - "size_t size" + + get_tensor: + frontend_return: "void" + frontend_extra_params: + - "apir_buffer_context_t *buffer_context" + - "const ggml_tensor *tensor" + - "void *data" + - "size_t offset" + - "size_t size" + + cpy_tensor: + frontend_return: "bool" + frontend_extra_params: + - "apir_buffer_context_t *buffer_context" + - "const ggml_tensor *src" + - "const ggml_tensor *dst" + + clear: + frontend_return: "void" + frontend_extra_params: + - "apir_buffer_context_t *buffer_context" + - "uint8_t value" + + free_buffer: + frontend_return: "void" + frontend_extra_params: + - "apir_buffer_context_t *buffer_context" + + backend: + group_description: "backend" + functions: + graph_compute: + frontend_return: "ggml_status" + frontend_extra_params: + - "ggml_cgraph *cgraph" + + graph_optimize: + frontend_return: "ggml_cgraph *" + frontend_extra_params: + - "ggml_cgraph *cgraph" + enabled: false + +# Naming patterns used for code generation +naming_patterns: + # How to generate enum names + enum_prefix: "APIR_COMMAND_TYPE_" + + # How to generate backend function names + backend_function_prefix: "backend_" + + # How to generate frontend function names + frontend_function_prefix: "apir_" + + # Standard frontend first parameter + frontend_base_param: "struct virtgpu *gpu" diff --git a/ggml/src/ggml-virtgpu/include/apir_hw.h b/ggml/src/ggml-virtgpu/include/apir_hw.h new file mode 100644 index 0000000000..33af045ca2 --- /dev/null +++ b/ggml/src/ggml-virtgpu/include/apir_hw.h @@ -0,0 +1,9 @@ +#pragma once + +#include + +struct virgl_renderer_capset_apir { + uint32_t apir_version; + uint32_t supports_blob_resources; + uint32_t reserved[4]; // For future expansion +}; diff --git a/ggml/src/ggml-virtgpu/regenerate_remoting.py b/ggml/src/ggml-virtgpu/regenerate_remoting.py new file mode 100755 index 0000000000..4174a24327 --- /dev/null +++ b/ggml/src/ggml-virtgpu/regenerate_remoting.py @@ -0,0 +1,322 @@ +#!/usr/bin/env python3 +""" +# Generated by Claude AI + +Script to completely regenerate the GGML remoting codebase from YAML configuration. + +This script reads api_functions.yaml and regenerates all the header files and +implementation templates for the GGML remoting layer. + +Usage: + python regenerate_remoting.py + +The script will: +1. Read ggmlremoting_functions.yaml configuration +2. Generate updated header files +3. Generate implementation templates in dedicated files +4. Show a summary of what was generated +""" + +import yaml +from typing import Dict, List, Any +from pathlib import Path +import os +import subprocess +import shutil +import logging + +NL = '\n' # can't have f"{'\n'}" in f-strings + + +class RemotingCodebaseGenerator: + def __init__(self, yaml_path: str = "ggmlremoting_functions.yaml"): + """Initialize the generator with the YAML configuration.""" + self.yaml_path = yaml_path + + if not Path(yaml_path).exists(): + raise FileNotFoundError(f"Configuration file {yaml_path} not found") + + with open(yaml_path, 'r') as f: + self.config = yaml.safe_load(f) + + self.functions = self.config['functions'] + self.naming_patterns = self.config['naming_patterns'] + self.config_data = self.config['config'] + + # Check if clang-format is available + self.clang_format_available = self._check_clang_format_available() + + def _check_clang_format_available(self) -> bool: + """Check if clang-format is available in the system PATH.""" + return shutil.which("clang-format") is not None + + def _format_file_with_clang_format(self, file_path: Path) -> bool: + """Format a file with clang-format -i. Returns True if successful, False otherwise.""" + if not self.clang_format_available: + return False + + try: + subprocess.run( + ["clang-format", "-i", str(file_path)], + check=True, + capture_output=True, + text=True + ) + return True + except subprocess.CalledProcessError: + logging.exception(f" ⚠️ clang-format failed for {file_path}") + return False + except Exception as e: + logging.exception(f" ⚠️ Unexpected error formatting {file_path}: {e}") + return False + + def generate_enum_name(self, group_name: str, function_name: str) -> str: + """Generate the APIR_COMMAND_TYPE enum name for a function.""" + prefix = self.naming_patterns['enum_prefix'] + return f"{prefix}{group_name.upper()}_{function_name.upper()}" + + def generate_backend_function_name(self, group_name: str, function_name: str) -> str: + """Generate the backend function name.""" + function_key = f"{group_name}_{function_name}" + overrides = self.naming_patterns.get('backend_function_overrides', {}) + + if function_key in overrides: + return overrides[function_key] + + prefix = self.naming_patterns['backend_function_prefix'] + return f"{prefix}{group_name}_{function_name}" + + def generate_frontend_function_name(self, group_name: str, function_name: str) -> str: + """Generate the frontend function name.""" + prefix = self.naming_patterns['frontend_function_prefix'] + return f"{prefix}{group_name}_{function_name}" + + def get_enabled_functions(self) -> List[Dict[str, Any]]: + """Get all enabled functions with their metadata.""" + functions = [] + enum_value = 0 + + for group_name, group_data in self.functions.items(): + group_description = group_data['group_description'] + + for function_name, func_metadata in group_data['functions'].items(): + # Handle case where func_metadata is None or empty (functions with only comments) + if func_metadata is None: + func_metadata = {} + + # Functions are enabled by default unless explicitly disabled + if func_metadata.get('enabled', True): + functions.append({ + 'group_name': group_name, + 'function_name': function_name, + 'enum_name': self.generate_enum_name(group_name, function_name), + 'enum_value': enum_value, + 'backend_function': self.generate_backend_function_name(group_name, function_name), + 'frontend_function': self.generate_frontend_function_name(group_name, function_name), + 'frontend_return': func_metadata.get('frontend_return', 'void'), + 'frontend_extra_params': func_metadata.get('frontend_extra_params', []), + 'group_description': group_description, + 'newly_added': func_metadata.get('newly_added', False) + }) + enum_value += 1 + + return functions + + def generate_apir_backend_header(self) -> str: + """Generate the complete apir_backend.h file.""" + functions = self.get_enabled_functions() + + # Generate the enum section + enum_lines = ["typedef enum ApirBackendCommandType {"] + current_group = None + + for func in functions: + # Add comment for new group + if func['group_name'] != current_group: + enum_lines.append("") + enum_lines.append(f" /* {func['group_description']} */") + current_group = func['group_name'] + + enum_lines.append(f" {func['enum_name']} = {func['enum_value']},") + + # Add the count + total_count = len(functions) + enum_lines.append("\n // last command_type index + 1") + enum_lines.append(f" APIR_BACKEND_DISPATCH_TABLE_COUNT = {total_count},") + enum_lines.append("} ApirBackendCommandType;") + + # Full header template + header_content = NL.join(enum_lines) + "\n" + + return header_content + + def generate_backend_dispatched_header(self) -> str: + """Generate the complete backend-dispatched.h file.""" + functions = self.get_enabled_functions() + + # Function declarations + decl_lines = [] + current_group = None + + for func in functions: + if func['group_name'] != current_group: + decl_lines.append(f"\n/* {func['group_description']} */") + current_group = func['group_name'] + + signature = "uint32_t" + params = "apir_encoder *enc, apir_decoder *dec, virgl_apir_context *ctx" + decl_lines.append(f"{signature} {func['backend_function']}({params});") + + # Switch cases + switch_lines = [] + current_group = None + + for func in functions: + if func['group_name'] != current_group: + switch_lines.append(f" /* {func['group_description']} */") + current_group = func['group_name'] + + switch_lines.append(f" case {func['enum_name']}: return \"{func['backend_function']}\";") + + # Dispatch table + table_lines = [] + current_group = None + + for func in functions: + if func['group_name'] != current_group: + table_lines.append(f"\n /* {func['group_description']} */") + table_lines.append("") + current_group = func['group_name'] + + table_lines.append(f" /* {func['enum_name']} = */ {func['backend_function']},") + + header_content = f'''\ +#pragma once + +{NL.join(decl_lines)} + +static inline const char *backend_dispatch_command_name(ApirBackendCommandType type) +{{ + switch (type) {{ +{NL.join(switch_lines)} + + default: return "unknown"; + }} +}} + +extern "C" {{ +static const backend_dispatch_t apir_backend_dispatch_table[APIR_BACKEND_DISPATCH_TABLE_COUNT] = {{ + {NL.join(table_lines)} +}}; +}} +''' + return header_content + + def generate_virtgpu_forward_header(self) -> str: + """Generate the complete virtgpu-forward.gen.h file.""" + functions = self.get_enabled_functions() + + decl_lines = [] + current_group = None + + for func in functions: + if func['group_name'] != current_group: + decl_lines.append("") + decl_lines.append(f"/* {func['group_description']} */") + current_group = func['group_name'] + + # Build parameter list + params = [self.naming_patterns['frontend_base_param']] + params.extend(func['frontend_extra_params']) + param_str = ', '.join(params) + + decl_lines.append(f"{func['frontend_return']} {func['frontend_function']}({param_str});") + + header_content = f'''\ +#pragma once +{NL.join(decl_lines)} +''' + return header_content + + def regenerate_codebase(self) -> None: + """Regenerate the entire remoting codebase.""" + logging.info("🔄 Regenerating GGML Remoting Codebase...") + logging.info("=" * 50) + + # Detect if we're running from frontend directory + current_dir = os.getcwd() + is_frontend_dir = current_dir.endswith('ggml-virtgpu') + + if is_frontend_dir: + # Running from ggml/src/ggml-virtgpu-apir + logging.info("📍 Detected frontend directory execution") + frontend_base = Path(".") + else: + # Running from project root (fallback to original behavior) + logging.info("📍 Detected project root execution") + base_path = self.config_data.get('base_path', 'ggml/src') + frontend_base = Path(base_path) / "ggml-virtgpu" + + # Compute final file paths + backend_base = frontend_base / "backend" + apir_backend_path = backend_base / "shared" / "apir_backend.gen.h" + backend_dispatched_path = backend_base / "backend-dispatched.gen.h" + virtgpu_forward_path = frontend_base / "virtgpu-forward.gen.h" + + # Create output directories for each file + apir_backend_path.parent.mkdir(parents=True, exist_ok=True) + backend_dispatched_path.parent.mkdir(parents=True, exist_ok=True) + virtgpu_forward_path.parent.mkdir(parents=True, exist_ok=True) + + # Generate header files + logging.info("📁 Generating header files...") + + apir_backend_content = self.generate_apir_backend_header() + apir_backend_path.write_text(apir_backend_content) + logging.info(f" ✅ {apir_backend_path.resolve()}") + + backend_dispatched_content = self.generate_backend_dispatched_header() + backend_dispatched_path.write_text(backend_dispatched_content) + logging.info(f" ✅ {backend_dispatched_path.resolve()}") + + virtgpu_forward_content = self.generate_virtgpu_forward_header() + virtgpu_forward_path.write_text(virtgpu_forward_content) + logging.info(f" ✅ {virtgpu_forward_path.resolve()}") + + # Format generated files with clang-format + generated_files = [apir_backend_path, backend_dispatched_path, virtgpu_forward_path] + + if not self.clang_format_available: + logging.warning("\n⚠️clang-format not found in PATH. Generated files will not be formatted." + " Install clang-format to enable automatic code formatting.") + else: + logging.info("\n🎨 Formatting files with clang-format...") + for file_path in generated_files: + if self._format_file_with_clang_format(file_path): + logging.info(f" ✅ Formatted {file_path.name}") + else: + logging.warning(f" ❌ Failed to format {file_path.name}") + + # Generate summary + functions = self.get_enabled_functions() + total_functions = len(functions) + + logging.info("\n📊 Generation Summary:") + logging.info("=" * 50) + logging.info(f" Total functions: {total_functions}") + logging.info(f" Function groups: {len(self.functions)}") + logging.info(" Header files: 3") + logging.info(f" Working directory: {current_dir}") + + +def main(): + try: + generator = RemotingCodebaseGenerator() + generator.regenerate_codebase() + except Exception as e: + logging.exception(f"❌ Error: {e}") + exit(1) + + +if __name__ == "__main__": + main() diff --git a/ggml/src/ggml-virtgpu/virtgpu-apir.h b/ggml/src/ggml-virtgpu/virtgpu-apir.h new file mode 100644 index 0000000000..238f960acd --- /dev/null +++ b/ggml/src/ggml-virtgpu/virtgpu-apir.h @@ -0,0 +1,15 @@ +#include "backend/shared/apir_backend.h" +#include "ggml-alloc.h" +#include "ggml-impl.h" +#include "ggml.h" +#include "virtgpu-shm.h" +#include "virtgpu-utils.h" + +struct apir_buffer_context_t { + apir_buffer_host_handle_t host_handle; + + struct virtgpu_shmem shmem; + apir_buffer_type_host_handle_t buft_host_handle; +}; + +#include "virtgpu-forward.gen.h" diff --git a/ggml/src/ggml-virtgpu/virtgpu-forward-backend.cpp b/ggml/src/ggml-virtgpu/virtgpu-forward-backend.cpp new file mode 100644 index 0000000000..bf3c41011a --- /dev/null +++ b/ggml/src/ggml-virtgpu/virtgpu-forward-backend.cpp @@ -0,0 +1,50 @@ +#include "virtgpu-forward-impl.h" + +static long long current_time_ms() { + timespec ts; + clock_gettime(CLOCK_REALTIME, &ts); // Use CLOCK_MONOTONIC for elapsed time + return (long long) ts.tv_sec * 1000000000LL + ts.tv_nsec; +} + +ggml_status apir_backend_graph_compute(virtgpu * gpu, ggml_cgraph * cgraph) { + apir_encoder * encoder; + apir_decoder * decoder; + ApirForwardReturnCode ret; + + REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_BACKEND_GRAPH_COMPUTE); + + std::vector cgraph_data; + size_t cgraph_size = apir_serialize_ggml_cgraph(cgraph, cgraph_data); + + virtgpu_shmem temp_shmem; // Local storage for large buffers + virtgpu_shmem * shmem = &temp_shmem; + + if (cgraph_size <= gpu->data_shmem.mmap_size) { + // prefer the init-time allocated page, if large enough + shmem = &gpu->data_shmem; + } else if (virtgpu_shmem_create(gpu, cgraph_size, shmem)) { + GGML_ABORT("Couldn't allocate the guest-host shared buffer"); + } + + apir_encode_virtgpu_shmem_res_id(encoder, shmem->res_id); + + apir_encode_size_t(encoder, &cgraph_size); + + char * shmem_data = (char *) shmem->mmap_ptr; + apir_encoder secondary_enc = apir_new_encoder(shmem_data, cgraph_size); + + apir_encode_cgraph_data(&secondary_enc, cgraph_data); + + REMOTE_CALL(gpu, encoder, decoder, ret); + + ggml_status status = GGML_STATUS_ABORTED; + apir_decode_ggml_status(decoder, &status); + + remote_call_finish(gpu, encoder, decoder); + + if (shmem != &gpu->data_shmem) { + virtgpu_shmem_destroy(gpu, shmem); + } + + return status; +} diff --git a/ggml/src/ggml-virtgpu/virtgpu-forward-buffer-type.cpp b/ggml/src/ggml-virtgpu/virtgpu-forward-buffer-type.cpp new file mode 100644 index 0000000000..03cb09e064 --- /dev/null +++ b/ggml/src/ggml-virtgpu/virtgpu-forward-buffer-type.cpp @@ -0,0 +1,125 @@ +#include "virtgpu-forward-impl.h" + +const char * apir_buffer_type_get_name(virtgpu * gpu, ggml_backend_buffer_type_t buft) { + apir_encoder * encoder; + apir_decoder * decoder; + ApirForwardReturnCode ret; + + REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_BUFFER_TYPE_GET_NAME); + + apir_encode_ggml_buffer_type(encoder, buft); + + REMOTE_CALL(gpu, encoder, decoder, ret); + + const size_t string_size = apir_decode_array_size_unchecked(decoder); + char * string = (char *) apir_decoder_alloc_array(sizeof(char), string_size); + if (!string) { + GGML_LOG_ERROR("%s: Could not allocate the device name buffer\n", __func__); + apir_decoder_set_fatal(decoder); + } + apir_decode_char_array(decoder, string, string_size); + + remote_call_finish(gpu, encoder, decoder); + + return string; +} + +size_t apir_buffer_type_get_alignment(virtgpu * gpu, ggml_backend_buffer_type_t buft) { + apir_encoder * encoder; + apir_decoder * decoder; + ApirForwardReturnCode ret; + + REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_BUFFER_TYPE_GET_ALIGNMENT); + + apir_encode_ggml_buffer_type(encoder, buft); + + REMOTE_CALL(gpu, encoder, decoder, ret); + + size_t alignment; + apir_decode_size_t(decoder, &alignment); + + remote_call_finish(gpu, encoder, decoder); + + return alignment; +} + +size_t apir_buffer_type_get_max_size(virtgpu * gpu, ggml_backend_buffer_type_t buft) { + apir_encoder * encoder; + apir_decoder * decoder; + ApirForwardReturnCode ret; + + REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_BUFFER_TYPE_GET_MAX_SIZE); + + apir_encode_ggml_buffer_type(encoder, buft); + + REMOTE_CALL(gpu, encoder, decoder, ret); + + size_t max_size; + apir_decode_size_t(decoder, &max_size); + + remote_call_finish(gpu, encoder, decoder); + + return max_size; +} + +bool apir_buffer_type_is_host(virtgpu * gpu, ggml_backend_buffer_type_t buft) { + apir_encoder * encoder; + apir_decoder * decoder; + ApirForwardReturnCode ret; + + REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_BUFFER_TYPE_IS_HOST); + + apir_encode_ggml_buffer_type(encoder, buft); + + REMOTE_CALL(gpu, encoder, decoder, ret); + + bool is_host; + apir_decode_bool_t(decoder, &is_host); + + remote_call_finish(gpu, encoder, decoder); + + return is_host; +} + +apir_buffer_context_t apir_buffer_type_alloc_buffer(virtgpu * gpu, ggml_backend_buffer_type_t buft, size_t size) { + apir_encoder * encoder; + apir_decoder * decoder; + ApirForwardReturnCode ret; + + apir_buffer_context_t buffer_context; + + REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_BUFFER_TYPE_ALLOC_BUFFER); + + apir_encode_ggml_buffer_type(encoder, buft); + + apir_encode_size_t(encoder, &size); + + REMOTE_CALL(gpu, encoder, decoder, ret); + + apir_decode_apir_buffer_host_handle_t(decoder, &buffer_context.host_handle); + + remote_call_finish(gpu, encoder, decoder); + + return buffer_context; +} + +size_t apir_buffer_type_get_alloc_size(virtgpu * gpu, ggml_backend_buffer_type_t buft, const ggml_tensor * op) { + apir_encoder * encoder; + apir_decoder * decoder; + ApirForwardReturnCode ret; + + REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_BUFFER_TYPE_GET_ALLOC_SIZE); + + apir_encode_ggml_buffer_type(encoder, buft); + + apir_encode_ggml_tensor_inline(encoder, op); + + REMOTE_CALL(gpu, encoder, decoder, ret); + + size_t alloc_size; + apir_decode_size_t(decoder, &alloc_size); + + remote_call_finish(gpu, encoder, decoder); + + return alloc_size; +} diff --git a/ggml/src/ggml-virtgpu/virtgpu-forward-buffer.cpp b/ggml/src/ggml-virtgpu/virtgpu-forward-buffer.cpp new file mode 100644 index 0000000000..3181e39440 --- /dev/null +++ b/ggml/src/ggml-virtgpu/virtgpu-forward-buffer.cpp @@ -0,0 +1,157 @@ +#include "virtgpu-forward-impl.h" + +void * apir_buffer_get_base(virtgpu * gpu, apir_buffer_context_t * buffer_context) { + apir_encoder * encoder; + apir_decoder * decoder; + ApirForwardReturnCode ret; + + REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_BUFFER_GET_BASE); + + apir_encode_apir_buffer_host_handle_t(encoder, &buffer_context->host_handle); + + REMOTE_CALL(gpu, encoder, decoder, ret); + + uintptr_t base; + apir_decode_uintptr_t(decoder, &base); + + remote_call_finish(gpu, encoder, decoder); + + return (void *) base; +} + +void apir_buffer_set_tensor(virtgpu * gpu, + apir_buffer_context_t * buffer_context, + ggml_tensor * tensor, + const void * data, + size_t offset, + size_t size) { + apir_encoder * encoder; + apir_decoder * decoder; + ApirForwardReturnCode ret; + + REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_BUFFER_SET_TENSOR); + + apir_encode_apir_buffer_host_handle_t(encoder, &buffer_context->host_handle); + apir_encode_ggml_tensor(encoder, tensor); + + virtgpu_shmem temp_shmem; // Local storage for large buffers + virtgpu_shmem * shmem = &temp_shmem; + + if (size <= gpu->data_shmem.mmap_size) { + // prefer the init-time allocated page, if large enough + shmem = &gpu->data_shmem; + + } else if (virtgpu_shmem_create(gpu, size, shmem)) { + GGML_ABORT("Couldn't allocate the guest-host shared buffer"); + } + + memcpy(shmem->mmap_ptr, data, size); + apir_encode_virtgpu_shmem_res_id(encoder, shmem->res_id); + + apir_encode_size_t(encoder, &offset); + apir_encode_size_t(encoder, &size); + + REMOTE_CALL(gpu, encoder, decoder, ret); + + remote_call_finish(gpu, encoder, decoder); + + if (shmem != &gpu->data_shmem) { + virtgpu_shmem_destroy(gpu, shmem); + } + + return; +} + +void apir_buffer_get_tensor(virtgpu * gpu, + apir_buffer_context_t * buffer_context, + const ggml_tensor * tensor, + void * data, + size_t offset, + size_t size) { + apir_encoder * encoder; + apir_decoder * decoder; + ApirForwardReturnCode ret; + + REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_BUFFER_GET_TENSOR); + + apir_encode_apir_buffer_host_handle_t(encoder, &buffer_context->host_handle); + apir_encode_ggml_tensor(encoder, tensor); + + virtgpu_shmem temp_shmem; // Local storage for large buffers + virtgpu_shmem * shmem = &temp_shmem; + + if (size <= gpu->data_shmem.mmap_size) { + // prefer the init-time allocated page, if large enough + shmem = &gpu->data_shmem; + + } else if (virtgpu_shmem_create(gpu, size, shmem)) { + GGML_ABORT("Couldn't allocate the guest-host shared buffer"); + } + + apir_encode_virtgpu_shmem_res_id(encoder, shmem->res_id); + apir_encode_size_t(encoder, &offset); + apir_encode_size_t(encoder, &size); + + REMOTE_CALL(gpu, encoder, decoder, ret); + + memcpy(data, shmem->mmap_ptr, size); + + remote_call_finish(gpu, encoder, decoder); + + if (shmem != &gpu->data_shmem) { + virtgpu_shmem_destroy(gpu, shmem); + } +} + +bool apir_buffer_cpy_tensor(virtgpu * gpu, + apir_buffer_context_t * buffer_context, + const ggml_tensor * src, + const ggml_tensor * dst) { + apir_encoder * encoder; + apir_decoder * decoder; + ApirForwardReturnCode ret; + + REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_BUFFER_CPY_TENSOR); + + apir_encode_apir_buffer_host_handle_t(encoder, &buffer_context->host_handle); + apir_encode_ggml_tensor(encoder, src); + apir_encode_ggml_tensor(encoder, dst); + + REMOTE_CALL(gpu, encoder, decoder, ret); + + bool ret_val; + apir_decode_bool_t(decoder, &ret_val); + + remote_call_finish(gpu, encoder, decoder); + + return ret_val; +} + +void apir_buffer_clear(virtgpu * gpu, apir_buffer_context_t * buffer_context, uint8_t value) { + apir_encoder * encoder; + apir_decoder * decoder; + ApirForwardReturnCode ret; + + REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_BUFFER_CLEAR); + + apir_encode_apir_buffer_host_handle_t(encoder, &buffer_context->host_handle); + apir_encode_uint8_t(encoder, &value); + + REMOTE_CALL(gpu, encoder, decoder, ret); + + remote_call_finish(gpu, encoder, decoder); +} + +void apir_buffer_free_buffer(virtgpu * gpu, apir_buffer_context_t * buffer_context) { + apir_encoder * encoder; + apir_decoder * decoder; + ApirForwardReturnCode ret; + + REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_BUFFER_FREE_BUFFER); + + apir_encode_apir_buffer_host_handle_t(encoder, &buffer_context->host_handle); + + REMOTE_CALL(gpu, encoder, decoder, ret); + + remote_call_finish(gpu, encoder, decoder); +} diff --git a/ggml/src/ggml-virtgpu/virtgpu-forward-device.cpp b/ggml/src/ggml-virtgpu/virtgpu-forward-device.cpp new file mode 100644 index 0000000000..3e45e55bdc --- /dev/null +++ b/ggml/src/ggml-virtgpu/virtgpu-forward-device.cpp @@ -0,0 +1,200 @@ +#include "virtgpu-forward-impl.h" +#include "virtgpu-shm.h" + +int apir_device_get_count(virtgpu * gpu) { + static int32_t dev_count = -1; + if (dev_count != -1) { + return dev_count; + } + + apir_encoder * encoder; + apir_decoder * decoder; + ApirForwardReturnCode ret; + + REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_DEVICE_GET_COUNT); + REMOTE_CALL(gpu, encoder, decoder, ret); + + apir_decode_int32_t(decoder, &dev_count); + + remote_call_finish(gpu, encoder, decoder); + + return dev_count; +} + +const char * apir_device_get_name(virtgpu * gpu) { + static char * string = nullptr; + if (string) { + return string; + } + apir_encoder * encoder; + apir_decoder * decoder; + ApirForwardReturnCode ret; + + REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_DEVICE_GET_NAME); + REMOTE_CALL(gpu, encoder, decoder, ret); + + const size_t string_size = apir_decode_array_size_unchecked(decoder); + string = (char *) apir_decoder_alloc_array(sizeof(char), string_size); + if (!string) { + GGML_LOG_ERROR("%s: Could not allocate the device name buffer\n", __func__); + return NULL; + } + apir_decode_char_array(decoder, string, string_size); + + remote_call_finish(gpu, encoder, decoder); + + return string; +} + +const char * apir_device_get_description(virtgpu * gpu) { + apir_encoder * encoder; + apir_decoder * decoder; + ApirForwardReturnCode ret; + + REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_DEVICE_GET_DESCRIPTION); + + REMOTE_CALL(gpu, encoder, decoder, ret); + + const size_t string_size = apir_decode_array_size_unchecked(decoder); + char * string = (char *) apir_decoder_alloc_array(sizeof(char), string_size); + if (!string) { + GGML_LOG_ERROR("%s: Could not allocate the device description buffer\n", __func__); + + return NULL; + } + apir_decode_char_array(decoder, string, string_size); + + remote_call_finish(gpu, encoder, decoder); + + return string; +} + +uint32_t apir_device_get_type(virtgpu * gpu) { + static uint32_t dev_type = 255; + if (dev_type != 255) { + return dev_type; + } + + apir_encoder * encoder; + apir_decoder * decoder; + ApirForwardReturnCode ret; + + REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_DEVICE_GET_TYPE); + + REMOTE_CALL(gpu, encoder, decoder, ret); + + apir_decode_uint32_t(decoder, &dev_type); + + remote_call_finish(gpu, encoder, decoder); + + return dev_type; +} + +void apir_device_get_memory(virtgpu * gpu, size_t * free, size_t * total) { + static size_t dev_free = 0; + static size_t dev_total = 0; + apir_encoder * encoder; + apir_decoder * decoder; + ApirForwardReturnCode ret; + + REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_DEVICE_GET_MEMORY); + + REMOTE_CALL(gpu, encoder, decoder, ret); + + apir_decode_size_t(decoder, &dev_free); + apir_decode_size_t(decoder, &dev_total); + + *free = dev_free; + *total = dev_total; + + remote_call_finish(gpu, encoder, decoder); + + return; +} + +bool apir_device_supports_op(virtgpu * gpu, const ggml_tensor * op) { + apir_encoder * encoder; + apir_decoder * decoder; + ApirForwardReturnCode ret; + + REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_DEVICE_SUPPORTS_OP); + + apir_encode_ggml_tensor_inline(encoder, op); + + REMOTE_CALL(gpu, encoder, decoder, ret); + + bool supports_op; + apir_decode_bool_t(decoder, &supports_op); + + remote_call_finish(gpu, encoder, decoder); + + return supports_op; +} + +apir_buffer_type_host_handle_t apir_device_get_buffer_type(virtgpu * gpu) { + apir_encoder * encoder; + apir_decoder * decoder; + ApirForwardReturnCode ret; + + REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_DEVICE_GET_BUFFER_TYPE); + + REMOTE_CALL(gpu, encoder, decoder, ret); + + apir_buffer_type_host_handle_t buft_handle; + apir_decode_apir_buffer_type_host_handle_t(decoder, &buft_handle); + + remote_call_finish(gpu, encoder, decoder); + + return buft_handle; +} + +void apir_device_get_props(virtgpu * gpu, + bool * async, + bool * host_buffer, + bool * buffer_from_host_ptr, + bool * events) { + apir_encoder * encoder; + apir_decoder * decoder; + ApirForwardReturnCode ret; + + REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_DEVICE_GET_PROPS); + + REMOTE_CALL(gpu, encoder, decoder, ret); + + apir_decode_bool_t(decoder, async); + apir_decode_bool_t(decoder, host_buffer); + apir_decode_bool_t(decoder, buffer_from_host_ptr); + apir_decode_bool_t(decoder, events); + + remote_call_finish(gpu, encoder, decoder); + + return; +} + +apir_buffer_context_t apir_device_buffer_from_ptr(virtgpu * gpu, size_t size, size_t max_tensor_size) { + apir_encoder * encoder; + apir_decoder * decoder; + ApirForwardReturnCode ret; + + apir_buffer_context_t buffer_context; + + REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_DEVICE_BUFFER_FROM_PTR); + + if (virtgpu_shmem_create(gpu, size, &buffer_context.shmem)) { + GGML_ABORT("Couldn't allocate the guest-host shared buffer"); + } + + apir_encode_virtgpu_shmem_res_id(encoder, buffer_context.shmem.res_id); + + apir_encode_size_t(encoder, &size); + apir_encode_size_t(encoder, &max_tensor_size); + + REMOTE_CALL(gpu, encoder, decoder, ret); + + apir_decode_apir_buffer_host_handle_t(decoder, &buffer_context.host_handle); + buffer_context.buft_host_handle = apir_decode_apir_buffer_type_host_handle(decoder); + + remote_call_finish(gpu, encoder, decoder); + + return buffer_context; +} diff --git a/ggml/src/ggml-virtgpu/virtgpu-forward-impl.h b/ggml/src/ggml-virtgpu/virtgpu-forward-impl.h new file mode 100644 index 0000000000..eea3e7e5a9 --- /dev/null +++ b/ggml/src/ggml-virtgpu/virtgpu-forward-impl.h @@ -0,0 +1,29 @@ +#include "virtgpu.h" + +#include "ggml-remoting.h" +#include "backend/shared/apir_backend.h" +#include "backend/shared/apir_cs_ggml.h" + +#include "ggml-backend-impl.h" + +#define REMOTE_CALL_PREPARE(gpu_dev_name, encoder_name, apir_command_type__) \ + do { \ + int32_t forward_flag = (int32_t) apir_command_type__; \ + encoder_name = remote_call_prepare(gpu_dev_name, APIR_COMMAND_TYPE_FORWARD, forward_flag); \ + if (!encoder_name) { \ + GGML_ABORT("%s: failed to prepare the remote call encoder", __func__); \ + } \ + } while (0) + +#define REMOTE_CALL(gpu_dev_name, encoder_name, decoder_name, ret_name) \ + do { \ + ret_name = (ApirForwardReturnCode) remote_call(gpu_dev_name, encoder_name, &decoder_name, 0, NULL); \ + if (!decoder_name) { \ + GGML_ABORT("%s: failed to kick the remote call", __func__); \ + } \ + if (ret_name < APIR_FORWARD_BASE_INDEX) { \ + GGML_ABORT("%s: failed to forward the API call: %s: code %d", __func__, \ + apir_forward_error(ret_name), ret_name); \ + } \ + ret_name = (ApirForwardReturnCode) (ret_name - APIR_FORWARD_BASE_INDEX); \ + } while (0) diff --git a/ggml/src/ggml-virtgpu/virtgpu-forward.gen.h b/ggml/src/ggml-virtgpu/virtgpu-forward.gen.h new file mode 100644 index 0000000000..c27c07f086 --- /dev/null +++ b/ggml/src/ggml-virtgpu/virtgpu-forward.gen.h @@ -0,0 +1,51 @@ +#pragma once + +/* device */ +void apir_device_get_device_count(struct virtgpu * gpu); +int apir_device_get_count(struct virtgpu * gpu); +const char * apir_device_get_name(struct virtgpu * gpu); +const char * apir_device_get_description(struct virtgpu * gpu); +uint32_t apir_device_get_type(struct virtgpu * gpu); +void apir_device_get_memory(struct virtgpu * gpu, size_t * free, size_t * total); +bool apir_device_supports_op(struct virtgpu * gpu, const ggml_tensor * op); +apir_buffer_type_host_handle_t apir_device_get_buffer_type(struct virtgpu * gpu); +void apir_device_get_props(struct virtgpu * gpu, + bool * async, + bool * host_buffer, + bool * buffer_from_host_ptr, + bool * events); +apir_buffer_context_t apir_device_buffer_from_ptr(struct virtgpu * gpu, size_t size, size_t max_tensor_size); + +/* buffer-type */ +const char * apir_buffer_type_get_name(struct virtgpu * gpu, ggml_backend_buffer_type_t buft); +size_t apir_buffer_type_get_alignment(struct virtgpu * gpu, ggml_backend_buffer_type_t buft); +size_t apir_buffer_type_get_max_size(struct virtgpu * gpu, ggml_backend_buffer_type_t buft); +bool apir_buffer_type_is_host(struct virtgpu * gpu, ggml_backend_buffer_type_t buft); +apir_buffer_context_t apir_buffer_type_alloc_buffer(struct virtgpu * gpu, + ggml_backend_buffer_type_t buffer_buft, + size_t size); +size_t apir_buffer_type_get_alloc_size(struct virtgpu * gpu, ggml_backend_buffer_type_t buft, const ggml_tensor * op); + +/* buffer */ +void * apir_buffer_get_base(struct virtgpu * gpu, apir_buffer_context_t * buffer_context); +void apir_buffer_set_tensor(struct virtgpu * gpu, + apir_buffer_context_t * buffer_context, + ggml_tensor * tensor, + const void * data, + size_t offset, + size_t size); +void apir_buffer_get_tensor(struct virtgpu * gpu, + apir_buffer_context_t * buffer_context, + const ggml_tensor * tensor, + void * data, + size_t offset, + size_t size); +bool apir_buffer_cpy_tensor(struct virtgpu * gpu, + apir_buffer_context_t * buffer_context, + const ggml_tensor * src, + const ggml_tensor * dst); +void apir_buffer_clear(struct virtgpu * gpu, apir_buffer_context_t * buffer_context, uint8_t value); +void apir_buffer_free_buffer(struct virtgpu * gpu, apir_buffer_context_t * buffer_context); + +/* backend */ +ggml_status apir_backend_graph_compute(struct virtgpu * gpu, ggml_cgraph * cgraph); diff --git a/ggml/src/ggml-virtgpu/virtgpu-shm.cpp b/ggml/src/ggml-virtgpu/virtgpu-shm.cpp new file mode 100644 index 0000000000..4def405a62 --- /dev/null +++ b/ggml/src/ggml-virtgpu/virtgpu-shm.cpp @@ -0,0 +1,99 @@ +#include "virtgpu-shm.h" + +#include "virtgpu.h" + +#include + +static uint32_t virtgpu_ioctl_resource_create_blob(virtgpu * gpu, + uint32_t blob_mem, + uint32_t blob_flags, + size_t blob_size, + uint64_t blob_id, + uint32_t * res_id) { +#ifdef SIMULATE_BO_SIZE_FIX + blob_size = align64(blob_size, 4096); +#endif + + drm_virtgpu_resource_create_blob args = { + .blob_mem = blob_mem, + .blob_flags = blob_flags, + .bo_handle = 0, + .res_handle = 0, + .size = blob_size, + .pad = 0, + .cmd_size = 0, + .cmd = 0, + .blob_id = blob_id, + }; + + if (virtgpu_ioctl(gpu, DRM_IOCTL_VIRTGPU_RESOURCE_CREATE_BLOB, &args)) { + return 0; + } + + *res_id = args.res_handle; + return args.bo_handle; +} + +static void virtgpu_ioctl_gem_close(virtgpu * gpu, uint32_t gem_handle) { + drm_gem_close args = { + .handle = gem_handle, + .pad = 0, + }; + + const int ret = virtgpu_ioctl(gpu, DRM_IOCTL_GEM_CLOSE, &args); + assert(!ret); +#ifdef NDEBUG + UNUSED(ret); +#endif +} + +static void * virtgpu_ioctl_map(virtgpu * gpu, uint32_t gem_handle, size_t size) { + drm_virtgpu_map args = { + .offset = 0, + .handle = gem_handle, + .pad = 0, + }; + + if (virtgpu_ioctl(gpu, DRM_IOCTL_VIRTGPU_MAP, &args)) { + return NULL; + } + + void * ptr = mmap(NULL, size, PROT_READ | PROT_WRITE, MAP_SHARED, gpu->fd, args.offset); + if (ptr == MAP_FAILED) { + return NULL; + } + + return ptr; +} + +void virtgpu_shmem_destroy(virtgpu * gpu, virtgpu_shmem * shmem) { + munmap(shmem->mmap_ptr, shmem->mmap_size); + virtgpu_ioctl_gem_close(gpu, shmem->gem_handle); +} + +int virtgpu_shmem_create(virtgpu * gpu, size_t size, virtgpu_shmem * shmem) { + size = align64(size, 16384); + + uint32_t res_id; + uint32_t gem_handle = virtgpu_ioctl_resource_create_blob(gpu, VIRTGPU_BLOB_MEM_HOST3D, + VIRTGPU_BLOB_FLAG_USE_MAPPABLE, size, 0, &res_id); + + if (!gem_handle) { + return 1; + } + + void * ptr = virtgpu_ioctl_map(gpu, gem_handle, size); + if (!ptr) { + virtgpu_ioctl_gem_close(gpu, gem_handle); + GGML_LOG_ERROR("virtgpu_ioctl_map FAILED\n"); + exit(1); + return 1; + } + + shmem->res_id = res_id; + shmem->mmap_size = size; + shmem->mmap_ptr = ptr; + shmem->gem_handle = gem_handle; + + return 0; +} diff --git a/ggml/src/ggml-virtgpu/virtgpu-shm.h b/ggml/src/ggml-virtgpu/virtgpu-shm.h new file mode 100644 index 0000000000..606860a094 --- /dev/null +++ b/ggml/src/ggml-virtgpu/virtgpu-shm.h @@ -0,0 +1,23 @@ +#pragma once + +#include "virtgpu-utils.h" + +#include + +#include +#include +#include +#include + +struct virtgpu; + +struct virtgpu_shmem { + uint32_t res_id; + size_t mmap_size; + void * mmap_ptr; + + uint32_t gem_handle; +}; + +int virtgpu_shmem_create(virtgpu * gpu, size_t size, virtgpu_shmem * shmem); +void virtgpu_shmem_destroy(virtgpu * gpu, virtgpu_shmem * shmem); diff --git a/ggml/src/ggml-virtgpu/virtgpu-utils.cpp b/ggml/src/ggml-virtgpu/virtgpu-utils.cpp new file mode 100644 index 0000000000..8a2805e990 --- /dev/null +++ b/ggml/src/ggml-virtgpu/virtgpu-utils.cpp @@ -0,0 +1,179 @@ +#include "virtgpu-utils.h" + +#include +#include + +#include + +#define NODE_ALLOC_ALIGN 64 +#define NODE_PTR_MASK (~((uintptr_t) NODE_ALLOC_ALIGN - 1)) +#define NODE_LEVEL_MASK ((uintptr_t) NODE_ALLOC_ALIGN - 1) +#define NULL_NODE 0 + +#define os_malloc_aligned(_size, _align) _aligned_malloc(_size, _align) +#define os_free_aligned(_ptr) free(_ptr) +#define p_atomic_cmpxchg(v, old, _new) __sync_val_compare_and_swap((v), (old), (_new)) + +static inline uint64_t util_logbase2_64(uint64_t n) { +#if defined(HAVE___BUILTIN_CLZLL) + return ((sizeof(uint64_t) * 8 - 1) - __builtin_clzll(n | 1)); +#else + uint64_t pos = 0ull; + if (n >= 1ull << 32) { + n >>= 32; + pos += 32; + } + if (n >= 1ull << 16) { + n >>= 16; + pos += 16; + } + if (n >= 1ull << 8) { + n >>= 8; + pos += 8; + } + if (n >= 1ull << 4) { + n >>= 4; + pos += 4; + } + if (n >= 1ull << 2) { + n >>= 2; + pos += 2; + } + if (n >= 1ull << 1) { + pos += 1; + } + return pos; +#endif +} + +void util_sparse_array_init(util_sparse_array * arr, size_t elem_size, size_t node_size) { + memset(arr, 0, sizeof(*arr)); + arr->elem_size = elem_size; + arr->node_size_log2 = util_logbase2_64(node_size); + assert(node_size >= 2 && node_size == (1ull << arr->node_size_log2)); +} + +static inline void * os_malloc_aligned(size_t size, size_t alignment) { + void * ptr; + alignment = (alignment + sizeof(void *) - 1) & ~(sizeof(void *) - 1); + if (posix_memalign(&ptr, alignment, size) != 0) { + return NULL; + } + return ptr; +} + +static inline void * _util_sparse_array_node_data(uintptr_t handle) { + return (void *) (handle & NODE_PTR_MASK); +} + +static inline unsigned _util_sparse_array_node_level(uintptr_t handle) { + return handle & NODE_LEVEL_MASK; +} + +static inline void _util_sparse_array_node_finish(util_sparse_array * arr, uintptr_t node) { + if (_util_sparse_array_node_level(node) > 0) { + uintptr_t * children = (uintptr_t *) _util_sparse_array_node_data(node); + size_t node_size = 1ull << arr->node_size_log2; + for (size_t i = 0; i < node_size; i++) { + if (children[i]) { + _util_sparse_array_node_finish(arr, children[i]); + } + } + } + + os_free_aligned(_util_sparse_array_node_data(node)); +} + +static inline uintptr_t _util_sparse_array_node(void * data, unsigned level) { + assert(data != NULL); + assert(((uintptr_t) data & NODE_LEVEL_MASK) == 0); + assert((level & NODE_PTR_MASK) == 0); + return (uintptr_t) data | level; +} + +inline uintptr_t _util_sparse_array_node_alloc(util_sparse_array * arr, unsigned level) { + size_t size; + if (level == 0) { + size = arr->elem_size << arr->node_size_log2; + } else { + size = sizeof(uintptr_t) << arr->node_size_log2; + } + + void * data = os_malloc_aligned(size, NODE_ALLOC_ALIGN); + memset(data, 0, size); + + return _util_sparse_array_node(data, level); +} + +static inline uintptr_t _util_sparse_array_set_or_free_node(uintptr_t * node_ptr, uintptr_t cmp_node, uintptr_t node) { + uintptr_t prev_node = p_atomic_cmpxchg(node_ptr, cmp_node, node); + + if (prev_node != cmp_node) { + /* We lost the race. Free this one and return the one that was already + * allocated. + */ + os_free_aligned(_util_sparse_array_node_data(node)); + return prev_node; + } else { + return node; + } +} + +void * util_sparse_array_get(util_sparse_array * arr, uint64_t idx) { + const unsigned node_size_log2 = arr->node_size_log2; + uintptr_t root = p_atomic_read(&arr->root); + if (unlikely(!root)) { + unsigned root_level = 0; + uint64_t idx_iter = idx >> node_size_log2; + while (idx_iter) { + idx_iter >>= node_size_log2; + root_level++; + } + uintptr_t new_root = _util_sparse_array_node_alloc(arr, root_level); + root = _util_sparse_array_set_or_free_node(&arr->root, NULL_NODE, new_root); + } + + while (1) { + unsigned root_level = _util_sparse_array_node_level(root); + uint64_t root_idx = idx >> (root_level * node_size_log2); + if (likely(root_idx < (1ull << node_size_log2))) { + break; + } + + /* In this case, we have a root but its level is low enough that the + * requested index is out-of-bounds. + */ + uintptr_t new_root = _util_sparse_array_node_alloc(arr, root_level + 1); + + uintptr_t * new_root_children = (uintptr_t *) _util_sparse_array_node_data(new_root); + new_root_children[0] = root; + + /* We only add one at a time instead of the whole tree because it's + * easier to ensure correctness of both the tree building and the + * clean-up path. Because we're only adding one node we never have to + * worry about trying to free multiple things without freeing the old + * things. + */ + root = _util_sparse_array_set_or_free_node(&arr->root, root, new_root); + } + + void * node_data = _util_sparse_array_node_data(root); + unsigned node_level = _util_sparse_array_node_level(root); + while (node_level > 0) { + uint64_t child_idx = (idx >> (node_level * node_size_log2)) & ((1ull << node_size_log2) - 1); + + uintptr_t * children = (uintptr_t *) node_data; + uintptr_t child = p_atomic_read(&children[child_idx]); + + if (unlikely(!child)) { + child = _util_sparse_array_node_alloc(arr, node_level - 1); + child = _util_sparse_array_set_or_free_node(&children[child_idx], NULL_NODE, child); + } + + node_data = _util_sparse_array_node_data(child); + node_level = _util_sparse_array_node_level(child); + } + + uint64_t elem_idx = idx & ((1ull << node_size_log2) - 1); + return (void *) ((char *) node_data + (elem_idx * arr->elem_size)); +} diff --git a/ggml/src/ggml-virtgpu/virtgpu-utils.h b/ggml/src/ggml-virtgpu/virtgpu-utils.h new file mode 100644 index 0000000000..a0036b4e2b --- /dev/null +++ b/ggml/src/ggml-virtgpu/virtgpu-utils.h @@ -0,0 +1,86 @@ +#pragma once + +#include +#include +#include +#include +#include +#include +#include +#include +#include + +#define unlikely(x) __builtin_expect(!!(x), 0) +#define likely(x) __builtin_expect(!!(x), 1) + +#ifndef UNUSED +# define UNUSED(x) (void) (x) +#endif + +/** Checks is a value is a power of two. Does not handle zero. */ +#define IS_POT(v) (((v) & ((v) - 1)) == 0) + +/** Checks is a value is a power of two. Zero handled. */ +#define IS_POT_NONZERO(v) ((v) != 0 && IS_POT(v)) + +/** Align a value to a power of two */ +#define ALIGN_POT(x, pot_align) (((x) + (pot_align) - 1) & ~((pot_align) - 1)) + +#define p_atomic_read(_v) __atomic_load_n((_v), __ATOMIC_ACQUIRE) + +static inline bool util_is_power_of_two_nonzero64(uint64_t v) { + return IS_POT_NONZERO(v); +} + +static inline uint64_t align64(uint64_t value, uint64_t alignment) { + assert(util_is_power_of_two_nonzero64(alignment)); + return ALIGN_POT(value, alignment); +} + +struct list_head { + list_head * prev; + list_head * next; +}; + +struct util_sparse_array { + size_t elem_size; + unsigned node_size_log2; + + uintptr_t root; +}; + +void * util_sparse_array_get(util_sparse_array * arr, uint64_t idx); +void util_sparse_array_init(util_sparse_array * arr, size_t elem_size, size_t node_size); + +inline void os_time_sleep(int64_t usecs) { + timespec time; + time.tv_sec = usecs / 1000000; + time.tv_nsec = (usecs % 1000000) * 1000; + while (clock_nanosleep(CLOCK_MONOTONIC, 0, &time, &time) == EINTR) + ; +} + +struct timer_data { + long long start; + long long total; + long long count; +}; + +static inline void start_timer(timer_data * timer) { + timespec ts; + clock_gettime(CLOCK_MONOTONIC, &ts); + timer->start = (long long) ts.tv_sec * 1000000000LL + ts.tv_nsec; +} + +// returns the duration in ns +static inline long long stop_timer(timer_data * timer) { + timespec ts; + clock_gettime(CLOCK_MONOTONIC, &ts); + long long timer_end = (long long) ts.tv_sec * 1000000000LL + ts.tv_nsec; + + long long duration = (timer_end - timer->start); + timer->total += duration; + timer->count += 1; + + return duration; +} diff --git a/ggml/src/ggml-virtgpu/virtgpu.cpp b/ggml/src/ggml-virtgpu/virtgpu.cpp new file mode 100644 index 0000000000..005c8e21db --- /dev/null +++ b/ggml/src/ggml-virtgpu/virtgpu.cpp @@ -0,0 +1,498 @@ +#include "virtgpu.h" + +#include +#include + +#include +#include +#include + +static virt_gpu_result_t virtgpu_open_device(virtgpu * gpu, const drmDevicePtr dev); +static virt_gpu_result_t virtgpu_open(virtgpu * gpu); + +static virt_gpu_result_t virtgpu_init_capset(virtgpu * gpu); +static virt_gpu_result_t virtgpu_init_context(virtgpu * gpu); + +static int virtgpu_ioctl_context_init(virtgpu * gpu, virgl_renderer_capset capset_id); +static int virtgpu_ioctl_get_caps(virtgpu * gpu, + virgl_renderer_capset id, + uint32_t version, + void * capset, + size_t capset_size); +static uint64_t virtgpu_ioctl_getparam(virtgpu * gpu, uint64_t param); +static void virtgpu_init_renderer_info(virtgpu * gpu); + +static void log_call_duration(long long call_duration_ns, const char * name); + +const uint64_t APIR_HANDSHAKE_MAX_WAIT_MS = 2 * 1000; // 2s +const uint64_t APIR_LOADLIBRARY_MAX_WAIT_MS = 60 * 1000; // 60s + +static int virtgpu_handshake(virtgpu * gpu) { + apir_encoder * encoder; + apir_decoder * decoder; + + encoder = remote_call_prepare(gpu, APIR_COMMAND_TYPE_HANDSHAKE, 0); + if (!encoder) { + GGML_ABORT("%s: failed to prepare the remote call encoder", __func__); + return 1; + } + + /* write handshake props */ + + uint32_t guest_major = APIR_PROTOCOL_MAJOR; + uint32_t guest_minor = APIR_PROTOCOL_MINOR; + apir_encode_uint32_t(encoder, &guest_major); + apir_encode_uint32_t(encoder, &guest_minor); + + /* *** */ + + uint32_t ret_magic; + long long call_duration_ns; + ret_magic = remote_call(gpu, encoder, &decoder, APIR_HANDSHAKE_MAX_WAIT_MS, &call_duration_ns); + log_call_duration(call_duration_ns, "API Remoting handshake"); + + if (!decoder) { + GGML_ABORT( + "%s: failed to initiate the communication with the virglrenderer library. " + "Most likely, the wrong virglrenderer library was loaded in the hypervisor.", + __func__); + return 1; + } + + /* read handshake return values */ + + uint32_t host_major; + uint32_t host_minor; + + if (ret_magic != APIR_HANDSHAKE_MAGIC) { + GGML_ABORT("%s: handshake with the virglrenderer failed (code=%d | %s)", __func__, ret_magic, + apir_backend_initialize_error(ret_magic)); + } else { + apir_decode_uint32_t(decoder, &host_major); + apir_decode_uint32_t(decoder, &host_minor); + } + + remote_call_finish(gpu, encoder, decoder); + + if (ret_magic != APIR_HANDSHAKE_MAGIC) { + return 1; + } + + GGML_LOG_INFO("%s: Guest is running with %u.%u\n", __func__, guest_major, guest_minor); + GGML_LOG_INFO("%s: Host is running with %u.%u\n", __func__, host_major, host_minor); + + if (guest_major != host_major) { + GGML_LOG_ERROR("Host major (%d) and guest major (%d) version differ\n", host_major, guest_major); + } else if (guest_minor != host_minor) { + GGML_LOG_WARN("Host minor (%d) and guest minor (%d) version differ\n", host_minor, guest_minor); + } + + return 0; +} + +static ApirLoadLibraryReturnCode virtgpu_load_library(virtgpu * gpu) { + apir_encoder * encoder; + apir_decoder * decoder; + ApirLoadLibraryReturnCode ret; + + encoder = remote_call_prepare(gpu, APIR_COMMAND_TYPE_LOADLIBRARY, 0); + if (!encoder) { + GGML_ABORT("%s: hypercall error: failed to prepare the remote call encoder", __func__); + return APIR_LOAD_LIBRARY_HYPERCALL_INITIALIZATION_ERROR; + } + + long long call_duration_ns; + + ret = (ApirLoadLibraryReturnCode) remote_call(gpu, encoder, &decoder, APIR_LOADLIBRARY_MAX_WAIT_MS, + &call_duration_ns); + log_call_duration(call_duration_ns, "API Remoting LoadLibrary"); + + if (!decoder) { + GGML_ABORT("%s: hypercall error: failed to kick the API remoting hypercall.\n", __func__); + return APIR_LOAD_LIBRARY_HYPERCALL_INITIALIZATION_ERROR; + } + + remote_call_finish(gpu, encoder, decoder); + + if (ret == APIR_LOAD_LIBRARY_SUCCESS) { + GGML_LOG_INFO("%s: The API Remoting backend was successfully loaded and initialized\n", __func__); + + return ret; + } + + // something wrong happened, find out what. + + if (ret < APIR_LOAD_LIBRARY_INIT_BASE_INDEX) { + GGML_ABORT("%s: virglrenderer could not load the API Remoting backend library: %s (code %d)", __func__, + apir_load_library_error(ret), ret); + return ret; + } + + GGML_LOG_INFO("%s: virglrenderer successfully loaded the API Remoting backend library", __func__); + + ApirLoadLibraryReturnCode apir_ret = (ApirLoadLibraryReturnCode) (ret - APIR_LOAD_LIBRARY_INIT_BASE_INDEX); + + if (apir_ret < APIR_LOAD_LIBRARY_INIT_BASE_INDEX) { + GGML_ABORT("%s: the API Remoting backend library couldn't load the backend library: apir code=%d | %s)", + __func__, apir_ret, apir_load_library_error(apir_ret)); + } else { + uint32_t lib_ret = apir_ret - APIR_LOAD_LIBRARY_INIT_BASE_INDEX; + GGML_ABORT("%s: the API Remoting backend library initialize its backend library: apir code=%d)", __func__, + lib_ret); + } + return ret; +} + +virtgpu * create_virtgpu() { + virtgpu * gpu = new virtgpu(); + + gpu->use_apir_capset = getenv("GGML_REMOTING_USE_APIR_CAPSET") != nullptr; + util_sparse_array_init(&gpu->shmem_array, sizeof(virtgpu_shmem), 1024); + + if (virtgpu_open(gpu) != APIR_SUCCESS) { + GGML_ABORT("%s: failed to open the virtgpu device", __func__); + return NULL; + } + + if (virtgpu_init_capset(gpu) != APIR_SUCCESS) { + GGML_ABORT("%s: failed to initialize the GPU capset", __func__); + return NULL; + } + + if (virtgpu_init_context(gpu) != APIR_SUCCESS) { + GGML_ABORT("%s: failed to initialize the GPU context", __func__); + return NULL; + } + + if (virtgpu_shmem_create(gpu, SHMEM_REPLY_SIZE, &gpu->reply_shmem)) { + GGML_ABORT("%s: failed to create the shared reply memory pages", __func__); + return NULL; + } + + if (virtgpu_shmem_create(gpu, SHMEM_DATA_SIZE, &gpu->data_shmem)) { + GGML_ABORT("%s: failed to create the shared data memory pages", __func__); + return NULL; + } + + if (virtgpu_handshake(gpu)) { + GGML_ABORT("%s: failed to handshake with the virglrenderer library", __func__); + return NULL; + } + + if (virtgpu_load_library(gpu) != APIR_LOAD_LIBRARY_SUCCESS) { + GGML_ABORT("%s: failed to load the backend library", __func__); + return NULL; + } + + return gpu; +} + +static virt_gpu_result_t virtgpu_open(virtgpu * gpu) { + drmDevicePtr devs[8]; + int count = drmGetDevices2(0, devs, ARRAY_SIZE(devs)); + if (count < 0) { + GGML_LOG_ERROR("%s: failed to enumerate DRM devices\n", __func__); + return APIR_ERROR_INITIALIZATION_FAILED; + } + + virt_gpu_result_t result = APIR_ERROR_INITIALIZATION_FAILED; + for (int i = 0; i < count; i++) { + result = virtgpu_open_device(gpu, devs[i]); + if (result == APIR_SUCCESS) { + break; + } + } + + drmFreeDevices(devs, count); + + return result; +} + +static virt_gpu_result_t virtgpu_open_device(virtgpu * gpu, const drmDevicePtr dev) { + const char * node_path = dev->nodes[DRM_NODE_RENDER]; + + int fd = open(node_path, O_RDWR | O_CLOEXEC); + if (fd < 0) { + GGML_ABORT("failed to open %s", node_path); + return APIR_ERROR_INITIALIZATION_FAILED; + } + + drmVersionPtr version = drmGetVersion(fd); + if (!version || strcmp(version->name, "virtio_gpu") || version->version_major != 0) { + if (version) { + GGML_ABORT("unknown DRM driver %s version %d", version->name, version->version_major); + } else { + GGML_ABORT("failed to get DRM driver version"); + } + + if (version) { + drmFreeVersion(version); + } + close(fd); + return APIR_ERROR_INITIALIZATION_FAILED; + } + + gpu->fd = fd; + + drmFreeVersion(version); + + GGML_LOG_INFO("using DRM device %s\n", node_path); + + return APIR_SUCCESS; +} + +static virt_gpu_result_t virtgpu_init_context(virtgpu * gpu) { + assert(!gpu->capset.version); + const int ret = virtgpu_ioctl_context_init(gpu, gpu->capset.id); + if (ret) { + GGML_LOG_INFO("failed to initialize context: %s\n", strerror(errno)); + return APIR_ERROR_INITIALIZATION_FAILED; + } + + return APIR_SUCCESS; +} + +static virt_gpu_result_t virtgpu_init_capset(virtgpu * gpu) { + if (gpu->use_apir_capset) { + GGML_LOG_INFO("Using the APIR capset\n"); + gpu->capset.id = VIRTGPU_DRM_CAPSET_APIR; + } else { + GGML_LOG_INFO("Using the Venus capset\n"); + gpu->capset.id = VIRTGPU_DRM_CAPSET_VENUS; + } + gpu->capset.version = 0; + + int ret = + virtgpu_ioctl_get_caps(gpu, gpu->capset.id, gpu->capset.version, &gpu->capset.data, sizeof(gpu->capset.data)); + + if (ret) { + GGML_LOG_INFO("failed to get APIR v%d capset: %s\n", gpu->capset.version, strerror(errno)); + return APIR_ERROR_INITIALIZATION_FAILED; + } + + assert(gpu->capset.data.supports_blob_resources); + + return APIR_SUCCESS; +} + +static int virtgpu_ioctl_context_init(virtgpu * gpu, virgl_renderer_capset capset_id) { + drm_virtgpu_context_set_param ctx_set_params[3] = { + { + .param = VIRTGPU_CONTEXT_PARAM_CAPSET_ID, + .value = capset_id, + }, + { + .param = VIRTGPU_CONTEXT_PARAM_NUM_RINGS, + .value = 1, + }, + { + .param = VIRTGPU_CONTEXT_PARAM_POLL_RINGS_MASK, + .value = 0, /* don't generate drm_events on fence signaling */ + }, + }; + + drm_virtgpu_context_init args = { + .num_params = ARRAY_SIZE(ctx_set_params), + .pad = 0, + .ctx_set_params = (uintptr_t) &ctx_set_params, + }; + + return virtgpu_ioctl(gpu, DRM_IOCTL_VIRTGPU_CONTEXT_INIT, &args); +} + +static int virtgpu_ioctl_get_caps(virtgpu * gpu, + virgl_renderer_capset id, + uint32_t version, + void * capset, + size_t capset_size) { + drm_virtgpu_get_caps args = { + .cap_set_id = id, + .cap_set_ver = version, + .addr = (uintptr_t) capset, + .size = (__u32) capset_size, + .pad = 0, + }; + + return virtgpu_ioctl(gpu, DRM_IOCTL_VIRTGPU_GET_CAPS, &args); +} + +static uint64_t virtgpu_ioctl_getparam(virtgpu * gpu, uint64_t param) { + /* val must be zeroed because kernel only writes the lower 32 bits */ + uint64_t val = 0; + drm_virtgpu_getparam args = { + .param = param, + .value = (uintptr_t) &val, + }; + + const int ret = virtgpu_ioctl(gpu, DRM_IOCTL_VIRTGPU_GETPARAM, &args); + return ret ? 0 : val; +} + +apir_encoder * remote_call_prepare(virtgpu * gpu, ApirCommandType apir_cmd_type, int32_t cmd_flags) { + /* + * Prepare the command encoder and its buffer + */ + + static char encoder_buffer[4096]; + + static apir_encoder enc; + enc = { + .cur = encoder_buffer, + .start = encoder_buffer, + .end = encoder_buffer + sizeof(encoder_buffer), + .fatal = false, + }; + + /* + * Fill the command encoder with the common args: + * - cmd_type (int32_t) + * - cmd_flags (int32_t) + * - reply res id (uint32_t) + */ + + int32_t cmd_type = apir_cmd_type; + + // for testing during the hypervisor transition + if (!gpu->use_apir_capset) { + cmd_type += VENUS_COMMAND_TYPE_LENGTH; + } + apir_encode_int32_t(&enc, &cmd_type); + apir_encode_int32_t(&enc, &cmd_flags); + + uint32_t reply_res_id = gpu->reply_shmem.res_id; + apir_encode_uint32_t(&enc, &reply_res_id); + + return &enc; +} + +void remote_call_finish(virtgpu * gpu, apir_encoder * enc, apir_decoder * dec) { + UNUSED(gpu); + + if (!enc) { + GGML_LOG_ERROR("Invalid (null) encoder\n"); + } + + if (!dec) { + GGML_LOG_ERROR("Invalid (null) decoder\n"); + } + + if (apir_encoder_get_fatal(enc)) { + GGML_LOG_ERROR("Failed to encode the output parameters.\n"); + } + + if (apir_decoder_get_fatal(dec)) { + GGML_LOG_ERROR("Failed to decode the input parameters.\n"); + } +} + +uint32_t remote_call(virtgpu * gpu, + apir_encoder * encoder, + apir_decoder ** decoder, + float max_wait_ms, + long long * call_duration_ns) { + /* + * Prepare the reply notification pointer + */ + + volatile std::atomic_uint * atomic_reply_notif = (volatile std::atomic_uint *) gpu->reply_shmem.mmap_ptr; + *atomic_reply_notif = 0; + + /* + * Trigger the execbuf ioctl + */ + + drm_virtgpu_execbuffer args = { + .flags = VIRTGPU_EXECBUF_RING_IDX, + .size = (uint32_t) (encoder->cur - encoder->start), + .command = (uintptr_t) encoder->start, + + .bo_handles = 0, + .num_bo_handles = 0, + + .fence_fd = 0, + .ring_idx = 0, + .syncobj_stride = 0, + .num_in_syncobjs = 0, + .num_out_syncobjs = 0, + .in_syncobjs = 0, + .out_syncobjs = 0, + }; + + *decoder = NULL; + + int ret = drmIoctl(gpu->fd, DRM_IOCTL_VIRTGPU_EXECBUFFER, &args); + + if (ret != 0) { + GGML_ABORT("%s: the virtgpu EXECBUFFER ioctl failed (%d)", __func__, ret); + } + + /* + * Wait for the response notification + */ + timer_data wait_host_reply_timer = { 0, 0, 0 }; + + start_timer(&wait_host_reply_timer); + + timespec ts_start, ts_end; + clock_gettime(CLOCK_MONOTONIC, &ts_start); + long long start_time = (long long) ts_start.tv_sec * 1000000000LL + ts_start.tv_nsec; + + bool timedout = false; + uint32_t notif_value = 0; + while (true) { + notif_value = std::atomic_load_explicit(atomic_reply_notif, std::memory_order_acquire); + + if (notif_value != 0) { + break; + } + + int64_t base_sleep_us = 15; + + os_time_sleep(base_sleep_us); + + if (max_wait_ms) { + clock_gettime(CLOCK_MONOTONIC, &ts_end); + long long end_time = (long long) ts_end.tv_sec * 1000000000LL + ts_end.tv_nsec; + float duration_ms = (end_time - start_time) / 1000000; + + if (duration_ms > max_wait_ms) { + timedout = true; + break; + } + } + } + + if (call_duration_ns) { + *call_duration_ns = stop_timer(&wait_host_reply_timer); + } + + if (max_wait_ms && timedout) { + GGML_LOG_ERROR("timed out waiting for the host answer...\n"); + return APIR_FORWARD_TIMEOUT; + } + + /* + * Prepare the decoder + */ + static apir_decoder response_dec; + response_dec.cur = (char *) gpu->reply_shmem.mmap_ptr + sizeof(*atomic_reply_notif); + response_dec.end = (char *) gpu->reply_shmem.mmap_ptr + gpu->reply_shmem.mmap_size; + *decoder = &response_dec; + + // extract the actual return value from the notif flag + uint32_t returned_value = notif_value - 1; + return returned_value; +} + +static void log_call_duration(long long call_duration_ns, const char * name) { + double call_duration_ms = (double) call_duration_ns / 1e6; // 1 millisecond = 1e6 nanoseconds + double call_duration_s = (double) call_duration_ns / 1e9; // 1 second = 1e9 nanoseconds + + if (call_duration_s > 1) { + GGML_LOG_INFO("%s: waited %.2fs for the %s host reply...\n", __func__, call_duration_s, name); + } else if (call_duration_ms > 1) { + GGML_LOG_INFO("%s: waited %.2fms for the %s host reply...\n", __func__, call_duration_ms, name); + } else { + GGML_LOG_INFO("%s: waited %lldns for the %s host reply...\n", __func__, call_duration_ns, name); + } +} diff --git a/ggml/src/ggml-virtgpu/virtgpu.h b/ggml/src/ggml-virtgpu/virtgpu.h new file mode 100644 index 0000000000..d4bb42e20b --- /dev/null +++ b/ggml/src/ggml-virtgpu/virtgpu.h @@ -0,0 +1,92 @@ +#pragma once + +#include "virtgpu-utils.h" +#include "virtgpu-shm.h" +#include "virtgpu-apir.h" + +#include "backend/shared/api_remoting.h" +#include "backend/shared/apir_cs.h" + +#include +#include +#include +#include +#include +#include +#include + +#include + +#define VIRGL_RENDERER_UNSTABLE_APIS 1 +#include "apir_hw.h" +#include +#include "venus_hw.h" + +#ifndef VIRTGPU_DRM_CAPSET_APIR +// Will be defined include/drm/virtgpu_drm.h when +// https://gitlab.freedesktop.org/virgl/virglrenderer/-/merge_requests/1590/diffs +// is merged +#define VIRTGPU_DRM_CAPSET_APIR 10 +#endif + +// Mesa/Virlgrenderer Venus internal. Only necessary during the +// Venus->APIR transition in Virglrenderer +#define VENUS_COMMAND_TYPE_LENGTH 331 + +#ifndef VIRTGPU_DRM_CAPSET_VENUS // only available with Linux >= v6.16 +#define VIRTGPU_DRM_CAPSET_VENUS 4 +#endif + +typedef uint32_t virgl_renderer_capset; + +/* from src/virtio/vulkan/vn_renderer_virtgpu.c */ +#define VIRTGPU_PCI_VENDOR_ID 0x1af4 +#define VIRTGPU_PCI_DEVICE_ID 0x1050 +#define VIRTGPU_BLOB_MEM_GUEST_VRAM 0x0004 +#define VIRTGPU_PARAM_GUEST_VRAM 9 + +#define SHMEM_DATA_SIZE 0x1830000 // 24MiB +#define SHMEM_REPLY_SIZE 0x4000 + +#define ARRAY_SIZE(x) (sizeof(x) / sizeof((x)[0])) + +enum virt_gpu_result_t { + APIR_SUCCESS = 0, + APIR_ERROR_INITIALIZATION_FAILED = -1, +}; + +#define PRINTFLIKE(f, a) __attribute__((format(__printf__, f, a))) + +struct virtgpu { + bool use_apir_capset; + + int fd; + + struct { + virgl_renderer_capset id; + uint32_t version; + virgl_renderer_capset_apir data; + } capset; + + util_sparse_array shmem_array; + + /* APIR communication pages */ + virtgpu_shmem reply_shmem; + virtgpu_shmem data_shmem; +}; + +static inline int virtgpu_ioctl(virtgpu * gpu, unsigned long request, void * args) { + return drmIoctl(gpu->fd, request, args); +} + +virtgpu * create_virtgpu(); + +apir_encoder * remote_call_prepare(virtgpu * gpu, ApirCommandType apir_cmd_type, int32_t cmd_flags); + +uint32_t remote_call(virtgpu * gpu, + apir_encoder * enc, + apir_decoder ** dec, + float max_wait_ms, + long long * call_duration_ns); + +void remote_call_finish(virtgpu * gpu, apir_encoder * enc, apir_decoder * dec); diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp index b5e5dba95f..3852867c29 100644 --- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp +++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp @@ -3162,17 +3162,31 @@ static void ggml_vk_load_shaders(vk_device& device) { // For scalar, use 128 (arbitrary) // The same D_split value is used for both HSK and HSV, so just base it on the union of the LSBs. const uint32_t D = (hsk|hsv); - uint32_t wg_size = (path == FA_SCALAR || path == FA_COOPMAT1) - ? scalar_flash_attention_workgroup_size - : ((small_rows && (D % 32) == 0) ? 256 : 128); auto rows_cols = fa_rows_cols(path, hsk, hsv, clamp, type, small_rows, small_cache); + uint32_t wg_size; + switch (path) { + case FA_COOPMAT2: + wg_size = ((small_rows && (D % 32) == 0) ? 256 : 128); + break; + case FA_COOPMAT1: + wg_size = (rows_cols[1] / 16) * device->subgroup_size; // enough subgroups for Bc/MatBc + break; + default: + wg_size = scalar_flash_attention_workgroup_size; + break; + } + // D_split can't be larger than a subgroup because we use subgroupShuffle to reduce it. // D_split can't be larger than the LSB of D divided by 4 due to vectorization in the shader. const uint32_t D_lsb = D ^ (D & (D-1)); uint32_t D_split = std::min(std::min(device->subgroup_size, 8u), D_lsb / 4); - return {wg_size, rows_cols[0], rows_cols[1], hsk, hsv, clamp, D_split}; + // Nvidia prefers shared memory use to load large tiles of K + // AMD prefers loading K directly from global memory + const uint32_t k_load_shmem = device->vendor_id == VK_VENDOR_ID_NVIDIA ? 1 : 0; + + return {wg_size, rows_cols[0], rows_cols[1], hsk, hsv, clamp, D_split, device->subgroup_size, k_load_shmem}; }; #define CREATE_FA(TYPE, NAMELC, FAPATH, SUFFIX) \ @@ -3187,15 +3201,15 @@ static void ggml_vk_load_shaders(vk_device& device) { if (path == FAPATH) { \ if (aligned) { \ if (f32acc) { \ - ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_aligned_f32acc" #NAMELC, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _data, "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,0,TYPE,small_rows,small_cache), fa_spec_constants(FAPATH, HSK,HSV,0,TYPE,small_rows,small_cache), fa_align(FAPATH,HSK,HSV,TYPE,small_rows,small_cache), true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0)); \ + ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_aligned_f32acc" #NAMELC, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _data, "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,0,TYPE,small_rows,small_cache), fa_spec_constants(FAPATH, HSK,HSV,0,TYPE,small_rows,small_cache), fa_align(FAPATH,HSK,HSV,TYPE,small_rows,small_cache), true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? device->subgroup_size : 0)); \ } else { \ - ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_aligned_f16acc" #NAMELC, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data, "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,0,TYPE,small_rows,small_cache), fa_spec_constants(FAPATH, HSK,HSV,0,TYPE,small_rows,small_cache), fa_align(FAPATH,HSK,HSV,TYPE,small_rows,small_cache), true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0)); \ + ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_aligned_f16acc" #NAMELC, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data, "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,0,TYPE,small_rows,small_cache), fa_spec_constants(FAPATH, HSK,HSV,0,TYPE,small_rows,small_cache), fa_align(FAPATH,HSK,HSV,TYPE,small_rows,small_cache), true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? device->subgroup_size : 0)); \ } \ } else { \ if (f32acc) { \ - ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_f32acc" #NAMELC, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _data, "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,1,TYPE,small_rows,small_cache), fa_spec_constants(FAPATH, HSK,HSV,1,TYPE,small_rows,small_cache), 1, true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0)); \ + ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_f32acc" #NAMELC, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _data, "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,1,TYPE,small_rows,small_cache), fa_spec_constants(FAPATH, HSK,HSV,1,TYPE,small_rows,small_cache), 1, true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? device->subgroup_size : 0)); \ } else { \ - ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_f16acc" #NAMELC, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data, "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,1,TYPE,small_rows,small_cache), fa_spec_constants(FAPATH, HSK,HSV,1,TYPE,small_rows,small_cache), 1, true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0)); \ + ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_f16acc" #NAMELC, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data, "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,1,TYPE,small_rows,small_cache), fa_spec_constants(FAPATH, HSK,HSV,1,TYPE,small_rows,small_cache), 1, true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? device->subgroup_size : 0)); \ } \ } \ } \ @@ -5522,22 +5536,32 @@ static void ggml_vk_instance_init() { if ((new_props.properties.deviceType == vk::PhysicalDeviceType::eDiscreteGpu || new_props.properties.deviceType == vk::PhysicalDeviceType::eIntegratedGpu) && ggml_vk_device_is_supported(devices[i])) { // Check if there are two physical devices corresponding to the same GPU + // This handles the case where the same GPU appears with different drivers (e.g., RADV + AMDVLK on Linux), + // see https://github.com/ggml-org/llama.cpp/pull/7582 for original deduplication. + // However, for MoltenVK on macOS, multiple GPUs on the same card may report the same UUID, + // see https://github.com/KhronosGroup/MoltenVK/issues/2683. Until this is fixed, we'll only deduplicate + // when drivers differ (same driver + same UUID = likely different GPUs) auto old_device = std::find_if( vk_instance.device_indices.begin(), vk_instance.device_indices.end(), - [&devices, &new_id](const size_t k){ + [&devices, &new_id, &new_driver](const size_t k){ vk::PhysicalDeviceProperties2 old_props; + vk::PhysicalDeviceDriverProperties old_driver; vk::PhysicalDeviceIDProperties old_id; - old_props.pNext = &old_id; + old_props.pNext = &old_driver; + old_driver.pNext = &old_id; devices[k].getProperties2(&old_props); - bool equals = std::equal(std::begin(old_id.deviceUUID), std::end(old_id.deviceUUID), std::begin(new_id.deviceUUID)); - equals = equals || ( + bool same_uuid = std::equal(std::begin(old_id.deviceUUID), std::end(old_id.deviceUUID), std::begin(new_id.deviceUUID)); + same_uuid = same_uuid || ( old_id.deviceLUIDValid && new_id.deviceLUIDValid && std::equal(std::begin(old_id.deviceLUID), std::end(old_id.deviceLUID), std::begin(new_id.deviceLUID)) ); - return equals; + // Only deduplicate if same UUID AND different drivers + // (same driver + same UUID on MoltenVK = likely different GPUs on multi-GPU card) + bool different_driver = (old_driver.driverID != new_driver.driverID); + return same_uuid && different_driver; } ); if (old_device == vk_instance.device_indices.end()) { @@ -8334,41 +8358,49 @@ static bool ggml_vk_flash_attn_scalar_shmem_support(const vk_device& device, con const uint32_t total_size = tmpsh + tmpshv4 + masksh + Qf; const bool supported = total_size <= device->properties.limits.maxComputeSharedMemorySize; - VK_LOG_DEBUG("ggml_vk_flash_attn_coopmat_shmem_support(HSK=" << hsk << ", HSV=" << hsv << ", total_size=" << total_size << ", supported=" << supported); + VK_LOG_DEBUG("ggml_vk_flash_attn_scalar_shmem_support(HSK=" << hsk << ", HSV=" << hsv << ", total_size=" << total_size << ", supported=" << supported); return supported; } -static bool ggml_vk_flash_attn_coopmat_shmem_support(const vk_device& device, const uint32_t hsk, uint32_t hsv, bool f32acc) { +static bool ggml_vk_flash_attn_coopmat_shmem_support(const vk_device& device, const uint32_t hsk, uint32_t hsv, bool f32acc, ggml_type kv_type) { // Needs to be kept up to date on shader changes GGML_UNUSED(hsv); - const uint32_t wg_size = scalar_flash_attention_workgroup_size; - const uint32_t Br = coopmat1_flash_attention_num_large_rows; - const uint32_t Bc = scalar_flash_attention_Bc; + const auto rows_cols = fa_rows_cols(FA_COOPMAT1, hsk, hsv, 0, kv_type, false, false); + const uint32_t Br = rows_cols[0]; + const uint32_t Bc = rows_cols[1]; + + const uint32_t MatBr = 16, MatBc = 16; + + const uint32_t row_split = Bc / MatBc; const uint32_t hsk_pad = ROUNDUP_POW2(hsk, 16); const uint32_t acctype = f32acc ? 4 : 2; const uint32_t f16vec4 = 8; - const uint32_t tmpsh = wg_size * sizeof(float); - const uint32_t tmpshv4 = wg_size * 4 * acctype; + const uint32_t tmpsh = (Bc / MatBc) * sizeof(float); const uint32_t qstride = hsk_pad / 4 + 2; const uint32_t Qf = Br * qstride * f16vec4; + const uint32_t psh_stride = Br / 4 + 2; + const uint32_t Psh = Bc * psh_stride * f16vec4; + const uint32_t sfshstride = (hsk <= 128) ? (Br + 8) : Br; const uint32_t sfsh = Bc * sfshstride * acctype; - const uint32_t kshstride = hsk_pad / 4 + 2; - const uint32_t ksh = Bc * kshstride * f16vec4; + const bool k_load_shmem = device->vendor_id == VK_VENDOR_ID_NVIDIA; + const uint32_t kshstride = (k_load_shmem ? hsk_pad : MatBr) / 4 + 2; + const uint32_t vsh_stride = MatBc / 4 * row_split; + const uint32_t ksh = ((kshstride >= vsh_stride) ? (Bc * kshstride) : (Bc * vsh_stride)) * f16vec4; - const uint32_t slope = Br * sizeof(float); + const uint32_t slope = Br * acctype; - const uint32_t total_size = tmpsh + tmpshv4 + Qf + sfsh + ksh + slope; + const uint32_t total_size = tmpsh + Qf + Psh + sfsh + ksh + slope; const bool supported = total_size <= device->properties.limits.maxComputeSharedMemorySize; - VK_LOG_DEBUG("ggml_vk_flash_attn_coopmat_shmem_support(HSK=" << hsk << ", HSV=" << hsv << ", f32acc=" << f32acc << ", total_size=" << total_size << ", supported=" << supported); + VK_LOG_DEBUG("ggml_vk_flash_attn_coopmat_shmem_support(HSK=" << hsk << ", HSV=" << hsv << ", f32acc=" << f32acc << ", kv_type=" << kv_type << ", total_size=" << total_size << ", supported=" << supported); return supported; } @@ -8432,7 +8464,7 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx const bool coopmat_shape_supported = (dst->op_params[3] == GGML_PREC_F32 && ctx->device->coopmat_support_16x16x16_f32acc) || (dst->op_params[3] != GGML_PREC_F32 && ctx->device->coopmat_support_16x16x16_f16acc); - const bool coopmat_shmem_supported = ggml_vk_flash_attn_coopmat_shmem_support(ctx->device, HSK, HSV, dst->op_params[3] == GGML_PREC_F32); + const bool coopmat_shmem_supported = ggml_vk_flash_attn_coopmat_shmem_support(ctx->device, HSK, HSV, dst->op_params[3] == GGML_PREC_F32, k->type); if (!coopmat_shape_supported || !coopmat_shmem_supported) { path = FA_SCALAR; diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.glsl b/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.glsl index 29b5c7c3a4..23a4d2c005 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.glsl +++ b/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.glsl @@ -8,6 +8,8 @@ layout (constant_id = 3) const uint32_t HSK = 32; layout (constant_id = 4) const uint32_t HSV = 32; layout (constant_id = 5) const uint32_t Clamp = 0; layout (constant_id = 6) const uint32_t D_split = 16; +layout (constant_id = 7) const uint32_t SubGroupSize = 32; +layout (constant_id = 8) const uint32_t K_LOAD_SHMEM = 0; // Round up head sizes to a multiple of 16, for coopmat1/coopmat2 paths const uint32_t HSK_pad = (HSK + 15) & ~15; @@ -74,6 +76,10 @@ layout (binding = 1) readonly buffer K_PACKED16 {A_TYPE_PACKED16 k_data_packed16 layout (binding = 2) readonly buffer V_PACKED16 {A_TYPE_PACKED16 v_data_packed16[];} v_packed; #endif +#ifndef BLOCK_SIZE +#define BLOCK_SIZE 1 +#endif + #if defined(DATA_A_F32) #undef BLOCK_SIZE #define BLOCK_SIZE 4 diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp b/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp index 0eb50fe58f..83d52d19d6 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp @@ -7,6 +7,7 @@ #extension GL_EXT_shader_explicit_arithmetic_types_int32 : require #extension GL_KHR_shader_subgroup_basic : enable +#extension GL_KHR_shader_subgroup_arithmetic : enable #extension GL_KHR_shader_subgroup_vote : enable #extension GL_KHR_memory_scope_semantics : enable #extension GL_KHR_cooperative_matrix : enable @@ -14,12 +15,13 @@ #include "types.glsl" #include "flash_attn_base.glsl" -const uint32_t HSK_per_thread = HSK / D_split; -const uint32_t HSV_per_thread = HSV / D_split; +// These need to be supported N,M values for a MatBc x MatBr x 16 coopmatmuladd +const uint32_t MatBr = 16; +const uint32_t MatBc = 16; -const uint32_t row_split = 4; +const uint32_t row_split = Bc / MatBc; const uint32_t rows_per_thread = Br / row_split; -const uint32_t cols_per_iter = gl_WorkGroupSize.x / D_split / row_split; +const uint32_t cols_per_iter = gl_WorkGroupSize.x / row_split; const uint32_t cols_per_thread = Bc / cols_per_iter; @@ -40,24 +42,24 @@ D_TYPE perElemOpGqaStore(const in uint32_t r, const in uint32_t c, const in D_TY return elem; } -// These need to be supported N,M values for a MatBc x MatBr x 16 coopmatmuladd -const uint32_t MatBr = 16; -const uint32_t MatBc = 16; - -shared FLOAT_TYPE tmpsh[gl_WorkGroupSize.x]; -shared ACC_TYPEV4 tmpshv4[gl_WorkGroupSize.x]; +shared float tmpsh[row_split]; const uint32_t qstride = HSK_pad / 4 + 2; // in units of f16vec4 shared f16vec4 Qf[Br * qstride]; +const uint psh_stride = Br / 4 + 2; +shared f16vec4 Psh[Bc * psh_stride]; + // Avoid padding for hsk==256 to make it fit in 48KB shmem. -const uint32_t sfshstride = (HSK <= 128) ? (Br + 8) : Br; -shared ACC_TYPE sfsh[Bc * sfshstride]; +const uint32_t sfshstride = (HSK <= 128) ? (Br / 4 + 2) : Br / 4; +shared ACC_TYPEV4 sfsh[Bc * sfshstride]; -const uint32_t kshstride = HSK_pad / 4 + 2; // in units of f16vec4 -shared f16vec4 ksh[Bc * kshstride]; +const uint32_t kshstride = (K_LOAD_SHMEM != 0 ? HSK_pad : MatBr) / 4 + 2; // in units of f16vec4 +const uint v_cols = MatBc / 4 * row_split; // total cols, 4 vec4s per MatBc * number of subgroups +const uint vsh_stride = v_cols; +shared f16vec4 ksh[(kshstride >= vsh_stride) ? (Bc * kshstride) : (Bc * vsh_stride)]; -shared float slope[Br]; +shared ACC_TYPE slope[Br]; void main() { #ifdef NEEDS_INIT_IQ_SHMEM @@ -69,9 +71,9 @@ void main() { const uint32_t tid = gl_LocalInvocationIndex; const uint32_t threads_per_rowgroup = gl_WorkGroupSize.x / row_split; + const uint32_t d_per_thread = (HSV/4 + threads_per_rowgroup - 1) / threads_per_rowgroup; const uint32_t row_tid = gl_LocalInvocationIndex / threads_per_rowgroup; - const uint32_t d_tid = gl_LocalInvocationIndex % D_split; - const uint32_t col_tid = (gl_LocalInvocationIndex % threads_per_rowgroup) / D_split; + const uint32_t col_tid = gl_LocalInvocationIndex % threads_per_rowgroup; #define tile_row(r) (row_tid * rows_per_thread + (r)) @@ -102,9 +104,9 @@ void main() { } barrier(); - ACC_TYPEV4 Of[rows_per_thread][HSV_per_thread / 4]; - [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) { - [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) { + ACC_TYPEV4 Of[rows_per_thread][d_per_thread]; + [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) { + [[unroll]] for (uint32_t d = 0; d < d_per_thread; ++d) { Of[r][d] = ACC_TYPEV4(0.0); } } @@ -125,13 +127,11 @@ void main() { uint r = tid; slope[r] = perElemOpComputeSlope(r, col_tid, ACC_TYPE(0), iq2); } - barrier(); } else { if (tid < Br) { uint r = tid; - slope[r] = 1.0; + slope[r] = ACC_TYPE(1.0); } - barrier(); } #if BLOCK_SIZE > 1 @@ -149,19 +149,45 @@ void main() { [[dont_unroll]] for (uint32_t j = start_j; j < end_j; ++j) { - float mask_cache[Bc * Br / WorkGroupSize]; + f16vec4 mask_cache[Bc * Br / 4 / WorkGroupSize]; if ((p.mask_n_head_log2 & MASK_ENABLE_BIT) != 0) { bool nem1_bounds_check = !(p.gqa_ratio > 1) && (p.nem1 % Br) != 0; float max_mask = NEG_FLT_MAX_OVER_2; - [[unroll]] for (uint32_t idx = 0; idx < Bc * Br; idx += gl_WorkGroupSize.x) { - uint32_t c = (idx + tid) % Bc; - uint32_t r = (idx + tid) / Bc; - if (idx + tid < Bc * Br || idx + gl_WorkGroupSize.x <= Bc * Br) { - if ((!KV_bounds_check || j * Bc + c < KV) && (!nem1_bounds_check || i * Br + r < p.nem1)) { - float m = float(data_m[m_offset + (i * Br + r) * m_stride + (j * Bc + c)]); + [[unroll]] for (uint32_t idx = 0; idx < Bc * Br / 4; idx += gl_WorkGroupSize.x) { + uint32_t c = (idx + tid) / (Br / 4); + uint32_t r = (idx + tid) % (Br / 4); + if (idx + tid < Bc * Br / 4 || idx + gl_WorkGroupSize.x <= Bc * Br / 4) { + if ((!KV_bounds_check || j * Bc + c < KV)) { + f16vec4 m; + if (!nem1_bounds_check || i * Br + r * 4 + 3 < p.nem1) { + m = f16vec4(data_m[m_offset + (i * Br + r * 4 ) * m_stride + (j * Bc + c)], + data_m[m_offset + (i * Br + r * 4 + 1) * m_stride + (j * Bc + c)], + data_m[m_offset + (i * Br + r * 4 + 2) * m_stride + (j * Bc + c)], + data_m[m_offset + (i * Br + r * 4 + 3) * m_stride + (j * Bc + c)]); + max_mask = max(max(max(max(max_mask, float(m[0])), float(m[1])), float(m[2])), float(m[3])); + } else if (i * Br + r * 4 + 2 < p.nem1) { + m = f16vec4(data_m[m_offset + (i * Br + r * 4 ) * m_stride + (j * Bc + c)], + data_m[m_offset + (i * Br + r * 4 + 1) * m_stride + (j * Bc + c)], + data_m[m_offset + (i * Br + r * 4 + 2) * m_stride + (j * Bc + c)], + 0.0); + max_mask = max(max(max(max_mask, float(m[0])), float(m[1])), float(m[2])); + } else if (i * Br + r * 4 + 1 < p.nem1) { + m = f16vec4(data_m[m_offset + (i * Br + r * 4 ) * m_stride + (j * Bc + c)], + data_m[m_offset + (i * Br + r * 4 + 1) * m_stride + (j * Bc + c)], + 0.0, + 0.0); + max_mask = max(max(max_mask, float(m[0])), float(m[1])); + } else if (i * Br + r * 4 < p.nem1) { + m = f16vec4(data_m[m_offset + (i * Br + r * 4 ) * m_stride + (j * Bc + c)], + 0.0, + 0.0, + 0.0); + max_mask = max(max_mask, float(m[0])); + } else { + m = f16vec4(0.0); + } mask_cache[idx / WorkGroupSize] = m; - max_mask = max(max_mask, m); } } } @@ -180,26 +206,28 @@ void main() { } } - [[unroll]] for (uint32_t idx = 0; idx < Bc * HSK / 4; idx += gl_WorkGroupSize.x) { - uint32_t d = (idx + tid) % (HSK / 4); - uint32_t c = (idx + tid) / (HSK / 4); - if (c < Bc && d < HSK / 4) { - f16vec4 K_Tf = f16vec4(0); - if (!KV_bounds_check || j * Bc + c < KV) { + if (K_LOAD_SHMEM != 0) { + [[unroll]] for (uint32_t idx = 0; idx < Bc * HSK / 4; idx += gl_WorkGroupSize.x) { + uint32_t d = (idx + tid) % (HSK / 4); + uint32_t c = (idx + tid) / (HSK / 4); + if (c < Bc && d < HSK / 4) { + f16vec4 K_Tf = f16vec4(0); + if (!KV_bounds_check || j * Bc + c < KV) { #if BLOCK_SIZE > 1 - uint coord = (j * Bc + c) * k_stride * BLOCK_SIZE + 4 * d; - uint ib = coord / BLOCK_SIZE; - uint iqs = (coord % BLOCK_SIZE); - K_Tf = f16vec4(dequantize4(ib, iqs, k_offset, BINDING_IDX_K)); + uint coord = (j * Bc + c) * k_stride * BLOCK_SIZE + 4 * d; + uint ib = coord / BLOCK_SIZE; + uint iqs = (coord % BLOCK_SIZE); + K_Tf = f16vec4(dequantize4(ib, iqs, k_offset, BINDING_IDX_K)); #else - K_Tf = f16vec4(data_kv4[k_offset / 4 + (j * Bc + c) * k_stride / 4 + d]); + K_Tf = f16vec4(data_kv4[k_offset / 4 + (j * Bc + c) * k_stride / 4 + d]); #endif - } + } - ksh[c * kshstride + d] = K_Tf; + ksh[c * kshstride + d] = K_Tf; + } } + barrier(); } - barrier(); // K * Q^T -> S^T: Bc x HSK_pad * HSK_pad x Br -> Bc x Br // Bc split across workgroup (four subgroups), loop over HSK in chunks of 16: 16 x 16 * 16 x 16 -> 16 x 16 @@ -208,11 +236,55 @@ void main() { coopmat KMat; coopmat QMat; - for (uint32_t d = 0; d < HSK_pad / 16; ++d) { - coopMatLoad(QMat, Qf, d * 16 / 4, qstride, gl_CooperativeMatrixLayoutColumnMajor); + [[unroll]] for (uint32_t d = 0; d < HSK_pad / 16; ++d) { + if (K_LOAD_SHMEM == 0) { +#if BLOCK_SIZE == 1 + if (KV_bounds_check || d * 16 + 16 > HSK) { +#endif + barrier(); + [[unroll]] for (uint32_t idx = 0; idx < Bc * MatBr / 4; idx += gl_WorkGroupSize.x) { + uint32_t col_vec = (idx + tid) % (MatBr / 4); + uint32_t row = (idx + tid) / (MatBr / 4); + if (idx + tid < Bc * MatBr / 4) { + f16vec4 K_Tf = f16vec4(0); + if ((!KV_bounds_check || j * Bc + row < KV) && (HSK == HSK_pad || d * 16 + col_vec * 4 < HSK)) { +#if BLOCK_SIZE > 1 + uint coord = (j * Bc + row) * k_stride * BLOCK_SIZE + d * 16 + col_vec * 4; + uint ib = coord / BLOCK_SIZE; + uint iqs = (coord % BLOCK_SIZE); + K_Tf = f16vec4(dequantize4(ib, iqs, k_offset, BINDING_IDX_K)); +#else + K_Tf = f16vec4(data_kv4[k_offset / 4 + (j * Bc + row) * k_stride / 4 + d * 16 / 4 + col_vec]); +#endif + } - uint coord = (gl_SubgroupID * MatBc) * kshstride + d * 16 / 4; - coopMatLoad(KMat, ksh, coord, kshstride, gl_CooperativeMatrixLayoutRowMajor); + ksh[row * kshstride + col_vec] = K_Tf; + } + } + barrier(); +#if BLOCK_SIZE == 1 + } +#endif + +#if BLOCK_SIZE == 1 + if (KV_bounds_check || d * 16 + 16 > HSK) +#endif + { + uint coord = (gl_SubgroupID * MatBc) * kshstride; + coopMatLoad(KMat, ksh, coord, kshstride, gl_CooperativeMatrixLayoutRowMajor); + } +#if BLOCK_SIZE == 1 + else { + const uint coord = k_offset / 4 + (j * Bc + gl_SubgroupID * MatBc) * k_stride / 4 + d * 16 / 4; + coopMatLoad(KMat, data_kv4, coord, k_stride / 4, gl_CooperativeMatrixLayoutRowMajor); + } +#endif + } else { + uint coord = (gl_SubgroupID * MatBc) * kshstride + d * 16 / 4; + coopMatLoad(KMat, ksh, coord, kshstride, gl_CooperativeMatrixLayoutRowMajor); + } + + coopMatLoad(QMat, Qf, d * 16 / 4, qstride, gl_CooperativeMatrixLayoutColumnMajor); SfMat = coopMatMulAdd(KMat, QMat, SfMat); } @@ -222,26 +294,26 @@ void main() { barrier(); if (p.logit_softcap != 0.0f) { - [[unroll]] for (uint32_t idx = 0; idx < Bc * Br; idx += gl_WorkGroupSize.x) { - uint32_t c = (idx + tid) / Br; - uint32_t r = (idx + tid) % Br; - if (idx + tid < Bc * Br || idx + gl_WorkGroupSize.x <= Bc * Br) { - sfsh[c * sfshstride + r] = ACC_TYPE(p.logit_softcap * tanh(sfsh[c * sfshstride + r])); + [[unroll]] for (uint32_t idx = 0; idx < Bc * Br / 4; idx += gl_WorkGroupSize.x) { + uint32_t c = (idx + tid) / (Br / 4); + uint32_t r = (idx + tid) % (Br / 4); + if (idx + tid < Bc * Br / 4 || idx + gl_WorkGroupSize.x <= Bc * Br / 4) { + sfsh[c * sfshstride + r] = ACC_TYPEV4(p.logit_softcap * tanh(sfsh[c * sfshstride + r])); } } barrier(); } if ((p.mask_n_head_log2 & MASK_ENABLE_BIT) != 0) { - bool nem1_bounds_check = !(p.gqa_ratio > 1) && (p.nem1 % Br) != 0; - - [[unroll]] for (uint32_t idx = 0; idx < Bc * Br; idx += gl_WorkGroupSize.x) { - uint32_t c = (idx + tid) % Bc; - uint32_t r = (idx + tid) / Bc; - if (idx + tid < Bc * Br || idx + gl_WorkGroupSize.x <= Bc * Br) { - if ((!KV_bounds_check || j * Bc + c < KV) && (!nem1_bounds_check || i * Br + r < p.nem1)) { - float f = mask_cache[idx / WorkGroupSize]; - sfsh[c * sfshstride + r] += ACC_TYPE(slope[r] * f); + [[unroll]] for (uint32_t idx = 0; idx < Bc * Br / 4; idx += gl_WorkGroupSize.x) { + uint32_t c = (idx + tid) / (Br / 4); + uint32_t r = (idx + tid) % (Br / 4); + if (idx + tid < Bc * Br / 4 || idx + gl_WorkGroupSize.x <= Bc * Br / 4) { + if (!KV_bounds_check || j * Bc + c < KV) { + // Mask nem1 bounds check is handled when loading masks + ACC_TYPEV4 masks = ACC_TYPEV4(mask_cache[idx / WorkGroupSize]); + ACC_TYPEV4 slopes = ACC_TYPEV4(slope[r * 4], slope[r * 4 + 1], slope[r * 4 + 2], slope[r * 4 + 3]); + sfsh[c * sfshstride + r] += slopes * masks; } } } @@ -250,51 +322,145 @@ void main() { float eMf[rows_per_thread]; [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) { + const uint r_vec = tile_row(r) / 4; + const uint r_comp = tile_row(r) % 4; + float rowmaxf = NEG_FLT_MAX_OVER_2; [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) { if (KV_bounds_check && j * Bc + c * cols_per_iter + col_tid >= KV) { continue; } - rowmaxf = max(rowmaxf, float(sfsh[tile_row(r) + (c * cols_per_iter + col_tid) * sfshstride])); + rowmaxf = max(rowmaxf, float(sfsh[r_vec + (c * cols_per_iter + col_tid) * sfshstride][r_comp])); } float Moldf = Mf[r]; + // Compute max across the row + rowmaxf = subgroupMax(rowmaxf); + // M = max(rowmax, Mold) // P = e^(S - M) // eM = e^(Mold - M) Mf[r] = max(rowmaxf, Moldf); eMf[r] = exp(Moldf - Mf[r]); - } - [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) { - [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) { - Of[r][d] = ACC_TYPE(eMf[r]) * Of[r][d]; - } - } - [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) { Lf[r] = eMf[r]*Lf[r]; } - [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) { - if (KV_bounds_check && j * Bc + c * cols_per_iter + col_tid >= KV) { - continue; - } - float Pf[rows_per_thread]; + [[unroll]] for (uint32_t d0 = 0; d0 < HSV / 4; d0 += threads_per_rowgroup) { + const uint d_local = d0 / threads_per_rowgroup; [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) { - Pf[r] = exp(sfsh[tile_row(r) + (c * cols_per_iter + col_tid) * sfshstride] - Mf[r]); - Lf[r] += Pf[r]; + Of[r][d_local] = ACC_TYPE(eMf[r]) * Of[r][d_local]; } - [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) { -#if BLOCK_SIZE > 1 - uint coord = (j * Bc + c * cols_per_iter + col_tid) * v_stride * BLOCK_SIZE + 4 * (d * D_split + d_tid); - uint ib = coord / BLOCK_SIZE; - uint iqs = (coord % BLOCK_SIZE); - vec4 Vf = dequantize4(ib, iqs, v_offset, BINDING_IDX_V); -#else - vec4 Vf = vec4(data_vv4[v_offset / 4 + (j * Bc + c * cols_per_iter + col_tid) * v_stride / 4 + d * D_split + d_tid]); + } + + // Calculate and store Pf in Psh + [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) { + const uint col = c * cols_per_iter + col_tid; + + [[unroll]] for (uint32_t r = 0; r < rows_per_thread; r += 4) { + const uint row = tile_row(r); + if (KV_bounds_check && j * Bc + col >= KV) { + Psh[col * psh_stride + row / 4] = f16vec4(0.0f); + } else { + const vec4 mfvec = vec4(Mf[r], Mf[r + 1], Mf[r + 2], Mf[r + 3]); + const f16vec4 Pf = f16vec4(exp(vec4(sfsh[row / 4 + col * sfshstride]) - mfvec)); + [[unroll]] for (uint32_t vec_idx = 0; vec_idx < 4; ++vec_idx) { + Lf[r + vec_idx] += Pf[vec_idx]; + } + Psh[col * psh_stride + row / 4] = Pf; + } + } + } + + const uint num_hsv_tiles = (HSV + MatBc * row_split - 1) / (MatBc * row_split); // round up + + // Each subgroup handles HSV/4 columns + [[unroll]] for (uint32_t hsv_tile = 0; hsv_tile < num_hsv_tiles; ++hsv_tile) { + const uint hsv_offset = (hsv_tile * row_split + gl_SubgroupID) * 16; + + SfMat = coopmat(0); + + // Preload V tiles for [Bc, 16 * num subgroups] + const uint v_rows = Bc; + const uint v_total = v_rows * v_cols; + const uint v_loads_per_thread = v_total / gl_WorkGroupSize.x; + +#if BLOCK_SIZE == 1 + // For f16, only preload if not aligned + if (KV_bounds_check) { #endif - [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) { - Of[r][d] += ACC_TYPE(Pf[r]) * ACC_TYPEV4(Vf); + [[unroll]] for (uint32_t i = 0; i < v_loads_per_thread; ++i) { + const uint idx = i * gl_WorkGroupSize.x + tid; + const uint row = idx / v_cols; + const uint col = idx % v_cols; + + const uint v_row = j * Bc + row; + const uint v_col = hsv_tile * MatBc * row_split + col * 4; + + const uint coord = v_row * v_stride * BLOCK_SIZE + v_col; + const uint ib = coord / BLOCK_SIZE; + const uint iqs = coord % BLOCK_SIZE; + + if (!KV_bounds_check || (v_row < KV && v_col < HSV)) { +#if BLOCK_SIZE > 1 + ksh[row * vsh_stride + col] = f16vec4(dequantize4(ib, iqs, v_offset, BINDING_IDX_V)); +#else + ksh[row * vsh_stride + col] = data_vv4[(v_offset + v_row * v_stride + v_col) / 4]; +#endif + } else { + ksh[row * vsh_stride + col] = f16vec4(0.0f); + } + } +#if BLOCK_SIZE == 1 + } +#endif + + barrier(); + + [[unroll]] for (uint32_t bc_chunk = 0; bc_chunk < Bc / MatBc; ++bc_chunk) { + coopMatLoad(KMat, Psh, bc_chunk * MatBc * psh_stride, psh_stride, gl_CooperativeMatrixLayoutColumnMajor); + +#if BLOCK_SIZE == 1 + if (!KV_bounds_check) { + // F16 values can be loaded directly from global memory + const uint v_tile_row = j * Bc + bc_chunk * MatBc; + const uint v_tile_offset = v_offset / 4 + v_tile_row * v_stride / 4 + hsv_offset / 4; + coopMatLoad(QMat, data_vv4, v_tile_offset, v_stride / 4, gl_CooperativeMatrixLayoutRowMajor); + } else +#endif + { + const uint v_tile_offset = bc_chunk * MatBr * v_cols + gl_SubgroupID * (MatBc / 4); + coopMatLoad(QMat, ksh, v_tile_offset, vsh_stride, gl_CooperativeMatrixLayoutRowMajor); + } + + SfMat = coopMatMulAdd(KMat, QMat, SfMat); + } + + // Store SfMat to sfsh and load into Of + const uint osh_stride = row_split * MatBc / 4; + const uint o_offset = gl_SubgroupID * MatBc / 4; + coopMatStore(SfMat, sfsh, o_offset, osh_stride, gl_CooperativeMatrixLayoutRowMajor); + + barrier(); + + const uint hsv_per_tile = row_split * MatBc; + const uint hsv_base = hsv_tile * hsv_per_tile; + const uint d_values_per_tile = hsv_per_tile / 4; + + const uint d_start = hsv_tile * d_values_per_tile; + const uint d_end = min(d_start + d_values_per_tile, HSV / 4); + + [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) { + const uint row = tile_row(r); + + [[unroll]] for (uint32_t d_local = 0; d_local < d_per_thread; ++d_local) { + const uint d = d_local * threads_per_rowgroup + col_tid; + const uint hsv_col = 4 * d; + + if (hsv_col >= hsv_base && hsv_col < hsv_base + hsv_per_tile && hsv_col < HSV) { + const uint local_hsv = (hsv_col - hsv_base) / 4; + Of[r][d_local] += ACC_TYPEV4(sfsh[row * osh_stride + local_hsv]); + } } } } @@ -302,69 +468,8 @@ void main() { barrier(); } - // prevent race on tmpsh - barrier(); - - // reduce across threads - - float rowmaxf[rows_per_thread], eMf[rows_per_thread], Moldf[rows_per_thread]; [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) { - FLOAT_TYPE M = Mf[r]; - tmpsh[tid] = M; - // Compute max across the row - barrier(); - [[unroll]] for (int s = int(gl_WorkGroupSize.x / row_split) / 2; s >= D_split; s >>= 1) { - M = max(M, tmpsh[tid ^ s]); - barrier(); - tmpsh[tid] = M; - barrier(); - } - rowmaxf[r] = tmpsh[d_tid + row_tid * threads_per_rowgroup]; - barrier(); - } - - [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) { - Moldf[r] = Mf[r]; - - // M = max(rowmax, Mold) - // eM = e^(Mold - M) - Mf[r] = max(rowmaxf[r], Moldf[r]); - eMf[r] = exp(Moldf[r] - Mf[r]); - - Lf[r] = eMf[r]*Lf[r]; - } - - [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) { - FLOAT_TYPE L = Lf[r]; - tmpsh[tid] = L; - // Compute sum across the row - barrier(); - [[unroll]] for (int s = int(gl_WorkGroupSize.x / row_split) / 2; s >= D_split; s >>= 1) { - L += tmpsh[tid ^ s]; - barrier(); - tmpsh[tid] = L; - barrier(); - } - Lf[r] = tmpsh[d_tid + row_tid * threads_per_rowgroup]; - barrier(); - } - - [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) { - [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) { - - Of[r][d] = ACC_TYPE(eMf[r]) * Of[r][d]; - tmpshv4[tid] = Of[r][d]; - - barrier(); - [[unroll]] for (int s = int(gl_WorkGroupSize.x / row_split) / 2; s >= D_split; s >>= 1) { - Of[r][d] += tmpshv4[tid ^ s]; - barrier(); - tmpshv4[tid] = Of[r][d]; - barrier(); - } - Of[r][d] = tmpshv4[d_tid + row_tid * threads_per_rowgroup]; - barrier(); - } + Lf[r] = subgroupAdd(Lf[r]); } // If there is split_k, then the split_k resolve shader does the final @@ -375,9 +480,12 @@ void main() { [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) { if (tile_row(r) < N) { - [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) { + [[unroll]] for (uint32_t d0 = 0; d0 < HSV / 4; d0 += threads_per_rowgroup) { + const uint d = d0 + col_tid; + if (d >= HSV/4) break; + const uint d_local = d0 / threads_per_rowgroup; [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) { - perElemOpGqaStore(tile_row(r), 4*(d * D_split + d_tid) + comp, float(Of[r][d][comp]), o_offset, iq2, N); + perElemOpGqaStore(tile_row(r), 4 * d + comp, float(Of[r][d_local][comp]), o_offset, iq2, N); } } } @@ -404,8 +512,9 @@ void main() { if (sink > Mf[r]) { ms = exp(Mf[r] - sink); - [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) { - Of[r][d] *= ACC_TYPE(ms); + [[unroll]] for (uint32_t d0 = 0; d0 < HSV / 4; d0 += threads_per_rowgroup) { + const uint d_local = d0 / threads_per_rowgroup; + Of[r][d_local] *= ACC_TYPE(ms); } } else { vs = exp(sink - Mf[r]); @@ -420,11 +529,12 @@ void main() { Lfrcp[r] = (Lf[r] == 0.0) ? 0.0 : (1.0 / Lf[r]); } - [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) { + [[unroll]] for (uint32_t d0 = 0; d0 < HSV / 4; d0 += threads_per_rowgroup) { + const uint d_local = d0 / threads_per_rowgroup; [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) { - Of[r][d] *= ACC_TYPE(Lfrcp[r]); + Of[r][d_local] *= ACC_TYPE(Lfrcp[r]); #if defined(ACC_TYPE_MAX) - Of[r][d] = clamp(Of[r][d], -ACC_TYPE_MAX, ACC_TYPE_MAX); + Of[r][d_local] = clamp(Of[r][d_local], -ACC_TYPE_MAX, ACC_TYPE_MAX); #endif } } @@ -434,9 +544,12 @@ void main() { if (p.gqa_ratio > 1) { [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) { if (tile_row(r) < N) { - [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) { + [[unroll]] for (uint32_t d0 = 0; d0 < HSV / 4; d0 += threads_per_rowgroup) { + const uint d = d0 + col_tid; + if (d >= HSV / 4) break; + const uint d_local = d0 / threads_per_rowgroup; [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) { - perElemOpGqaStore(tile_row(r), 4*(d * D_split + d_tid) + comp, float(Of[r][d][comp]), o_offset, iq2, N); + perElemOpGqaStore(tile_row(r), 4 * d + comp, float(Of[r][d_local][comp]), o_offset, iq2, N); } } } @@ -444,9 +557,12 @@ void main() { } else { [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) { if (i * Br + tile_row(r) < N) { - [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) { + [[unroll]] for (uint32_t d0 = 0; d0 < HSV / 4; d0 += threads_per_rowgroup) { + const uint d = d0 + col_tid; + if (d >= HSV / 4) break; + const uint d_local = d0 / threads_per_rowgroup; [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) { - data_o[o_offset + iq2 * HSV + (i * Br + tile_row(r)) * p.ne1 * HSV + 4*(d * D_split + d_tid) + comp] = D_TYPE(Of[r][d][comp]); + data_o[o_offset + iq2 * HSV + (i * Br + tile_row(r)) * p.ne1 * HSV + 4 * d + comp] = D_TYPE(Of[r][d_local][comp]); } } } diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp b/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp index d49a8da65f..54f1b0b622 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp @@ -55,7 +55,7 @@ ACC_TYPE Max(const in uint32_t row, const in uint32_t col, const in ACC_TYPE ele return max(elem0, elem1); } -#if defined(BLOCK_SIZE) +#if BLOCK_SIZE > 1 #define DECODEFUNC , DEQUANTFUNC #else #define DECODEFUNC @@ -85,7 +85,7 @@ void main() { tensorViewNV<2, false, 1, 0> tensorViewTranspose = createTensorViewNV(2, false, 1, 0); -#if defined(BLOCK_SIZE) +#if BLOCK_SIZE > 1 tensorLayoutK = setTensorLayoutBlockSizeNV(tensorLayoutK, 1, BLOCK_SIZE); tensorLayoutV = setTensorLayoutBlockSizeNV(tensorLayoutV, 1, BLOCK_SIZE); #endif @@ -98,7 +98,7 @@ void main() { if (Clamp != gl_CooperativeMatrixClampModeConstantNV) { q_stride &= ~7; -#if !defined(BLOCK_SIZE) +#if BLOCK_SIZE == 1 k_stride &= ~7; v_stride &= ~7; #endif diff --git a/ggml/src/ggml-webgpu/ggml-webgpu.cpp b/ggml/src/ggml-webgpu/ggml-webgpu.cpp index 584cea7698..22e2bfeb4c 100644 --- a/ggml/src/ggml-webgpu/ggml-webgpu.cpp +++ b/ggml/src/ggml-webgpu/ggml-webgpu.cpp @@ -47,7 +47,6 @@ double cpu_total_time_##id = \ std::chrono::duration(cpu_total_end_##id - cpu_total_start_##id).count(); \ (ctx)->cpu_time_ms[#id] += cpu_total_time_##id; - // fine-grained timing (not included in totals) # define WEBGPU_CPU_PROFILE_DETAIL_START(id) auto cpu_detail_start_##id = std::chrono::high_resolution_clock::now(); @@ -74,13 +73,13 @@ #define WEBGPU_MAX_WG_SIZE 288 #define WEBGPU_MUL_MAT_WG_SIZE 256 -#define WEBGPU_NUM_PARAM_BUFS 32u +#define WEBGPU_NUM_PARAM_BUFS 16u #define WEBGPU_COMMAND_SUBMIT_BATCH_SIZE 8u #define WEBGPU_WAIT_ANY_TIMEOUT_MS 0 // Maximum number of in-flight submissions per-thread, to avoid exhausting the parameter buffer pool #define WEBGPU_MAX_INFLIGHT_SUBS_PER_THREAD WEBGPU_NUM_PARAM_BUFS / WEBGPU_COMMAND_SUBMIT_BATCH_SIZE #define WEBGPU_PARAMS_BUF_SIZE_BYTES 128 // enough for 32 parameters -#define WEBGPU_NUM_SET_ROWS_ERROR_BUFS 32 +#define WEBGPU_NUM_SET_ROWS_ERROR_BUFS 16 #define WEBGPU_SET_ROWS_ERROR_BUF_SIZE_BYTES 4 #define WEBGPU_STORAGE_BUF_BINDING_MULT 4 // a storage buffer binding size must be a multiple of 4 @@ -267,30 +266,67 @@ struct webgpu_command { #endif }; -// All the base objects needed to run operations on a WebGPU device -struct webgpu_context_struct { +struct webgpu_capabilities_base { + wgpu::Limits limits; + bool supports_subgroup_matrix = false; + + uint32_t sg_mat_m = 0; + uint32_t sg_mat_n = 0; + uint32_t sg_mat_k = 0; + + uint32_t subgroup_size = 0; + uint32_t max_subgroup_size = 0; + size_t memset_bytes_per_thread; +}; + +// Stores global webgpu members +struct webgpu_global_context_struct { wgpu::Instance instance; wgpu::Adapter adapter; wgpu::Device device; wgpu::Queue queue; - wgpu::Limits limits; - uint32_t max_subgroup_size; + webgpu_capabilities_base capabilities; + // Shared buffer to move data from device to host + wgpu::Buffer get_tensor_staging_buf; + // Global mutex for pipeline and staging buffer, will be refactored to exclude pipeline caches. + std::recursive_mutex mutex; - bool supports_subgroup_matrix = false; - uint32_t sg_mat_m; - uint32_t sg_mat_n; - uint32_t sg_mat_k; + webgpu_buf_pool memset_buf_pool; + std::map memset_pipelines; // variant or type index + std::atomic_uint inflight_threads = 0; - std::recursive_mutex mutex; - std::atomic_uint inflight_threads = 0; +#ifdef GGML_WEBGPU_CPU_PROFILE + // Profiling: labeled CPU time in ms (total) + std::unordered_map cpu_time_ms; + // Profiling: detailed CPU time in ms + std::unordered_map cpu_detail_ms; +#endif - webgpu_buf_pool param_buf_pool; - webgpu_buf_pool set_rows_error_buf_pool; +#ifdef GGML_WEBGPU_GPU_PROFILE + // Profiling: per-shader GPU time in ms + std::unordered_map shader_gpu_time_ms; + // Profiling: pool of timestamp query buffers (one per operation) + webgpu_gpu_profile_buf_pool timestamp_query_buf_pool; +#endif + +#ifdef GGML_WEBGPU_DEBUG + wgpu::Buffer debug_host_buf; + wgpu::Buffer debug_dev_buf; +#endif +}; + +typedef std::shared_ptr webgpu_global_context; + +// All the base objects needed to run operations on a WebGPU device +struct webgpu_context_struct { + // Points to global instances owned by ggml_backend_webgpu_reg_context + webgpu_global_context global_ctx; pre_wgsl::Preprocessor p; - std::map memset_pipelines; // variant or type index + webgpu_buf_pool param_buf_pool; + webgpu_buf_pool set_rows_error_buf_pool; std::map>> mul_mat_pipelines; // src0_type, src1_type, vectorized std::map>> @@ -326,57 +362,42 @@ struct webgpu_context_struct { size_t memset_bytes_per_thread; - // Staging buffer for reading data from the GPU - wgpu::Buffer get_tensor_staging_buf; - -#ifdef GGML_WEBGPU_DEBUG - wgpu::Buffer debug_host_buf; - wgpu::Buffer debug_dev_buf; -#endif - -#ifdef GGML_WEBGPU_CPU_PROFILE - // Profiling: labeled CPU time in ms (total) - std::unordered_map cpu_time_ms; - // Profiling: detailed CPU time in ms - std::unordered_map cpu_detail_ms; -#endif - -#ifdef GGML_WEBGPU_GPU_PROFILE - // Profiling: per-shader GPU time in ms - std::unordered_map shader_gpu_time_ms; - // Profiling: pool of timestamp query buffers (one per operation) - webgpu_gpu_profile_buf_pool timestamp_query_buf_pool; -#endif }; typedef std::shared_ptr webgpu_context; +// Metadata required for the ggml backend registration/discovery interface struct ggml_backend_webgpu_reg_context { - webgpu_context webgpu_ctx; - size_t device_count; - const char * name; + // Since the Instance is a global entrypoint into the WebGPU API, it lives here + webgpu_global_context webgpu_global_ctx; + size_t device_count; + const char * name; }; +// Per-device struct for the global logical device interface struct ggml_backend_webgpu_device_context { - webgpu_context webgpu_ctx; - std::string device_name; - std::string device_desc; + webgpu_global_context webgpu_global_ctx; + std::string device_name; + std::string device_desc; }; +// Per-thread data required to actually run WebGPU operations in a backend instance struct ggml_backend_webgpu_context { - webgpu_context webgpu_ctx; - std::string name; + webgpu_context webgpu_ctx; + std::once_flag init_once; + std::string name; }; +// Per-thread data related to buffers struct ggml_backend_webgpu_buffer_context { - webgpu_context webgpu_ctx; - wgpu::Buffer buffer; - std::string label; + wgpu::Buffer buffer; + std::string label; + webgpu_global_context global_ctx; - ggml_backend_webgpu_buffer_context(webgpu_context ctx, wgpu::Buffer buf, std::string lbl) : - webgpu_ctx(std::move(ctx)), + ggml_backend_webgpu_buffer_context(wgpu::Buffer buf, std::string lbl, webgpu_global_context global_ctx_) : buffer(std::move(buf)), - label(std::move(lbl)) {} + label(std::move(lbl)), + global_ctx(std::move(global_ctx_)) {} }; /* WebGPU object initializations */ @@ -444,7 +465,7 @@ static void ggml_webgpu_create_buffer(wgpu::Device & device, /** WebGPU Actions */ // Wait for the queue to finish processing all submitted work -static void ggml_backend_webgpu_wait(webgpu_context & ctx, +static void ggml_backend_webgpu_wait(webgpu_global_context & ctx, std::vector & futures, bool block = true) { // If we have too many in-flight submissions, wait on the oldest one first. If there are many threads, @@ -476,11 +497,11 @@ static void ggml_backend_webgpu_wait(webgpu_context & ct } } -static void ggml_backend_webgpu_map_buffer(webgpu_context & ctx, - wgpu::Buffer & buffer, - wgpu::MapMode mode, - size_t offset, - size_t size) { +static void ggml_backend_webgpu_map_buffer(webgpu_global_context & ctx, + wgpu::Buffer & buffer, + wgpu::MapMode mode, + size_t offset, + size_t size) { ctx->instance.WaitAny(buffer.MapAsync(mode, offset, size, wgpu::CallbackMode::AllowSpontaneous, [](wgpu::MapAsyncStatus status, wgpu::StringView message) { if (status != wgpu::MapAsyncStatus::Success) { @@ -495,7 +516,7 @@ static void ggml_backend_webgpu_map_buffer(webgpu_context & ctx, // This function adds debugging information to shaders, as WebGPU does not support printing directly. // To use, add a bind group entry to the setup for the shader you are debugging, add the buffer and // debug statements in the shader, and then call this function after encoding the commands and submitting them. -static void ggml_backend_webgpu_debug(webgpu_context & ctx) { +static void ggml_backend_webgpu_debug(webgpu_global_context & ctx) { wgpu::CommandEncoder encoder = ctx->device.CreateCommandEncoder(); encoder.CopyBufferToBuffer(ctx->debug_dev_buf, 0, ctx->debug_host_buf, 0, ctx->debug_host_buf.GetSize()); wgpu::CommandBuffer commands = encoder.Finish(); @@ -507,7 +528,10 @@ static void ggml_backend_webgpu_debug(webgpu_context & ctx) { } #endif -static webgpu_submission_futures ggml_backend_webgpu_submit(webgpu_context ctx, std::vector commands) { +static webgpu_submission_futures ggml_backend_webgpu_submit(webgpu_global_context ctx, + std::vector commands, + webgpu_buf_pool & param_buf_pool, + webgpu_buf_pool * set_rows_error_buf_pool = nullptr) { std::vector command_buffers; std::vector params_bufs; std::vector set_rows_error_bufs; @@ -528,19 +552,19 @@ static webgpu_submission_futures ggml_backend_webgpu_submit(webgpu_context ctx, wgpu::Future p_f = ctx->queue.OnSubmittedWorkDone( wgpu::CallbackMode::AllowSpontaneous, - [ctx, params_bufs](wgpu::QueueWorkDoneStatus status, wgpu::StringView message) { + [¶m_buf_pool, params_bufs](wgpu::QueueWorkDoneStatus status, wgpu::StringView message) { if (status != wgpu::QueueWorkDoneStatus::Success) { GGML_LOG_ERROR("ggml_webgpu: Failed to submit commands: %s\n", std::string(message).c_str()); } // Free the staged buffers - ctx->param_buf_pool.free_bufs(params_bufs); + param_buf_pool.free_bufs(params_bufs); }); futures.push_back({ p_f }); for (const auto & bufs : set_rows_error_bufs) { wgpu::Future f = bufs.host_buf.MapAsync( wgpu::MapMode::Read, 0, bufs.host_buf.GetSize(), wgpu::CallbackMode::AllowSpontaneous, - [ctx, bufs](wgpu::MapAsyncStatus status, wgpu::StringView message) { + [set_rows_error_buf_pool, bufs](wgpu::MapAsyncStatus status, wgpu::StringView message) { if (status != wgpu::MapAsyncStatus::Success) { GGML_LOG_ERROR("ggml_webgpu: Failed to map error buffer: %s\n", std::string(message).c_str()); } else { @@ -549,7 +573,9 @@ static webgpu_submission_futures ggml_backend_webgpu_submit(webgpu_context ctx, GGML_ABORT("ggml_webgpu: SET_ROWS index > 2^32, unsupported."); } // We can't unmap in here due to WebGPU reentrancy limitations. - ctx->set_rows_error_buf_pool.free_bufs({ bufs }); + if (set_rows_error_buf_pool) { + set_rows_error_buf_pool->free_bufs({ bufs }); + } } }); futures.push_back({ f }); @@ -581,7 +607,8 @@ static webgpu_submission_futures ggml_backend_webgpu_submit(webgpu_context ctx, } static webgpu_command ggml_backend_webgpu_build_multi( - webgpu_context & ctx, + webgpu_global_context & ctx, + webgpu_buf_pool & param_buf_pool, const std::vector & pipelines, const std::vector> & params_list, const std::vector> & bind_group_entries_list, @@ -595,7 +622,7 @@ static webgpu_command ggml_backend_webgpu_build_multi( std::vector bind_groups; for (size_t i = 0; i < pipelines.size(); i++) { - webgpu_pool_bufs params_bufs = ctx->param_buf_pool.alloc_bufs(); + webgpu_pool_bufs params_bufs = param_buf_pool.alloc_bufs(); ggml_backend_webgpu_map_buffer(ctx, params_bufs.host_buf, wgpu::MapMode::Write, 0, params_bufs.host_buf.GetSize()); @@ -672,34 +699,37 @@ static webgpu_command ggml_backend_webgpu_build_multi( return result; } -static webgpu_command ggml_backend_webgpu_build(webgpu_context & ctx, +static webgpu_command ggml_backend_webgpu_build(webgpu_global_context & ctx, + webgpu_buf_pool & param_buf_pool, webgpu_pipeline & pipeline, std::vector params, std::vector bind_group_entries, uint32_t wg_x, uint32_t wg_y = 1, std::optional set_rows_error_bufs = std::nullopt) { - return ggml_backend_webgpu_build_multi(ctx, + return ggml_backend_webgpu_build_multi(ctx, param_buf_pool, { pipeline }, { params }, { bind_group_entries }, { { wg_x, wg_y } }, set_rows_error_bufs); } -static void ggml_backend_webgpu_buffer_memset(webgpu_context & ctx, - wgpu::Buffer & buf, - uint32_t value, - size_t offset, - size_t size) { +static void ggml_backend_webgpu_buffer_memset(webgpu_global_context & ctx, + wgpu::Buffer & buf, + uint32_t value, + size_t offset, + size_t size) { std::vector params = { (uint32_t) offset, (uint32_t) size, value }; std::vector entries = { { .binding = 0, .buffer = buf, .offset = 0, .size = buf.GetSize() } }; - size_t bytes_per_wg = WEBGPU_MAX_WG_SIZE * ctx->memset_bytes_per_thread; + size_t bytes_per_wg = WEBGPU_MAX_WG_SIZE * ctx->capabilities.memset_bytes_per_thread; uint32_t wg_x = CEIL_DIV(size + 3, bytes_per_wg); - webgpu_command command = ggml_backend_webgpu_build(ctx, ctx->memset_pipelines[0], params, entries, wg_x); - std::vector futures = { ggml_backend_webgpu_submit(ctx, { command }) }; + webgpu_command command = + ggml_backend_webgpu_build(ctx, ctx->memset_buf_pool, ctx->memset_pipelines[0], params, entries, wg_x); + std::vector futures = { ggml_backend_webgpu_submit(ctx, { command }, + ctx->memset_buf_pool) }; ggml_backend_webgpu_wait(ctx, futures); } @@ -720,19 +750,19 @@ static void ggml_backend_webgpu_free(ggml_backend_t backend) { #ifdef GGML_WEBGPU_CPU_PROFILE std::cout << "\n[ggml_webgpu cpu profiling summary]\n"; double total_cpu = 0.0; - for (const auto & kv : ctx->webgpu_ctx->cpu_time_ms) { + for (const auto & kv : ctx->webgpu_ctx->global_ctx->cpu_time_ms) { total_cpu += kv.second; } std::cout << "ggml_webgpu: total cpu time: " << total_cpu << " ms\n"; std::cout << "ggml_webgpu: cpu breakdown:\n"; - for (const auto & kv : ctx->webgpu_ctx->cpu_time_ms) { + for (const auto & kv : ctx->webgpu_ctx->global_ctx->cpu_time_ms) { double pct = (total_cpu > 0.0) ? (kv.second / total_cpu * 100.0) : 0.0; std::cout << "ggml_webgpu: " << kv.first << ": " << kv.second << " ms (" << pct << "%)\n"; } - if (ctx->webgpu_ctx->cpu_detail_ms.size() > 0) { + if (ctx->webgpu_ctx->global_ctx->cpu_detail_ms.size() > 0) { std::cout << "ggml_webgpu: cpu detailed breakdown:\n"; } - for (const auto & kv : ctx->webgpu_ctx->cpu_detail_ms) { + for (const auto & kv : ctx->webgpu_ctx->global_ctx->cpu_detail_ms) { double pct = (total_cpu > 0.0) ? (kv.second / total_cpu * 100.0) : 0.0; std::cout << "ggml_webgpu: " << kv.first << ": " << kv.second << " ms (" << pct << "%)\n"; } @@ -741,12 +771,12 @@ static void ggml_backend_webgpu_free(ggml_backend_t backend) { #ifdef GGML_WEBGPU_GPU_PROFILE std::cout << "\n[ggml_webgpu gpu profiling summary]\n"; double total_gpu = 0.0; - for (const auto & kv : ctx->webgpu_ctx->shader_gpu_time_ms) { + for (const auto & kv : ctx->webgpu_ctx->global_ctx->shader_gpu_time_ms) { total_gpu += kv.second; } std::cout << "ggml_webgpu: total gpu time (all shaders): " << total_gpu << " ms\n"; std::cout << "\nggml_webgpu: gpu breakdown:\n"; - for (const auto & kv : ctx->webgpu_ctx->shader_gpu_time_ms) { + for (const auto & kv : ctx->webgpu_ctx->global_ctx->shader_gpu_time_ms) { double pct = (total_gpu > 0.0) ? (kv.second / total_gpu * 100.0) : 0.0; std::cout << "ggml_webgpu: " << kv.first << ": " << kv.second << " ms (" << pct << "%)\n"; } @@ -772,12 +802,12 @@ static wgpu::Buffer ggml_webgpu_tensor_buf(const ggml_tensor * tensor) { static size_t ggml_webgpu_tensor_misalignment(webgpu_context & ctx, const ggml_tensor * t) { size_t offset = ggml_webgpu_tensor_offset(t); - return offset & (ctx->limits.minStorageBufferOffsetAlignment - 1); + return offset & (ctx->global_ctx->capabilities.limits.minStorageBufferOffsetAlignment - 1); } static size_t ggml_webgpu_tensor_align_offset(webgpu_context & ctx, const ggml_tensor * t) { size_t offset = ggml_webgpu_tensor_offset(t); - return offset & ~(ctx->limits.minStorageBufferOffsetAlignment - 1); + return offset & ~(ctx->global_ctx->capabilities.limits.minStorageBufferOffsetAlignment - 1); } static size_t ggml_webgpu_tensor_binding_size(webgpu_context & ctx, ggml_tensor * t) { @@ -818,28 +848,30 @@ static webgpu_command ggml_webgpu_cpy(webgpu_context & ctx, ggml_tensor * src, g }; uint32_t wg_x = CEIL_DIV(ne, WEBGPU_MAX_WG_SIZE); - return ggml_backend_webgpu_build(ctx, ctx->cpy_pipelines[src->type][dst->type], params, entries, wg_x); + return ggml_backend_webgpu_build(ctx->global_ctx, ctx->param_buf_pool, ctx->cpy_pipelines[src->type][dst->type], + params, entries, wg_x); } static webgpu_command ggml_webgpu_pad(webgpu_context & ctx, ggml_tensor * src, ggml_tensor * dst) { const bool circular = ggml_get_op_params_i32(dst, 8) != 0; ggml_webgpu_pad_pipeline_key pipeline_key = { .circular = circular }; - ggml_webgpu_pad_shader_lib_context shader_lib_ctx = { .key = pipeline_key, - .max_wg_size = - ctx->limits.maxComputeInvocationsPerWorkgroup }; + ggml_webgpu_pad_shader_lib_context shader_lib_ctx = { + .key = pipeline_key, .max_wg_size = ctx->global_ctx->capabilities.limits.maxComputeInvocationsPerWorkgroup + }; webgpu_pipeline pipeline; { // TODO: remove guard once pipeline caches are per-thread - std::lock_guard lock(ctx->mutex); + std::lock_guard lock(ctx->global_ctx->mutex); auto it = ctx->pad_pipelines.find(pipeline_key); if (it != ctx->pad_pipelines.end()) { pipeline = it->second; } else { ggml_webgpu_processed_shader processed = ggml_webgpu_preprocess_pad_shader(ctx->p, wgsl_pad, shader_lib_ctx); - pipeline = ggml_webgpu_create_pipeline(ctx->device, processed.wgsl.c_str(), processed.variant.c_str()); + pipeline = + ggml_webgpu_create_pipeline(ctx->global_ctx->device, processed.wgsl.c_str(), processed.variant.c_str()); pipeline.context = processed.decisions; ctx->pad_pipelines.emplace(pipeline_key, pipeline); } @@ -891,7 +923,7 @@ static webgpu_command ggml_webgpu_pad(webgpu_context & ctx, ggml_tensor * src, g }; uint32_t wg_x = CEIL_DIV(ne, decisions.wg_size); - return ggml_backend_webgpu_build(ctx, pipeline, params, entries, wg_x); + return ggml_backend_webgpu_build(ctx->global_ctx, ctx->param_buf_pool, pipeline, params, entries, wg_x); } static std::optional ggml_webgpu_set_rows(webgpu_context & ctx, @@ -907,21 +939,22 @@ static std::optional ggml_webgpu_set_rows(webgpu_context & ctx, .vec4 = src->ne[0] % 4 == 0, .i64_idx = idx->type == GGML_TYPE_I64 }; - ggml_webgpu_set_rows_shader_lib_context shader_lib_ctx = { .key = key, - .max_wg_size = - ctx->limits.maxComputeInvocationsPerWorkgroup }; + ggml_webgpu_set_rows_shader_lib_context shader_lib_ctx = { + .key = key, .max_wg_size = ctx->global_ctx->capabilities.limits.maxComputeInvocationsPerWorkgroup + }; webgpu_pipeline pipeline; // TODO: remove guard once pipeline caches are per-thread { - std::lock_guard lock(ctx->mutex); + std::lock_guard lock(ctx->global_ctx->mutex); auto it = ctx->set_rows_pipelines.find(key); if (it != ctx->set_rows_pipelines.end()) { pipeline = it->second; } else { ggml_webgpu_processed_shader processed = ggml_webgpu_preprocess_set_rows_shader(ctx->p, wgsl_set_rows, shader_lib_ctx); - pipeline = ggml_webgpu_create_pipeline(ctx->device, processed.wgsl.c_str(), processed.variant.c_str()); + pipeline = + ggml_webgpu_create_pipeline(ctx->global_ctx->device, processed.wgsl.c_str(), processed.variant.c_str()); pipeline.context = processed.decisions; ctx->set_rows_pipelines.emplace(key, pipeline); } @@ -981,7 +1014,8 @@ static std::optional ggml_webgpu_set_rows(webgpu_context & ctx, threads = src->ne[0] * src->ne[1] * src->ne[2] * src->ne[3]; } uint32_t wg_x = CEIL_DIV(threads, decisions.wg_size); - return ggml_backend_webgpu_build(ctx, pipeline, params, entries, wg_x, 1, error_bufs); + return ggml_backend_webgpu_build(ctx->global_ctx, ctx->param_buf_pool, pipeline, params, entries, wg_x, 1, + error_bufs); } static webgpu_command ggml_webgpu_get_rows(webgpu_context & ctx, @@ -1023,7 +1057,7 @@ static webgpu_command ggml_webgpu_get_rows(webgpu_context & ctx, uint32_t vectorized = src->type == GGML_TYPE_F32 && dst->ne[0] % 4 == 0; webgpu_pipeline pipeline = ctx->get_rows_pipelines[src->type][vectorized]; - return ggml_backend_webgpu_build(ctx, pipeline, params, entries, wg_x); + return ggml_backend_webgpu_build(ctx->global_ctx, ctx->param_buf_pool, pipeline, params, entries, wg_x); } static webgpu_command ggml_webgpu_mul_mat(webgpu_context & ctx, @@ -1098,19 +1132,21 @@ static webgpu_command ggml_webgpu_mul_mat(webgpu_context & ctx, uint32_t batches = dst->ne[2] * dst->ne[3]; uint32_t output_groups = CEIL_DIV(dst->ne[0], WEBGPU_MUL_MAT_VEC_OUTPUTS_PER_WG); uint32_t total_wg = output_groups * batches; - wg_x = total_wg % ctx->limits.maxComputeWorkgroupsPerDimension; - wg_y = CEIL_DIV(total_wg, ctx->limits.maxComputeWorkgroupsPerDimension); + wg_x = total_wg % ctx->global_ctx->capabilities.limits.maxComputeWorkgroupsPerDimension; + wg_y = CEIL_DIV(total_wg, ctx->global_ctx->capabilities.limits.maxComputeWorkgroupsPerDimension); } else { pipeline = ctx->mul_mat_pipelines[src0->type][src1->type][vectorized]; uint32_t wg_m; uint32_t wg_n; #ifndef __EMSCRIPTEN__ - if (ctx->supports_subgroup_matrix) { + if (ctx->global_ctx->capabilities.supports_subgroup_matrix) { // The total number of subgroups/workgroups needed per matrix. - uint32_t wg_m_sg_tile = WEBGPU_MUL_MAT_SUBGROUP_M * WEBGPU_MUL_MAT_SUBGROUP_MATRIX_M * ctx->sg_mat_m; + uint32_t wg_m_sg_tile = WEBGPU_MUL_MAT_SUBGROUP_M * WEBGPU_MUL_MAT_SUBGROUP_MATRIX_M * + ctx->global_ctx->capabilities.sg_mat_m; wg_m = CEIL_DIV(dst->ne[0], wg_m_sg_tile); - uint32_t wg_n_sg_tile = WEBGPU_MUL_MAT_SUBGROUP_N * WEBGPU_MUL_MAT_SUBGROUP_MATRIX_N * ctx->sg_mat_n; - wg_n = CEIL_DIV(dst->ne[1], wg_n_sg_tile); + uint32_t wg_n_sg_tile = WEBGPU_MUL_MAT_SUBGROUP_N * WEBGPU_MUL_MAT_SUBGROUP_MATRIX_N * + ctx->global_ctx->capabilities.sg_mat_n; + wg_n = CEIL_DIV(dst->ne[1], wg_n_sg_tile); } else { #endif uint32_t tile_m_s = WEBGPU_MUL_MAT_TILE_M * WEBGPU_MUL_MAT_WG_SIZE_M; @@ -1124,9 +1160,10 @@ static webgpu_command ggml_webgpu_mul_mat(webgpu_context & ctx, wg_x = wg_m * wg_n * dst->ne[2] * dst->ne[3]; } } - return ggml_backend_webgpu_build(ctx, pipeline, params, entries, wg_x, wg_y); + return ggml_backend_webgpu_build(ctx->global_ctx, ctx->param_buf_pool, pipeline, params, entries, wg_x, wg_y); } +#ifndef __EMSCRIPTEN__ static webgpu_command ggml_webgpu_flash_attn(webgpu_context & ctx, ggml_tensor * Q, ggml_tensor * K, @@ -1210,8 +1247,8 @@ static webgpu_command ggml_webgpu_flash_attn(webgpu_context & ctx, .offset = ggml_webgpu_tensor_align_offset(ctx, dst), .size = ggml_webgpu_tensor_binding_size(ctx, dst) }); - bool kv_direct = - (K->type == GGML_TYPE_F16) && (Q->ne[0] % ctx->sg_mat_k == 0) && (K->ne[1] % GGML_WEBGPU_KV_SEQ_PAD == 0); + bool kv_direct = (K->type == GGML_TYPE_F16) && (Q->ne[0] % ctx->global_ctx->capabilities.sg_mat_k == 0) && + (K->ne[1] % GGML_WEBGPU_KV_SEQ_PAD == 0); ggml_webgpu_flash_attn_pipeline_key key = { .kv_type = K->type, @@ -1223,25 +1260,27 @@ static webgpu_command ggml_webgpu_flash_attn(webgpu_context & ctx, .uses_logit_softcap = logit_softcap != 0.0f, }; - webgpu_pipeline pipeline; + webgpu_pipeline pipeline; // TODO: remove guard once pipeline caches are per-thread { - std::lock_guard lock(ctx->mutex); + std::lock_guard lock(ctx->global_ctx->mutex); auto it = ctx->flash_attn_pipelines.find(key); if (it != ctx->flash_attn_pipelines.end()) { - pipeline = it->second; + pipeline = it->second; } else { - ggml_webgpu_flash_attn_shader_lib_context shader_lib_ctx = { .key = key, - .sg_mat_m = ctx->sg_mat_m, - .sg_mat_n = ctx->sg_mat_n, - .sg_mat_k = ctx->sg_mat_k, - .wg_mem_limit_bytes = - ctx->limits.maxComputeWorkgroupStorageSize, - .max_subgroup_size = ctx->max_subgroup_size }; + ggml_webgpu_flash_attn_shader_lib_context shader_lib_ctx = { + .key = key, + .sg_mat_m = ctx->global_ctx->capabilities.sg_mat_m, + .sg_mat_n = ctx->global_ctx->capabilities.sg_mat_n, + .sg_mat_k = ctx->global_ctx->capabilities.sg_mat_k, + .wg_mem_limit_bytes = ctx->global_ctx->capabilities.limits.maxComputeWorkgroupStorageSize, + .max_subgroup_size = ctx->global_ctx->capabilities.max_subgroup_size + }; ggml_webgpu_processed_shader processed = ggml_webgpu_preprocess_flash_attn_shader(ctx->p, wgsl_flash_attn, shader_lib_ctx); - pipeline = ggml_webgpu_create_pipeline(ctx->device, processed.wgsl.c_str(), processed.variant.c_str()); + pipeline = + ggml_webgpu_create_pipeline(ctx->global_ctx->device, processed.wgsl.c_str(), processed.variant.c_str()); pipeline.context = processed.decisions; ctx->flash_attn_pipelines.emplace(key, pipeline); } @@ -1250,11 +1289,11 @@ static webgpu_command ggml_webgpu_flash_attn(webgpu_context & ctx, ggml_webgpu_flash_attn_shader_decisions decisions = *static_cast(pipeline.context); - uint32_t wg_per_head = CEIL_DIV(Q->ne[1], decisions.q_tile); uint32_t wg_x = wg_per_head * Q->ne[2] * Q->ne[3]; // wg per head * number of heads * number of batches - return ggml_backend_webgpu_build(ctx, pipeline, params, entries, wg_x); + return ggml_backend_webgpu_build(ctx->global_ctx, ctx->param_buf_pool, pipeline, params, entries, wg_x); } +#endif static webgpu_command ggml_webgpu_unary_op(webgpu_context & ctx, ggml_tensor * src, ggml_tensor * dst) { bool is_unary = dst->op == GGML_OP_UNARY; @@ -1264,21 +1303,22 @@ static webgpu_command ggml_webgpu_unary_op(webgpu_context & ctx, ggml_tensor * s ggml_webgpu_unary_pipeline_key pipeline_key = { .type = dst->type, .op = op, .is_unary = is_unary, .inplace = inplace }; - ggml_webgpu_unary_shader_lib_context shader_lib_ctx = { .key = pipeline_key, - .max_wg_size = - ctx->limits.maxComputeInvocationsPerWorkgroup }; + ggml_webgpu_unary_shader_lib_context shader_lib_ctx = { + .key = pipeline_key, .max_wg_size = ctx->global_ctx->capabilities.limits.maxComputeInvocationsPerWorkgroup + }; webgpu_pipeline pipeline; { // TODO: remove guard once pipeline caches are per-thread - std::lock_guard lock(ctx->mutex); + std::lock_guard lock(ctx->global_ctx->mutex); auto it = ctx->unary_pipelines.find(pipeline_key); if (it != ctx->unary_pipelines.end()) { pipeline = it->second; } else { ggml_webgpu_processed_shader processed = ggml_webgpu_preprocess_unary_shader(ctx->p, wgsl_unary, shader_lib_ctx); - pipeline = ggml_webgpu_create_pipeline(ctx->device, processed.wgsl.c_str(), processed.variant.c_str()); + pipeline = + ggml_webgpu_create_pipeline(ctx->global_ctx->device, processed.wgsl.c_str(), processed.variant.c_str()); pipeline.context = processed.decisions; ctx->unary_pipelines.emplace(pipeline_key, pipeline); } @@ -1346,7 +1386,7 @@ static webgpu_command ggml_webgpu_unary_op(webgpu_context & ctx, ggml_tensor * s } uint32_t wg_x = CEIL_DIV(ne, decisions.wg_size); - return ggml_backend_webgpu_build(ctx, pipeline, params, entries, wg_x); + return ggml_backend_webgpu_build(ctx->global_ctx, ctx->param_buf_pool, pipeline, params, entries, wg_x); } static webgpu_command ggml_webgpu_binary_op(webgpu_context & ctx, @@ -1391,7 +1431,7 @@ static webgpu_command ggml_webgpu_binary_op(webgpu_context & ctx, } uint32_t wg_x = CEIL_DIV(ggml_nelements(dst), WEBGPU_MAX_WG_SIZE); - return ggml_backend_webgpu_build(ctx, pipeline, params, entries, wg_x); + return ggml_backend_webgpu_build(ctx->global_ctx, ctx->param_buf_pool, pipeline, params, entries, wg_x); } static webgpu_command ggml_webgpu_rms_norm(webgpu_context & ctx, ggml_tensor * src, ggml_tensor * dst) { @@ -1426,7 +1466,8 @@ static webgpu_command ggml_webgpu_rms_norm(webgpu_context & ctx, ggml_tensor * s .size = ggml_webgpu_tensor_binding_size(ctx, dst) }); } - return ggml_backend_webgpu_build(ctx, ctx->rms_norm_pipelines[inplace], params, entries, ggml_nrows(src)); + return ggml_backend_webgpu_build(ctx->global_ctx, ctx->param_buf_pool, ctx->rms_norm_pipelines[inplace], params, + entries, ggml_nrows(src)); } static webgpu_command ggml_webgpu_rope(webgpu_context & ctx, @@ -1513,7 +1554,7 @@ static webgpu_command ggml_webgpu_rope(webgpu_context & ctx, webgpu_pipeline pipeline = ctx->rope_pipelines[dst->type][has_freq_factor][inplace]; uint32_t wg_x = CEIL_DIV(ggml_nelements(dst), WEBGPU_MAX_WG_SIZE); - return ggml_backend_webgpu_build(ctx, pipeline, params, entries, wg_x); + return ggml_backend_webgpu_build(ctx->global_ctx, ctx->param_buf_pool, pipeline, params, entries, wg_x); } static webgpu_command ggml_webgpu_glu(webgpu_context & ctx, ggml_tensor * src0, ggml_tensor * src1, ggml_tensor * dst) { @@ -1565,7 +1606,7 @@ static webgpu_command ggml_webgpu_glu(webgpu_context & ctx, ggml_tensor * src0, webgpu_pipeline pipeline = ctx->glu_pipelines[ggml_get_glu_op(dst)][dst->type][split]; uint32_t wg_x = CEIL_DIV(ggml_nelements(dst), WEBGPU_MAX_WG_SIZE); - return ggml_backend_webgpu_build(ctx, pipeline, params, entries, wg_x); + return ggml_backend_webgpu_build(ctx->global_ctx, ctx->param_buf_pool, pipeline, params, entries, wg_x); } static webgpu_command ggml_webgpu_scale(webgpu_context & ctx, ggml_tensor * src, ggml_tensor * dst) { @@ -1602,7 +1643,8 @@ static webgpu_command ggml_webgpu_scale(webgpu_context & ctx, ggml_tensor * src, } uint32_t wg_x = CEIL_DIV(ggml_nelements(dst), WEBGPU_MAX_WG_SIZE); - return ggml_backend_webgpu_build(ctx, ctx->scale_pipelines[inplace], params, entries, wg_x); + return ggml_backend_webgpu_build(ctx->global_ctx, ctx->param_buf_pool, ctx->scale_pipelines[inplace], params, + entries, wg_x); } static webgpu_command ggml_webgpu_soft_max(webgpu_context & ctx, @@ -1674,7 +1716,8 @@ static webgpu_command ggml_webgpu_soft_max(webgpu_context & ctx, .size = ggml_webgpu_tensor_binding_size(ctx, dst) }); } - return ggml_backend_webgpu_build(ctx, ctx->soft_max_pipelines[mask_type][has_sink][inplace], params, entries, + return ggml_backend_webgpu_build(ctx->global_ctx, ctx->param_buf_pool, + ctx->soft_max_pipelines[mask_type][has_sink][inplace], params, entries, ggml_nrows(dst)); } @@ -1696,25 +1739,26 @@ static webgpu_command ggml_webgpu_argmax(webgpu_context & ctx, ggml_tensor * src ggml_webgpu_generic_shader_lib_context shader_lib_ctx = { .vec4 = src->ne[0] % 4 == 0, - .max_wg_size = ctx->limits.maxComputeInvocationsPerWorkgroup, + .max_wg_size = ctx->global_ctx->capabilities.limits.maxComputeInvocationsPerWorkgroup, }; webgpu_pipeline pipeline; { // TODO: remove guard once pipeline caches are per-thread - std::lock_guard lock(ctx->mutex); + std::lock_guard lock(ctx->global_ctx->mutex); auto it = ctx->argmax_pipelines.find(shader_lib_ctx.vec4); if (it != ctx->argmax_pipelines.end()) { pipeline = it->second; } else { ggml_webgpu_processed_shader processed = ggml_webgpu_preprocess_generic_shader(ctx->p, wgsl_argmax, shader_lib_ctx, "argmax"); - pipeline = ggml_webgpu_create_pipeline(ctx->device, processed.wgsl.c_str(), processed.variant.c_str()); + pipeline = + ggml_webgpu_create_pipeline(ctx->global_ctx->device, processed.wgsl.c_str(), processed.variant.c_str()); ctx->argmax_pipelines.emplace(shader_lib_ctx.vec4, pipeline); } } uint32_t wg_x = ggml_nelements(dst); - return ggml_backend_webgpu_build(ctx, pipeline, params, entries, wg_x); + return ggml_backend_webgpu_build(ctx->global_ctx, ctx->param_buf_pool, pipeline, params, entries, wg_x); } static webgpu_command ggml_webgpu_argsort(webgpu_context & ctx, ggml_tensor * src, ggml_tensor * dst) { @@ -1722,13 +1766,13 @@ static webgpu_command ggml_webgpu_argsort(webgpu_context & ctx, ggml_tensor * sr // ascending order is 0, descending order is 1 const int32_t order = is_top_k ? (int32_t) GGML_SORT_ORDER_DESC : (int32_t) ggml_get_op_params_i32(dst, 0); - ggml_webgpu_argsort_shader_lib_context shader_lib_ctx = { .max_wg_size = - ctx->limits.maxComputeInvocationsPerWorkgroup, - .wg_mem_limit_bytes = - ctx->limits.maxComputeWorkgroupStorageSize, - .order = order }; + ggml_webgpu_argsort_shader_lib_context shader_lib_ctx = { + .max_wg_size = ctx->global_ctx->capabilities.limits.maxComputeInvocationsPerWorkgroup, + .wg_mem_limit_bytes = ctx->global_ctx->capabilities.limits.maxComputeWorkgroupStorageSize, + .order = order + }; - std::lock_guard lock(ctx->mutex); + std::lock_guard lock(ctx->global_ctx->mutex); webgpu_pipeline argsort_pipeline; auto it = ctx->argsort_pipelines.find(order); if (it != ctx->argsort_pipelines.end()) { @@ -1736,7 +1780,8 @@ static webgpu_command ggml_webgpu_argsort(webgpu_context & ctx, ggml_tensor * sr } else { ggml_webgpu_processed_shader processed = ggml_webgpu_preprocess_argsort_shader(ctx->p, wgsl_argsort, shader_lib_ctx); - argsort_pipeline = ggml_webgpu_create_pipeline(ctx->device, processed.wgsl.c_str(), processed.variant.c_str()); + argsort_pipeline = + ggml_webgpu_create_pipeline(ctx->global_ctx->device, processed.wgsl.c_str(), processed.variant.c_str()); argsort_pipeline.context = processed.decisions; ctx->argsort_pipelines.emplace(order, argsort_pipeline); } @@ -1751,7 +1796,7 @@ static webgpu_command ggml_webgpu_argsort(webgpu_context & ctx, ggml_tensor * sr ggml_webgpu_processed_shader processed = ggml_webgpu_preprocess_argsort_merge_shader(ctx->p, wgsl_argsort_merge, shader_lib_ctx); argsort_merge_pipeline = - ggml_webgpu_create_pipeline(ctx->device, processed.wgsl.c_str(), processed.variant.c_str()); + ggml_webgpu_create_pipeline(ctx->global_ctx->device, processed.wgsl.c_str(), processed.variant.c_str()); argsort_merge_pipeline.context = processed.decisions; ctx->argsort_merge_pipelines.emplace(order, argsort_merge_pipeline); } @@ -1780,9 +1825,10 @@ static webgpu_command ggml_webgpu_argsort(webgpu_context & ctx, ggml_tensor * sr const bool start_in_tmp = (merge_passes % 2) == 1; - const size_t dst_offset = ggml_webgpu_tensor_offset(dst); - const size_t idx_nbytes = out_ne0 * ggml_nrows(dst) * sizeof(int32_t); - const size_t tmp_offset = ROUNDUP_POW2(dst_offset + idx_nbytes, ctx->limits.minStorageBufferOffsetAlignment); + const size_t dst_offset = ggml_webgpu_tensor_offset(dst); + const size_t idx_nbytes = out_ne0 * ggml_nrows(dst) * sizeof(int32_t); + const size_t tmp_offset = + ROUNDUP_POW2(dst_offset + idx_nbytes, ctx->global_ctx->capabilities.limits.minStorageBufferOffsetAlignment); const size_t tmp_binding_size = ROUNDUP_POW2(idx_nbytes, WEBGPU_STORAGE_BUF_BINDING_MULT); const size_t dst_binding_size = ROUNDUP_POW2(idx_nbytes + ggml_webgpu_tensor_misalignment(ctx, dst), WEBGPU_STORAGE_BUF_BINDING_MULT); @@ -1813,10 +1859,10 @@ static webgpu_command ggml_webgpu_argsort(webgpu_context & ctx, ggml_tensor * sr }; const uint32_t total_wg_init = npr * nrows; - const uint32_t max_wg = ctx->limits.maxComputeWorkgroupsPerDimension; - const uint32_t wg_x_init = std::min(total_wg_init, max_wg); - const uint32_t wg_y_init = CEIL_DIV(total_wg_init, wg_x_init); - std::vector init_entries = { + const uint32_t max_wg = ctx->global_ctx->capabilities.limits.maxComputeWorkgroupsPerDimension; + const uint32_t wg_x_init = std::min(total_wg_init, max_wg); + const uint32_t wg_y_init = CEIL_DIV(total_wg_init, wg_x_init); + std::vector init_entries = { { .binding = 0, .buffer = ggml_webgpu_tensor_buf(src), .offset = ggml_webgpu_tensor_align_offset(ctx, src), @@ -1830,7 +1876,8 @@ static webgpu_command ggml_webgpu_argsort(webgpu_context & ctx, ggml_tensor * sr workgroups_list.push_back({ wg_x_init, wg_y_init }); if (merge_passes == 0) { - return ggml_backend_webgpu_build_multi(ctx, pipelines, params_list, entries_list, workgroups_list); + return ggml_backend_webgpu_build_multi(ctx->global_ctx, ctx->param_buf_pool, pipelines, params_list, + entries_list, workgroups_list); } bool in_is_tmp = start_in_tmp; @@ -1891,7 +1938,8 @@ static webgpu_command ggml_webgpu_argsort(webgpu_context & ctx, ggml_tensor * sr in_is_tmp = !in_is_tmp; } - return ggml_backend_webgpu_build_multi(ctx, pipelines, params_list, entries_list, workgroups_list); + return ggml_backend_webgpu_build_multi(ctx->global_ctx, ctx->param_buf_pool, pipelines, params_list, entries_list, + workgroups_list); } static webgpu_command ggml_webgpu_cumsum(webgpu_context & ctx, ggml_tensor * src, ggml_tensor * dst) { @@ -1912,24 +1960,25 @@ static webgpu_command ggml_webgpu_cumsum(webgpu_context & ctx, ggml_tensor * src ggml_webgpu_generic_shader_lib_context shader_lib_ctx = { .vec4 = false, - .max_wg_size = ctx->limits.maxComputeInvocationsPerWorkgroup, + .max_wg_size = ctx->global_ctx->capabilities.limits.maxComputeInvocationsPerWorkgroup, }; webgpu_pipeline pipeline; // TODO: remove guard once pipeline caches are per-thread { - std::lock_guard lock(ctx->mutex); + std::lock_guard lock(ctx->global_ctx->mutex); auto it = ctx->cumsum_pipelines.find(1); if (it != ctx->cumsum_pipelines.end()) { pipeline = it->second; } else { ggml_webgpu_processed_shader processed = ggml_webgpu_preprocess_generic_shader(ctx->p, wgsl_cumsum, shader_lib_ctx, "cumsum"); - pipeline = ggml_webgpu_create_pipeline(ctx->device, processed.wgsl.c_str(), processed.variant.c_str()); + pipeline = + ggml_webgpu_create_pipeline(ctx->global_ctx->device, processed.wgsl.c_str(), processed.variant.c_str()); ctx->cumsum_pipelines.emplace(1, pipeline); } } uint32_t wg_x = ggml_nrows(dst); - return ggml_backend_webgpu_build(ctx, pipeline, params, entries, wg_x); + return ggml_backend_webgpu_build(ctx->global_ctx, ctx->param_buf_pool, pipeline, params, entries, wg_x); } static webgpu_command ggml_webgpu_sum_rows(webgpu_context & ctx, ggml_tensor * src, ggml_tensor * dst) { @@ -1956,25 +2005,26 @@ static webgpu_command ggml_webgpu_sum_rows(webgpu_context & ctx, ggml_tensor * s ggml_webgpu_generic_shader_lib_context shader_lib_ctx = { .vec4 = false, - .max_wg_size = ctx->limits.maxComputeInvocationsPerWorkgroup, + .max_wg_size = ctx->global_ctx->capabilities.limits.maxComputeInvocationsPerWorkgroup, }; webgpu_pipeline pipeline; { // TODO: remove guard once pipeline caches are per-thread - std::lock_guard lock(ctx->mutex); + std::lock_guard lock(ctx->global_ctx->mutex); auto it = ctx->sum_rows_pipelines.find(1); if (it != ctx->sum_rows_pipelines.end()) { pipeline = it->second; } else { ggml_webgpu_processed_shader processed = ggml_webgpu_preprocess_generic_shader(ctx->p, wgsl_sum_rows, shader_lib_ctx, "sum_rows"); - pipeline = ggml_webgpu_create_pipeline(ctx->device, processed.wgsl.c_str(), processed.variant.c_str()); + pipeline = + ggml_webgpu_create_pipeline(ctx->global_ctx->device, processed.wgsl.c_str(), processed.variant.c_str()); ctx->sum_rows_pipelines.emplace(1, pipeline); } } uint32_t wg_x = total_sum ? 1 : ggml_nrows(dst); - return ggml_backend_webgpu_build(ctx, pipeline, params, entries, wg_x); + return ggml_backend_webgpu_build(ctx->global_ctx, ctx->param_buf_pool, pipeline, params, entries, wg_x); } // Returns the encoded command, or std::nullopt if the operation is a no-op @@ -2009,7 +2059,11 @@ static std::optional ggml_webgpu_encode_node(webgpu_context ctx, case GGML_OP_MUL_MAT: return ggml_webgpu_mul_mat(ctx, src0, src1, node); case GGML_OP_FLASH_ATTN_EXT: +#ifndef __EMSCRIPTEN__ return ggml_webgpu_flash_attn(ctx, src0, src1, src2, node->src[3], node->src[4], node); +#else + return std::nullopt; +#endif case GGML_OP_ADD: { int inplace = ggml_webgpu_tensor_equal(src0, node); @@ -2070,12 +2124,12 @@ static std::optional ggml_webgpu_encode_node(webgpu_context ctx, static ggml_status ggml_backend_webgpu_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) { WEBGPU_LOG_DEBUG("ggml_backend_webgpu_graph_compute(" << cgraph->n_nodes << " nodes)"); - ggml_backend_webgpu_context * backend_ctx = static_cast(backend->context); + ggml_backend_webgpu_context * backend_ctx = (ggml_backend_webgpu_context *) backend->context; webgpu_context ctx = backend_ctx->webgpu_ctx; WEBGPU_CPU_PROFILE_TOTAL_START(graph_compute); - ctx->inflight_threads++; + ctx->global_ctx->inflight_threads++; std::vector commands; std::vector futures; @@ -2084,25 +2138,27 @@ static ggml_status ggml_backend_webgpu_graph_compute(ggml_backend_t backend, str commands.push_back(*cmd); } // compute the batch size based on the number of inflight threads - uint32_t inflight_threads = ctx->inflight_threads; + uint32_t inflight_threads = ctx->global_ctx->inflight_threads; uint32_t batch_size = std::min(std::max(1u, WEBGPU_NUM_PARAM_BUFS / std::max(inflight_threads, 1u)), WEBGPU_COMMAND_SUBMIT_BATCH_SIZE); if (commands.size() >= batch_size) { - futures.push_back(ggml_backend_webgpu_submit(ctx, commands)); + futures.push_back(ggml_backend_webgpu_submit(ctx->global_ctx, commands, ctx->param_buf_pool, + &ctx->set_rows_error_buf_pool)); // Process events and check for completed submissions - ctx->instance.ProcessEvents(); - ggml_backend_webgpu_wait(ctx, futures, false); + ctx->global_ctx->instance.ProcessEvents(); + ggml_backend_webgpu_wait(ctx->global_ctx, futures, false); commands.clear(); } } if (!commands.empty()) { - webgpu_submission_futures new_futures = ggml_backend_webgpu_submit(ctx, commands); + webgpu_submission_futures new_futures = + ggml_backend_webgpu_submit(ctx->global_ctx, commands, ctx->param_buf_pool, &ctx->set_rows_error_buf_pool); futures.push_back(new_futures); } - ggml_backend_webgpu_wait(ctx, futures); - ctx->inflight_threads--; - WEBGPU_CPU_PROFILE_TOTAL_END(graph_compute, ctx); + ggml_backend_webgpu_wait(ctx->global_ctx, futures); + ctx->global_ctx->inflight_threads--; + WEBGPU_CPU_PROFILE_TOTAL_END(graph_compute, ctx->global_ctx); return GGML_STATUS_SUCCESS; } @@ -2159,8 +2215,8 @@ static void ggml_backend_webgpu_buffer_memset_tensor(ggml_backend_buffer_t buffe // This is a trick to set all bytes of a u32 to the same 1 byte value. uint32_t val32 = (uint32_t) value * 0x01010101; - ggml_backend_webgpu_buffer_memset(buf_ctx->webgpu_ctx, buf_ctx->buffer, val32, total_offset, size); - WEBGPU_CPU_PROFILE_TOTAL_END(memset_tensor, buf_ctx->webgpu_ctx); + ggml_backend_webgpu_buffer_memset(buf_ctx->global_ctx, buf_ctx->buffer, val32, total_offset, size); + WEBGPU_CPU_PROFILE_TOTAL_END(memset_tensor, buf_ctx->global_ctx); } static void ggml_backend_webgpu_buffer_set_tensor(ggml_backend_buffer_t buffer, @@ -2169,15 +2225,14 @@ static void ggml_backend_webgpu_buffer_set_tensor(ggml_backend_buffer_t buffer, size_t offset, size_t size) { WEBGPU_CPU_PROFILE_TOTAL_START(set_tensor); - ggml_backend_webgpu_buffer_context * buf_ctx = (ggml_backend_webgpu_buffer_context *) buffer->context; - webgpu_context webgpu_ctx = buf_ctx->webgpu_ctx; + ggml_backend_webgpu_buffer_context * buf_ctx = (ggml_backend_webgpu_buffer_context *) buffer->context; WEBGPU_LOG_DEBUG("ggml_backend_webgpu_buffer_set_tensor(" << buf_ctx->label << ", " << tensor << ", " << data << ", " << offset << ", " << size << ")"); size_t total_offset = webgpu_tensor_offset(tensor) + tensor->view_offs + offset; - webgpu_ctx->queue.WriteBuffer(buf_ctx->buffer, total_offset, data, (size / 4) * 4); + buf_ctx->global_ctx->queue.WriteBuffer(buf_ctx->buffer, total_offset, data, (size / 4) * 4); if (size % 4 != 0) { // If size is not a multiple of 4, we need to memset the remaining bytes @@ -2190,21 +2245,21 @@ static void ggml_backend_webgpu_buffer_set_tensor(ggml_backend_buffer_t buffer, ((uint8_t *) &val32)[i] = ((const uint8_t *) data)[size - remaining_size + i]; } // memset the remaining bytes - ggml_backend_webgpu_buffer_memset(webgpu_ctx, buf_ctx->buffer, val32, total_offset + (size - remaining_size), - remaining_size); + ggml_backend_webgpu_buffer_memset(buf_ctx->global_ctx, buf_ctx->buffer, val32, + total_offset + (size - remaining_size), remaining_size); } else { // wait for WriteBuffer to complete - webgpu_ctx->instance.WaitAny( - webgpu_ctx->queue.OnSubmittedWorkDone(wgpu::CallbackMode::AllowSpontaneous, + buf_ctx->global_ctx->instance.WaitAny(buf_ctx->global_ctx->queue.OnSubmittedWorkDone( + wgpu::CallbackMode::AllowSpontaneous, [](wgpu::QueueWorkDoneStatus status, wgpu::StringView message) { if (status != wgpu::QueueWorkDoneStatus::Success) { GGML_LOG_ERROR("ggml_webgpu: Failed to submit commands: %s\n", std::string(message).c_str()); } }), - UINT64_MAX); + UINT64_MAX); } - WEBGPU_CPU_PROFILE_TOTAL_END(set_tensor, webgpu_ctx); + WEBGPU_CPU_PROFILE_TOTAL_END(set_tensor, buf_ctx->global_ctx); } static void ggml_backend_webgpu_buffer_get_tensor(ggml_backend_buffer_t buffer, @@ -2216,8 +2271,7 @@ static void ggml_backend_webgpu_buffer_get_tensor(ggml_backend_buffer_t buffer, ggml_backend_webgpu_buffer_context * buf_ctx = (ggml_backend_webgpu_buffer_context *) buffer->context; WEBGPU_LOG_DEBUG("ggml_backend_webgpu_buffer_get_tensor(" << buf_ctx->label << ", " << tensor << ", " << data << ", " << offset << ", " << size << ")"); - webgpu_context webgpu_ctx = buf_ctx->webgpu_ctx; - wgpu::Device device = webgpu_ctx->device; + wgpu::Device device = buf_ctx->global_ctx->device; size_t total_offset = webgpu_tensor_offset(tensor) + tensor->view_offs + offset; @@ -2227,42 +2281,45 @@ static void ggml_backend_webgpu_buffer_get_tensor(ggml_backend_buffer_t buffer, final_size = size + (4 - (size % 4)); } - std::lock_guard lock(webgpu_ctx->mutex); + std::lock_guard lock(buf_ctx->global_ctx->mutex); - if (webgpu_ctx->get_tensor_staging_buf == nullptr || webgpu_ctx->get_tensor_staging_buf.GetSize() < final_size) { + if (buf_ctx->global_ctx->get_tensor_staging_buf == nullptr || + buf_ctx->global_ctx->get_tensor_staging_buf.GetSize() < final_size) { // Create a new staging buffer if it doesn't exist or is too small - if (webgpu_ctx->get_tensor_staging_buf) { - webgpu_ctx->get_tensor_staging_buf.Destroy(); + if (buf_ctx->global_ctx->get_tensor_staging_buf) { + buf_ctx->global_ctx->get_tensor_staging_buf.Destroy(); } - ggml_webgpu_create_buffer(device, webgpu_ctx->get_tensor_staging_buf, final_size, + ggml_webgpu_create_buffer(device, buf_ctx->global_ctx->get_tensor_staging_buf, final_size, wgpu::BufferUsage::CopyDst | wgpu::BufferUsage::MapRead, "get_tensor_staging_buf"); } // Copy the data from the buffer to the staging buffer wgpu::CommandEncoder encoder = device.CreateCommandEncoder(); - encoder.CopyBufferToBuffer(buf_ctx->buffer, total_offset, webgpu_ctx->get_tensor_staging_buf, 0, final_size); + encoder.CopyBufferToBuffer(buf_ctx->buffer, total_offset, buf_ctx->global_ctx->get_tensor_staging_buf, 0, + final_size); wgpu::CommandBuffer commands = encoder.Finish(); // Submit the command buffer to the queue - webgpu_ctx->queue.Submit(1, &commands); + buf_ctx->global_ctx->queue.Submit(1, &commands); // Map the staging buffer to read the data - ggml_backend_webgpu_map_buffer(webgpu_ctx, webgpu_ctx->get_tensor_staging_buf, wgpu::MapMode::Read, 0, final_size); + ggml_backend_webgpu_map_buffer(buf_ctx->global_ctx, buf_ctx->global_ctx->get_tensor_staging_buf, + wgpu::MapMode::Read, 0, final_size); // Must specify size here since the staging buffer might be larger than the tensor size - const void * mapped_range = webgpu_ctx->get_tensor_staging_buf.GetConstMappedRange(0, final_size); + const void * mapped_range = buf_ctx->global_ctx->get_tensor_staging_buf.GetConstMappedRange(0, final_size); // Copy the data from the mapped range to the output buffer std::memcpy(data, mapped_range, size); - webgpu_ctx->get_tensor_staging_buf.Unmap(); - WEBGPU_CPU_PROFILE_TOTAL_END(get_tensor, webgpu_ctx); + buf_ctx->global_ctx->get_tensor_staging_buf.Unmap(); + WEBGPU_CPU_PROFILE_TOTAL_END(get_tensor, buf_ctx->global_ctx); } static void ggml_backend_webgpu_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) { WEBGPU_LOG_DEBUG("ggml_backend_webgpu_buffer_clear(" << buffer << ", " << (uint32_t) value << ")"); WEBGPU_CPU_PROFILE_TOTAL_START(clear); ggml_backend_webgpu_buffer_context * buf_ctx = (ggml_backend_webgpu_buffer_context *) buffer->context; - ggml_backend_webgpu_buffer_memset(buf_ctx->webgpu_ctx, buf_ctx->buffer, value, 0, buffer->size); - WEBGPU_CPU_PROFILE_TOTAL_END(clear, buf_ctx->webgpu_ctx); + ggml_backend_webgpu_buffer_memset(buf_ctx->global_ctx, buf_ctx->buffer, value, 0, buffer->size); + WEBGPU_CPU_PROFILE_TOTAL_END(clear, buf_ctx->global_ctx); } static ggml_backend_buffer_i ggml_backend_webgpu_buffer_interface = { @@ -2292,28 +2349,30 @@ static ggml_backend_buffer_t ggml_backend_webgpu_buffer_type_alloc_buffer(ggml_b int buffer_id = buffer_count++; std::string buf_name = "tensor_buf" + std::to_string(buffer_id); WEBGPU_LOG_DEBUG("ggml_backend_webgpu_buffer_type_alloc_buffer_" << buffer_id << ": " << size << " bytes"); - ggml_backend_webgpu_device_context * ctx = static_cast(buft->device->context); - wgpu::Buffer buf; - ggml_webgpu_create_buffer(ctx->webgpu_ctx->device, buf, ROUNDUP_POW2(size, WEBGPU_STORAGE_BUF_BINDING_MULT), + ggml_backend_webgpu_device_context * ctx = static_cast(buft->device->context); + wgpu::Buffer buf; + ggml_webgpu_create_buffer(ctx->webgpu_global_ctx->device, buf, ROUNDUP_POW2(size, WEBGPU_STORAGE_BUF_BINDING_MULT), wgpu::BufferUsage::Storage | wgpu::BufferUsage::CopySrc | wgpu::BufferUsage::CopyDst, buf_name.c_str()); ggml_backend_webgpu_buffer_context * buf_ctx = - new ggml_backend_webgpu_buffer_context(ctx->webgpu_ctx, buf, buf_name); + new ggml_backend_webgpu_buffer_context(buf, buf_name, ctx->webgpu_global_ctx); return ggml_backend_buffer_init(buft, ggml_backend_webgpu_buffer_interface, buf_ctx, size); } static size_t ggml_backend_webgpu_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) { - ggml_backend_webgpu_device_context * ctx = static_cast(buft->device->context); - return ctx->webgpu_ctx->limits.minStorageBufferOffsetAlignment; + ggml_backend_webgpu_device_context * dev_ctx = + static_cast(buft->device->context); + return dev_ctx->webgpu_global_ctx->capabilities.limits.minStorageBufferOffsetAlignment; } // maxBufferSize might be larger, but you can't bind more than maxStorageBufferBindingSize to a single binding. static size_t ggml_backend_webgpu_buffer_type_get_max_size(ggml_backend_buffer_type_t buft) { - ggml_backend_webgpu_device_context * ctx = static_cast(buft->device->context); - return ctx->webgpu_ctx->limits.maxStorageBufferBindingSize; + ggml_backend_webgpu_device_context * dev_ctx = + static_cast(buft->device->context); + return dev_ctx->webgpu_global_ctx->capabilities.limits.maxStorageBufferBindingSize; } static size_t ggml_backend_webgpu_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, @@ -2322,7 +2381,7 @@ static size_t ggml_backend_webgpu_buffer_type_get_alloc_size(ggml_backend_buffer size_t res = ggml_nbytes(tensor); switch (tensor->op) { case GGML_OP_ARGSORT: - res = ROUNDUP_POW2(res * 2 + ctx->webgpu_ctx->limits.minStorageBufferOffsetAlignment, + res = ROUNDUP_POW2(res * 2 + ctx->webgpu_global_ctx->capabilities.limits.minStorageBufferOffsetAlignment, WEBGPU_STORAGE_BUF_BINDING_MULT); break; case GGML_OP_TOP_K: @@ -2330,8 +2389,9 @@ static size_t ggml_backend_webgpu_buffer_type_get_alloc_size(ggml_backend_buffer const ggml_tensor * src0 = tensor->src[0]; if (src0) { const size_t full = sizeof(int32_t) * ggml_nelements(src0); - res = ROUNDUP_POW2(full * 2 + ctx->webgpu_ctx->limits.minStorageBufferOffsetAlignment, - WEBGPU_STORAGE_BUF_BINDING_MULT); + res = ROUNDUP_POW2( + full * 2 + ctx->webgpu_global_ctx->capabilities.limits.minStorageBufferOffsetAlignment, + WEBGPU_STORAGE_BUF_BINDING_MULT); } } break; @@ -2359,7 +2419,7 @@ static void ggml_backend_webgpu_device_get_memory(ggml_backend_dev_t dev, size_t ggml_backend_webgpu_device_context * ctx = static_cast(dev->context); // TODO: for now, return maxBufferSize as both free and total memory // Track https://github.com/gpuweb/gpuweb/issues/5505 for updates. - uint64_t max_buffer_size = ctx->webgpu_ctx->limits.maxBufferSize; + uint64_t max_buffer_size = ctx->webgpu_global_ctx->capabilities.limits.maxBufferSize; // If we're on a 32-bit system, clamp to UINTPTR_MAX #if UINTPTR_MAX < UINT64_MAX uint64_t max_ptr_size = static_cast(UINTPTR_MAX); @@ -2402,66 +2462,67 @@ static std::vector ggml_webgpu_wg_size_entry(uint32_t wg_si return constants; } -static void ggml_webgpu_init_memset_pipeline(webgpu_context & webgpu_ctx) { +static void ggml_webgpu_init_memset_pipeline(webgpu_global_context & ctx) { // we use the maximum workgroup size for the memset pipeline - size_t max_threads = WEBGPU_MAX_WG_SIZE * webgpu_ctx->limits.maxComputeWorkgroupsPerDimension; + size_t max_threads = WEBGPU_MAX_WG_SIZE * ctx->capabilities.limits.maxComputeWorkgroupsPerDimension; // Size the bytes_per_thread so that the largest buffer size can be handled - webgpu_ctx->memset_bytes_per_thread = CEIL_DIV(webgpu_ctx->limits.maxStorageBufferBindingSize, max_threads); + ctx->capabilities.memset_bytes_per_thread = + CEIL_DIV(ctx->capabilities.limits.maxStorageBufferBindingSize, max_threads); std::vector constants(2); - constants[0].key = "wg_size"; - constants[0].value = WEBGPU_MAX_WG_SIZE; - constants[1].key = "bytes_per_thread"; - constants[1].value = webgpu_ctx->memset_bytes_per_thread; - webgpu_ctx->memset_pipelines[0] = ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_memset, "memset", constants); + constants[0].key = "wg_size"; + constants[0].value = WEBGPU_MAX_WG_SIZE; + constants[1].key = "bytes_per_thread"; + constants[1].value = ctx->capabilities.memset_bytes_per_thread; + ctx->memset_pipelines[0] = ggml_webgpu_create_pipeline(ctx->device, wgsl_memset, "memset", constants); } static void ggml_webgpu_init_mul_mat_pipeline(webgpu_context & webgpu_ctx) { // Q4/Q5/Q8 classic quantizations webgpu_ctx->mul_mat_pipelines[GGML_TYPE_Q4_0][GGML_TYPE_F32][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_mat_q4_0_f32, "mul_mat_q4_0_f32"); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_mul_mat_q4_0_f32, "mul_mat_q4_0_f32"); webgpu_ctx->mul_mat_pipelines[GGML_TYPE_Q4_1][GGML_TYPE_F32][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_mat_q4_1_f32, "mul_mat_q4_1_f32"); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_mul_mat_q4_1_f32, "mul_mat_q4_1_f32"); webgpu_ctx->mul_mat_pipelines[GGML_TYPE_Q5_0][GGML_TYPE_F32][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_mat_q5_0_f32, "mul_mat_q5_0_f32"); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_mul_mat_q5_0_f32, "mul_mat_q5_0_f32"); webgpu_ctx->mul_mat_pipelines[GGML_TYPE_Q5_1][GGML_TYPE_F32][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_mat_q5_1_f32, "mul_mat_q5_1_f32"); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_mul_mat_q5_1_f32, "mul_mat_q5_1_f32"); webgpu_ctx->mul_mat_pipelines[GGML_TYPE_Q8_0][GGML_TYPE_F32][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_mat_q8_0_f32, "mul_mat_q8_0_f32"); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_mul_mat_q8_0_f32, "mul_mat_q8_0_f32"); // K-quantizations webgpu_ctx->mul_mat_pipelines[GGML_TYPE_Q2_K][GGML_TYPE_F32][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_mat_q2_k_f32, "mul_mat_q2_k_f32"); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_mul_mat_q2_k_f32, "mul_mat_q2_k_f32"); webgpu_ctx->mul_mat_pipelines[GGML_TYPE_Q3_K][GGML_TYPE_F32][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_mat_q3_k_f32, "mul_mat_q3_k_f32"); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_mul_mat_q3_k_f32, "mul_mat_q3_k_f32"); webgpu_ctx->mul_mat_pipelines[GGML_TYPE_Q4_K][GGML_TYPE_F32][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_mat_q4_k_f32, "mul_mat_q4_k_f32"); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_mul_mat_q4_k_f32, "mul_mat_q4_k_f32"); webgpu_ctx->mul_mat_pipelines[GGML_TYPE_Q5_K][GGML_TYPE_F32][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_mat_q5_k_f32, "mul_mat_q5_k_f32"); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_mul_mat_q5_k_f32, "mul_mat_q5_k_f32"); webgpu_ctx->mul_mat_pipelines[GGML_TYPE_Q6_K][GGML_TYPE_F32][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_mat_q6_k_f32, "mul_mat_q6_k_f32"); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_mul_mat_q6_k_f32, "mul_mat_q6_k_f32"); // IQ quantizations (2-, 3-, 4-bit variants) webgpu_ctx->mul_mat_pipelines[GGML_TYPE_IQ2_XXS][GGML_TYPE_F32][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_mat_iq2_xxs_f32, "mul_mat_iq2_xxs_f32"); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_mul_mat_iq2_xxs_f32, "mul_mat_iq2_xxs_f32"); webgpu_ctx->mul_mat_pipelines[GGML_TYPE_IQ2_XS][GGML_TYPE_F32][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_mat_iq2_xs_f32, "mul_mat_iq2_xs_f32"); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_mul_mat_iq2_xs_f32, "mul_mat_iq2_xs_f32"); webgpu_ctx->mul_mat_pipelines[GGML_TYPE_IQ2_S][GGML_TYPE_F32][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_mat_iq2_s_f32, "mul_mat_iq2_s_f32"); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_mul_mat_iq2_s_f32, "mul_mat_iq2_s_f32"); webgpu_ctx->mul_mat_pipelines[GGML_TYPE_IQ3_XXS][GGML_TYPE_F32][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_mat_iq3_xxs_f32, "mul_mat_iq3_xxs_f32"); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_mul_mat_iq3_xxs_f32, "mul_mat_iq3_xxs_f32"); webgpu_ctx->mul_mat_pipelines[GGML_TYPE_IQ3_S][GGML_TYPE_F32][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_mat_iq3_s_f32, "mul_mat_iq3_s_f32"); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_mul_mat_iq3_s_f32, "mul_mat_iq3_s_f32"); // 1-bit and 4-bit IQ variants webgpu_ctx->mul_mat_pipelines[GGML_TYPE_IQ1_S][GGML_TYPE_F32][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_mat_iq1_s_f32, "mul_mat_iq1_s_f32"); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_mul_mat_iq1_s_f32, "mul_mat_iq1_s_f32"); webgpu_ctx->mul_mat_pipelines[GGML_TYPE_IQ1_M][GGML_TYPE_F32][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_mat_iq1_m_f32, "mul_mat_iq1_m_f32"); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_mul_mat_iq1_m_f32, "mul_mat_iq1_m_f32"); webgpu_ctx->mul_mat_pipelines[GGML_TYPE_IQ4_NL][GGML_TYPE_F32][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_mat_iq4_nl_f32, "mul_mat_iq4_nl_f32"); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_mul_mat_iq4_nl_f32, "mul_mat_iq4_nl_f32"); webgpu_ctx->mul_mat_pipelines[GGML_TYPE_IQ4_XS][GGML_TYPE_F32][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_mat_iq4_xs_f32, "mul_mat_iq4_xs_f32"); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_mul_mat_iq4_xs_f32, "mul_mat_iq4_xs_f32"); std::string proc_mul_mat_f32_f32; std::string proc_mul_mat_f32_f32_vec; @@ -2474,18 +2535,18 @@ static void ggml_webgpu_init_mul_mat_pipeline(webgpu_context & webgpu_ctx) { std::vector mul_mat_constants; #ifndef __EMSCRIPTEN__ - if (webgpu_ctx->supports_subgroup_matrix) { + if (webgpu_ctx->global_ctx->capabilities.supports_subgroup_matrix) { std::map sg_matrix_repls; - sg_matrix_repls["WEBGPU_MAX_SUBGROUP_SIZE"] = std::to_string(webgpu_ctx->max_subgroup_size); + sg_matrix_repls["WEBGPU_MAX_SUBGROUP_SIZE"] = + std::to_string(webgpu_ctx->global_ctx->capabilities.max_subgroup_size); sg_matrix_repls["WEBGPU_TILE_K"] = std::to_string(WEBGPU_MUL_MAT_TILE_K); sg_matrix_repls["WEBGPU_SUBGROUP_M"] = std::to_string(WEBGPU_MUL_MAT_SUBGROUP_M); sg_matrix_repls["WEBGPU_SUBGROUP_N"] = std::to_string(WEBGPU_MUL_MAT_SUBGROUP_N); sg_matrix_repls["WEBGPU_SUBGROUP_MATRIX_M"] = std::to_string(WEBGPU_MUL_MAT_SUBGROUP_MATRIX_M); sg_matrix_repls["WEBGPU_SUBGROUP_MATRIX_N"] = std::to_string(WEBGPU_MUL_MAT_SUBGROUP_MATRIX_N); - sg_matrix_repls["WEBGPU_SG_MAT_M_SIZE"] = std::to_string(webgpu_ctx->sg_mat_m); - sg_matrix_repls["WEBGPU_SG_MAT_N_SIZE"] = std::to_string(webgpu_ctx->sg_mat_n); - sg_matrix_repls["WEBGPU_SG_MAT_K_SIZE"] = std::to_string(webgpu_ctx->sg_mat_k); - + sg_matrix_repls["WEBGPU_SG_MAT_M_SIZE"] = std::to_string(webgpu_ctx->global_ctx->capabilities.sg_mat_m); + sg_matrix_repls["WEBGPU_SG_MAT_N_SIZE"] = std::to_string(webgpu_ctx->global_ctx->capabilities.sg_mat_n); + sg_matrix_repls["WEBGPU_SG_MAT_K_SIZE"] = std::to_string(webgpu_ctx->global_ctx->capabilities.sg_mat_k); proc_mul_mat_f32_f32 = ggml_webgpu_process_shader_repls(wgsl_mul_mat_subgroup_matrix_f32_f32, sg_matrix_repls); proc_mul_mat_f32_f32_vec = ggml_webgpu_process_shader_repls(wgsl_mul_mat_subgroup_matrix_f32_f32_vec, sg_matrix_repls); @@ -2522,21 +2583,21 @@ static void ggml_webgpu_init_mul_mat_pipeline(webgpu_context & webgpu_ctx) { #endif webgpu_ctx->mul_mat_pipelines[GGML_TYPE_F32][GGML_TYPE_F32][0] = ggml_webgpu_create_pipeline( - webgpu_ctx->device, proc_mul_mat_f32_f32.c_str(), "mul_mat_f32_f32", mul_mat_constants); + webgpu_ctx->global_ctx->device, proc_mul_mat_f32_f32.c_str(), "mul_mat_f32_f32", mul_mat_constants); webgpu_ctx->mul_mat_pipelines[GGML_TYPE_F32][GGML_TYPE_F32][1] = ggml_webgpu_create_pipeline( - webgpu_ctx->device, proc_mul_mat_f32_f32_vec.c_str(), "mul_mat_f32_f32_vec", mul_mat_constants); + webgpu_ctx->global_ctx->device, proc_mul_mat_f32_f32_vec.c_str(), "mul_mat_f32_f32_vec", mul_mat_constants); webgpu_ctx->mul_mat_pipelines[GGML_TYPE_F16][GGML_TYPE_F32][0] = ggml_webgpu_create_pipeline( - webgpu_ctx->device, proc_mul_mat_f16_f32.c_str(), "mul_mat_f16_f32", mul_mat_constants); + webgpu_ctx->global_ctx->device, proc_mul_mat_f16_f32.c_str(), "mul_mat_f16_f32", mul_mat_constants); webgpu_ctx->mul_mat_pipelines[GGML_TYPE_F16][GGML_TYPE_F32][1] = ggml_webgpu_create_pipeline( - webgpu_ctx->device, proc_mul_mat_f16_f32_vec.c_str(), "mul_mat_f16_f32_vec", mul_mat_constants); + webgpu_ctx->global_ctx->device, proc_mul_mat_f16_f32_vec.c_str(), "mul_mat_f16_f32_vec", mul_mat_constants); webgpu_ctx->mul_mat_pipelines[GGML_TYPE_F16][GGML_TYPE_F16][0] = ggml_webgpu_create_pipeline( - webgpu_ctx->device, proc_mul_mat_f16_f16.c_str(), "mul_mat_f16_f16", mul_mat_constants); + webgpu_ctx->global_ctx->device, proc_mul_mat_f16_f16.c_str(), "mul_mat_f16_f16", mul_mat_constants); webgpu_ctx->mul_mat_pipelines[GGML_TYPE_F16][GGML_TYPE_F16][1] = ggml_webgpu_create_pipeline( - webgpu_ctx->device, proc_mul_mat_f16_f16_vec.c_str(), "mul_mat_f16_f16_vec", mul_mat_constants); + webgpu_ctx->global_ctx->device, proc_mul_mat_f16_f16_vec.c_str(), "mul_mat_f16_f16_vec", mul_mat_constants); webgpu_ctx->mul_mat_pipelines[GGML_TYPE_Q4_0][GGML_TYPE_F32][0] = ggml_webgpu_create_pipeline( - webgpu_ctx->device, proc_mul_mat_q4_0_f32.c_str(), "mul_mat_q4_0_f32", mul_mat_constants); + webgpu_ctx->global_ctx->device, proc_mul_mat_q4_0_f32.c_str(), "mul_mat_q4_0_f32", mul_mat_constants); webgpu_ctx->mul_mat_pipelines[GGML_TYPE_Q4_0][GGML_TYPE_F32][1] = ggml_webgpu_create_pipeline( - webgpu_ctx->device, proc_mul_mat_q4_0_f32_vec.c_str(), "mul_mat_q4_0_f32_vec", mul_mat_constants); + webgpu_ctx->global_ctx->device, proc_mul_mat_q4_0_f32_vec.c_str(), "mul_mat_q4_0_f32_vec", mul_mat_constants); std::vector mul_mat_vec_constants(3); mul_mat_vec_constants[0].key = "WORKGROUP_SIZE"; @@ -2547,171 +2608,171 @@ static void ggml_webgpu_init_mul_mat_pipeline(webgpu_context & webgpu_ctx) { mul_mat_vec_constants[2].value = WEBGPU_MUL_MAT_VEC_OUTPUTS_PER_WG; webgpu_ctx->mul_mat_vec_pipelines[GGML_TYPE_F32][GGML_TYPE_F32][0] = ggml_webgpu_create_pipeline( - webgpu_ctx->device, wgsl_mul_mat_vec_f32_f32, "mul_mat_vec_f32_f32", mul_mat_vec_constants); + webgpu_ctx->global_ctx->device, wgsl_mul_mat_vec_f32_f32, "mul_mat_vec_f32_f32", mul_mat_vec_constants); webgpu_ctx->mul_mat_vec_pipelines[GGML_TYPE_F32][GGML_TYPE_F32][1] = ggml_webgpu_create_pipeline( - webgpu_ctx->device, wgsl_mul_mat_vec_f32_f32_vec, "mul_mat_vec_f32_f32_vec", mul_mat_vec_constants); + webgpu_ctx->global_ctx->device, wgsl_mul_mat_vec_f32_f32_vec, "mul_mat_vec_f32_f32_vec", mul_mat_vec_constants); webgpu_ctx->mul_mat_vec_pipelines[GGML_TYPE_F16][GGML_TYPE_F32][0] = ggml_webgpu_create_pipeline( - webgpu_ctx->device, wgsl_mul_mat_vec_f16_f32, "mul_mat_vec_f16_f32", mul_mat_vec_constants); + webgpu_ctx->global_ctx->device, wgsl_mul_mat_vec_f16_f32, "mul_mat_vec_f16_f32", mul_mat_vec_constants); webgpu_ctx->mul_mat_vec_pipelines[GGML_TYPE_F16][GGML_TYPE_F32][1] = ggml_webgpu_create_pipeline( - webgpu_ctx->device, wgsl_mul_mat_vec_f16_f32_vec, "mul_mat_vec_f16_f32_vec", mul_mat_vec_constants); + webgpu_ctx->global_ctx->device, wgsl_mul_mat_vec_f16_f32_vec, "mul_mat_vec_f16_f32_vec", mul_mat_vec_constants); webgpu_ctx->mul_mat_vec_pipelines[GGML_TYPE_F16][GGML_TYPE_F16][0] = ggml_webgpu_create_pipeline( - webgpu_ctx->device, wgsl_mul_mat_vec_f16_f16, "mul_mat_vec_f16_f16", mul_mat_vec_constants); + webgpu_ctx->global_ctx->device, wgsl_mul_mat_vec_f16_f16, "mul_mat_vec_f16_f16", mul_mat_vec_constants); webgpu_ctx->mul_mat_vec_pipelines[GGML_TYPE_F16][GGML_TYPE_F16][1] = ggml_webgpu_create_pipeline( - webgpu_ctx->device, wgsl_mul_mat_vec_f16_f16_vec, "mul_mat_vec_f16_f16_vec", mul_mat_vec_constants); + webgpu_ctx->global_ctx->device, wgsl_mul_mat_vec_f16_f16_vec, "mul_mat_vec_f16_f16_vec", mul_mat_vec_constants); webgpu_ctx->mul_mat_vec_pipelines[GGML_TYPE_Q4_0][GGML_TYPE_F32][0] = ggml_webgpu_create_pipeline( - webgpu_ctx->device, wgsl_mul_mat_vec_q4_0_f32, "mul_mat_vec_q4_0_f32", mul_mat_vec_constants); + webgpu_ctx->global_ctx->device, wgsl_mul_mat_vec_q4_0_f32, "mul_mat_vec_q4_0_f32", mul_mat_vec_constants); } static void ggml_webgpu_init_get_rows_pipeline(webgpu_context & webgpu_ctx) { std::vector constants = ggml_webgpu_wg_size_entry(WEBGPU_MAX_WG_SIZE); webgpu_ctx->get_rows_pipelines[GGML_TYPE_F32][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_f32, "get_rows_f32", constants); - webgpu_ctx->get_rows_pipelines[GGML_TYPE_F32][1] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_f32_vec, "get_rows_f32_vec", constants); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_get_rows_f32, "get_rows_f32", constants); + webgpu_ctx->get_rows_pipelines[GGML_TYPE_F32][1] = ggml_webgpu_create_pipeline( + webgpu_ctx->global_ctx->device, wgsl_get_rows_f32_vec, "get_rows_f32_vec", constants); webgpu_ctx->get_rows_pipelines[GGML_TYPE_F16][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_f16, "get_rows_f16", constants); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_get_rows_f16, "get_rows_f16", constants); webgpu_ctx->get_rows_pipelines[GGML_TYPE_I32][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_i32, "get_rows_i32", constants); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_get_rows_i32, "get_rows_i32", constants); webgpu_ctx->get_rows_pipelines[GGML_TYPE_Q4_0][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_q4_0, "get_rows_q4_0", constants); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_get_rows_q4_0, "get_rows_q4_0", constants); webgpu_ctx->get_rows_pipelines[GGML_TYPE_Q4_1][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_q4_1, "get_rows_q4_1", constants); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_get_rows_q4_1, "get_rows_q4_1", constants); webgpu_ctx->get_rows_pipelines[GGML_TYPE_Q5_0][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_q5_0, "get_rows_q5_0", constants); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_get_rows_q5_0, "get_rows_q5_0", constants); webgpu_ctx->get_rows_pipelines[GGML_TYPE_Q5_1][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_q5_1, "get_rows_q5_1", constants); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_get_rows_q5_1, "get_rows_q5_1", constants); webgpu_ctx->get_rows_pipelines[GGML_TYPE_Q8_0][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_q8_0, "get_rows_q8_0", constants); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_get_rows_q8_0, "get_rows_q8_0", constants); webgpu_ctx->get_rows_pipelines[GGML_TYPE_Q2_K][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_q2_k, "get_rows_q2_k", constants); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_get_rows_q2_k, "get_rows_q2_k", constants); webgpu_ctx->get_rows_pipelines[GGML_TYPE_Q3_K][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_q3_k, "get_rows_q3_k", constants); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_get_rows_q3_k, "get_rows_q3_k", constants); webgpu_ctx->get_rows_pipelines[GGML_TYPE_Q4_K][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_q4_k, "get_rows_q4_k", constants); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_get_rows_q4_k, "get_rows_q4_k", constants); webgpu_ctx->get_rows_pipelines[GGML_TYPE_Q5_K][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_q5_k, "get_rows_q5_k", constants); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_get_rows_q5_k, "get_rows_q5_k", constants); webgpu_ctx->get_rows_pipelines[GGML_TYPE_Q6_K][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_q6_k, "get_rows_q6_k", constants); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_get_rows_q6_k, "get_rows_q6_k", constants); - webgpu_ctx->get_rows_pipelines[GGML_TYPE_IQ2_XXS][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_iq2_xxs, "get_rows_iq2_xxs", constants); + webgpu_ctx->get_rows_pipelines[GGML_TYPE_IQ2_XXS][0] = ggml_webgpu_create_pipeline( + webgpu_ctx->global_ctx->device, wgsl_get_rows_iq2_xxs, "get_rows_iq2_xxs", constants); webgpu_ctx->get_rows_pipelines[GGML_TYPE_IQ2_XS][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_iq2_xs, "get_rows_iq2_xs", constants); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_get_rows_iq2_xs, "get_rows_iq2_xs", constants); webgpu_ctx->get_rows_pipelines[GGML_TYPE_IQ2_S][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_iq2_s, "get_rows_iq2_s", constants); - webgpu_ctx->get_rows_pipelines[GGML_TYPE_IQ3_XXS][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_iq3_xxs, "get_rows_iq3_xxs", constants); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_get_rows_iq2_s, "get_rows_iq2_s", constants); + webgpu_ctx->get_rows_pipelines[GGML_TYPE_IQ3_XXS][0] = ggml_webgpu_create_pipeline( + webgpu_ctx->global_ctx->device, wgsl_get_rows_iq3_xxs, "get_rows_iq3_xxs", constants); webgpu_ctx->get_rows_pipelines[GGML_TYPE_IQ3_S][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_iq3_s, "get_rows_iq3_s", constants); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_get_rows_iq3_s, "get_rows_iq3_s", constants); webgpu_ctx->get_rows_pipelines[GGML_TYPE_IQ1_S][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_iq1_s, "get_rows_iq1_s", constants); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_get_rows_iq1_s, "get_rows_iq1_s", constants); webgpu_ctx->get_rows_pipelines[GGML_TYPE_IQ1_M][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_iq1_m, "get_rows_iq1_m", constants); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_get_rows_iq1_m, "get_rows_iq1_m", constants); webgpu_ctx->get_rows_pipelines[GGML_TYPE_IQ4_NL][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_iq4_nl, "get_rows_iq4_nl", constants); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_get_rows_iq4_nl, "get_rows_iq4_nl", constants); webgpu_ctx->get_rows_pipelines[GGML_TYPE_IQ4_XS][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_iq4_xs, "get_rows_iq4_xs", constants); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_get_rows_iq4_xs, "get_rows_iq4_xs", constants); } static void ggml_webgpu_init_cpy_pipeline(webgpu_context & webgpu_ctx) { std::vector constants = ggml_webgpu_wg_size_entry(WEBGPU_MAX_WG_SIZE); webgpu_ctx->cpy_pipelines[GGML_TYPE_F32][GGML_TYPE_F32] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_cpy_f32_f32, "cpy_f32_f32", constants); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_cpy_f32_f32, "cpy_f32_f32", constants); webgpu_ctx->cpy_pipelines[GGML_TYPE_F32][GGML_TYPE_I32] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_cpy_f32_i32, "cpy_f32_i32", constants); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_cpy_f32_i32, "cpy_f32_i32", constants); webgpu_ctx->cpy_pipelines[GGML_TYPE_F32][GGML_TYPE_F16] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_cpy_f32_f16, "cpy_f32_f16", constants); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_cpy_f32_f16, "cpy_f32_f16", constants); webgpu_ctx->cpy_pipelines[GGML_TYPE_F16][GGML_TYPE_F32] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_cpy_f16_f32, "cpy_f16_f32", constants); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_cpy_f16_f32, "cpy_f16_f32", constants); webgpu_ctx->cpy_pipelines[GGML_TYPE_F16][GGML_TYPE_F16] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_cpy_f16_f16, "cpy_f16_f16", constants); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_cpy_f16_f16, "cpy_f16_f16", constants); } static void ggml_webgpu_init_add_pipeline(webgpu_context & webgpu_ctx) { std::vector constants = ggml_webgpu_wg_size_entry(WEBGPU_MAX_WG_SIZE); webgpu_ctx->add_pipelines[GGML_TYPE_F32][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_add_f32, "add_f32", constants); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_add_f32, "add_f32", constants); webgpu_ctx->add_pipelines[GGML_TYPE_F16][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_add_f16, "add_f16", constants); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_add_f16, "add_f16", constants); webgpu_ctx->add_pipelines[GGML_TYPE_F32][1] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_add_f32_inplace, "add_f32_inplace", constants); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_add_f32_inplace, "add_f32_inplace", constants); webgpu_ctx->add_pipelines[GGML_TYPE_F16][1] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_add_f16_inplace, "add_f16_inplace", constants); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_add_f16_inplace, "add_f16_inplace", constants); } static void ggml_webgpu_init_sub_pipeline(webgpu_context & webgpu_ctx) { std::vector constants = ggml_webgpu_wg_size_entry(WEBGPU_MAX_WG_SIZE); webgpu_ctx->sub_pipelines[GGML_TYPE_F32][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_sub_f32, "sub_f32", constants); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_sub_f32, "sub_f32", constants); webgpu_ctx->sub_pipelines[GGML_TYPE_F16][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_sub_f16, "sub_f16", constants); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_sub_f16, "sub_f16", constants); webgpu_ctx->sub_pipelines[GGML_TYPE_F32][1] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_sub_f32_inplace, "sub_f32_inplace", constants); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_sub_f32_inplace, "sub_f32_inplace", constants); webgpu_ctx->sub_pipelines[GGML_TYPE_F16][1] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_sub_f16_inplace, "sub_f16_inplace", constants); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_sub_f16_inplace, "sub_f16_inplace", constants); } static void ggml_webgpu_init_mul_pipeline(webgpu_context & webgpu_ctx) { std::vector constants = ggml_webgpu_wg_size_entry(WEBGPU_MAX_WG_SIZE); webgpu_ctx->mul_pipelines[GGML_TYPE_F32][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_f32, "mul_f32", constants); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_mul_f32, "mul_f32", constants); webgpu_ctx->mul_pipelines[GGML_TYPE_F16][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_f16, "mul_f16", constants); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_mul_f16, "mul_f16", constants); webgpu_ctx->mul_pipelines[GGML_TYPE_F32][1] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_f32_inplace, "mul_f32_inplace", constants); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_mul_f32_inplace, "mul_f32_inplace", constants); webgpu_ctx->mul_pipelines[GGML_TYPE_F16][1] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_f16_inplace, "mul_f16_inplace", constants); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_mul_f16_inplace, "mul_f16_inplace", constants); } static void ggml_webgpu_init_div_pipeline(webgpu_context & webgpu_ctx) { std::vector constants = ggml_webgpu_wg_size_entry(WEBGPU_MAX_WG_SIZE); webgpu_ctx->div_pipelines[GGML_TYPE_F32][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_div_f32, "div_f32", constants); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_div_f32, "div_f32", constants); webgpu_ctx->div_pipelines[GGML_TYPE_F16][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_div_f16, "div_f16", constants); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_div_f16, "div_f16", constants); webgpu_ctx->div_pipelines[GGML_TYPE_F32][1] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_div_f32_inplace, "div_f32_inplace", constants); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_div_f32_inplace, "div_f32_inplace", constants); webgpu_ctx->div_pipelines[GGML_TYPE_F16][1] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_div_f16_inplace, "div_f16_inplace", constants); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_div_f16_inplace, "div_f16_inplace", constants); } static void ggml_webgpu_init_rms_norm_pipeline(webgpu_context & webgpu_ctx) { std::vector constants = ggml_webgpu_wg_size_entry(WEBGPU_ROW_SPLIT_WG_SIZE); webgpu_ctx->rms_norm_pipelines[0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_rms_norm, "rms_norm", constants); - webgpu_ctx->rms_norm_pipelines[1] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_rms_norm_inplace, "rms_norm_inplace", constants); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_rms_norm, "rms_norm", constants); + webgpu_ctx->rms_norm_pipelines[1] = ggml_webgpu_create_pipeline( + webgpu_ctx->global_ctx->device, wgsl_rms_norm_inplace, "rms_norm_inplace", constants); } static void ggml_webgpu_init_rope_pipeline(webgpu_context & webgpu_ctx) { std::vector constants = ggml_webgpu_wg_size_entry(WEBGPU_MAX_WG_SIZE); webgpu_ctx->rope_pipelines[GGML_TYPE_F32][0][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_rope_f32, "rope_f32", constants); - webgpu_ctx->rope_pipelines[GGML_TYPE_F32][0][1] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_rope_f32_inplace, "rope_f32_inplace", constants); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_rope_f32, "rope_f32", constants); + webgpu_ctx->rope_pipelines[GGML_TYPE_F32][0][1] = ggml_webgpu_create_pipeline( + webgpu_ctx->global_ctx->device, wgsl_rope_f32_inplace, "rope_f32_inplace", constants); webgpu_ctx->rope_pipelines[GGML_TYPE_F32][1][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_rope_f32_ff, "rope_f32_ff", constants); - webgpu_ctx->rope_pipelines[GGML_TYPE_F32][1][1] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_rope_f32_ff_inplace, "rope_f32_ff_inplace", constants); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_rope_f32_ff, "rope_f32_ff", constants); + webgpu_ctx->rope_pipelines[GGML_TYPE_F32][1][1] = ggml_webgpu_create_pipeline( + webgpu_ctx->global_ctx->device, wgsl_rope_f32_ff_inplace, "rope_f32_ff_inplace", constants); webgpu_ctx->rope_pipelines[GGML_TYPE_F16][0][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_rope_f16, "rope_f16", constants); - webgpu_ctx->rope_pipelines[GGML_TYPE_F16][0][1] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_rope_f16_inplace, "rope_f16_inplace", constants); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_rope_f16, "rope_f16", constants); + webgpu_ctx->rope_pipelines[GGML_TYPE_F16][0][1] = ggml_webgpu_create_pipeline( + webgpu_ctx->global_ctx->device, wgsl_rope_f16_inplace, "rope_f16_inplace", constants); webgpu_ctx->rope_pipelines[GGML_TYPE_F16][1][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_rope_f16_ff, "rope_f16_ff", constants); - webgpu_ctx->rope_pipelines[GGML_TYPE_F16][1][1] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_rope_f16_ff_inplace, "rope_f16_ff_inplace", constants); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_rope_f16_ff, "rope_f16_ff", constants); + webgpu_ctx->rope_pipelines[GGML_TYPE_F16][1][1] = ggml_webgpu_create_pipeline( + webgpu_ctx->global_ctx->device, wgsl_rope_f16_ff_inplace, "rope_f16_ff_inplace", constants); } static void ggml_webgpu_init_glu_pipeline(webgpu_context & webgpu_ctx) { @@ -2719,68 +2780,68 @@ static void ggml_webgpu_init_glu_pipeline(webgpu_context & webgpu_ctx) { // REGLU webgpu_ctx->glu_pipelines[GGML_GLU_OP_REGLU][GGML_TYPE_F32][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_reglu_f32, "reglu_f32", constants); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_reglu_f32, "reglu_f32", constants); webgpu_ctx->glu_pipelines[GGML_GLU_OP_REGLU][GGML_TYPE_F16][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_reglu_f16, "reglu_f16", constants); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_reglu_f16, "reglu_f16", constants); webgpu_ctx->glu_pipelines[GGML_GLU_OP_REGLU][GGML_TYPE_F32][1] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_reglu_f32_split, "reglu_f32_split", constants); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_reglu_f32_split, "reglu_f32_split", constants); webgpu_ctx->glu_pipelines[GGML_GLU_OP_REGLU][GGML_TYPE_F16][1] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_reglu_f16_split, "reglu_f16_split", constants); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_reglu_f16_split, "reglu_f16_split", constants); // GEGLU webgpu_ctx->glu_pipelines[GGML_GLU_OP_GEGLU][GGML_TYPE_F32][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_geglu_f32, "geglu_f32", constants); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_geglu_f32, "geglu_f32", constants); webgpu_ctx->glu_pipelines[GGML_GLU_OP_GEGLU][GGML_TYPE_F16][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_geglu_f16, "geglu_f16", constants); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_geglu_f16, "geglu_f16", constants); webgpu_ctx->glu_pipelines[GGML_GLU_OP_GEGLU][GGML_TYPE_F32][1] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_geglu_f32_split, "geglu_f32_split", constants); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_geglu_f32_split, "geglu_f32_split", constants); webgpu_ctx->glu_pipelines[GGML_GLU_OP_GEGLU][GGML_TYPE_F16][1] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_geglu_f16_split, "geglu_f16_split", constants); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_geglu_f16_split, "geglu_f16_split", constants); // SWIGLU webgpu_ctx->glu_pipelines[GGML_GLU_OP_SWIGLU][GGML_TYPE_F32][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_swiglu_f32, "swiglu_f32", constants); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_swiglu_f32, "swiglu_f32", constants); webgpu_ctx->glu_pipelines[GGML_GLU_OP_SWIGLU][GGML_TYPE_F16][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_swiglu_f16, "swiglu_f16", constants); - webgpu_ctx->glu_pipelines[GGML_GLU_OP_SWIGLU][GGML_TYPE_F32][1] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_swiglu_f32_split, "swiglu_f32_split", constants); - webgpu_ctx->glu_pipelines[GGML_GLU_OP_SWIGLU][GGML_TYPE_F16][1] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_swiglu_f16_split, "swiglu_f16_split", constants); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_swiglu_f16, "swiglu_f16", constants); + webgpu_ctx->glu_pipelines[GGML_GLU_OP_SWIGLU][GGML_TYPE_F32][1] = ggml_webgpu_create_pipeline( + webgpu_ctx->global_ctx->device, wgsl_swiglu_f32_split, "swiglu_f32_split", constants); + webgpu_ctx->glu_pipelines[GGML_GLU_OP_SWIGLU][GGML_TYPE_F16][1] = ggml_webgpu_create_pipeline( + webgpu_ctx->global_ctx->device, wgsl_swiglu_f16_split, "swiglu_f16_split", constants); // SWIGLU_OAI webgpu_ctx->glu_pipelines[GGML_GLU_OP_SWIGLU_OAI][GGML_TYPE_F32][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_swiglu_oai_f32, "swiglu_oai_f32", constants); - webgpu_ctx->glu_pipelines[GGML_GLU_OP_SWIGLU_OAI][GGML_TYPE_F32][1] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_swiglu_oai_f32_split, "swiglu_oai_f32_split", constants); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_swiglu_oai_f32, "swiglu_oai_f32", constants); + webgpu_ctx->glu_pipelines[GGML_GLU_OP_SWIGLU_OAI][GGML_TYPE_F32][1] = ggml_webgpu_create_pipeline( + webgpu_ctx->global_ctx->device, wgsl_swiglu_oai_f32_split, "swiglu_oai_f32_split", constants); // GEGLU_ERF webgpu_ctx->glu_pipelines[GGML_GLU_OP_GEGLU_ERF][GGML_TYPE_F32][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_geglu_erf_f32, "geglu_erf_f32", constants); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_geglu_erf_f32, "geglu_erf_f32", constants); webgpu_ctx->glu_pipelines[GGML_GLU_OP_GEGLU_ERF][GGML_TYPE_F16][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_geglu_erf_f16, "geglu_erf_f16", constants); - webgpu_ctx->glu_pipelines[GGML_GLU_OP_GEGLU_ERF][GGML_TYPE_F32][1] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_geglu_erf_f32_split, "geglu_erf_f32_split", constants); - webgpu_ctx->glu_pipelines[GGML_GLU_OP_GEGLU_ERF][GGML_TYPE_F16][1] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_geglu_erf_f16_split, "geglu_erf_f16_split", constants); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_geglu_erf_f16, "geglu_erf_f16", constants); + webgpu_ctx->glu_pipelines[GGML_GLU_OP_GEGLU_ERF][GGML_TYPE_F32][1] = ggml_webgpu_create_pipeline( + webgpu_ctx->global_ctx->device, wgsl_geglu_erf_f32_split, "geglu_erf_f32_split", constants); + webgpu_ctx->glu_pipelines[GGML_GLU_OP_GEGLU_ERF][GGML_TYPE_F16][1] = ggml_webgpu_create_pipeline( + webgpu_ctx->global_ctx->device, wgsl_geglu_erf_f16_split, "geglu_erf_f16_split", constants); // GEGLU_QUICK webgpu_ctx->glu_pipelines[GGML_GLU_OP_GEGLU_QUICK][GGML_TYPE_F32][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_geglu_quick_f32, "geglu_quick_f32", constants); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_geglu_quick_f32, "geglu_quick_f32", constants); webgpu_ctx->glu_pipelines[GGML_GLU_OP_GEGLU_QUICK][GGML_TYPE_F16][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_geglu_quick_f16, "geglu_quick_f16", constants); - webgpu_ctx->glu_pipelines[GGML_GLU_OP_GEGLU_QUICK][GGML_TYPE_F32][1] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_geglu_quick_f32_split, "geglu_quick_f32_split", constants); - webgpu_ctx->glu_pipelines[GGML_GLU_OP_GEGLU_QUICK][GGML_TYPE_F16][1] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_geglu_quick_f16_split, "geglu_quick_f16_split", constants); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_geglu_quick_f16, "geglu_quick_f16", constants); + webgpu_ctx->glu_pipelines[GGML_GLU_OP_GEGLU_QUICK][GGML_TYPE_F32][1] = ggml_webgpu_create_pipeline( + webgpu_ctx->global_ctx->device, wgsl_geglu_quick_f32_split, "geglu_quick_f32_split", constants); + webgpu_ctx->glu_pipelines[GGML_GLU_OP_GEGLU_QUICK][GGML_TYPE_F16][1] = ggml_webgpu_create_pipeline( + webgpu_ctx->global_ctx->device, wgsl_geglu_quick_f16_split, "geglu_quick_f16_split", constants); } static void ggml_webgpu_init_scale_pipeline(webgpu_context & webgpu_ctx) { std::vector constants = ggml_webgpu_wg_size_entry(WEBGPU_MAX_WG_SIZE); webgpu_ctx->scale_pipelines[0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_scale_f32, "scale_f32", constants); - webgpu_ctx->scale_pipelines[1] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_scale_f32_inplace, "scale_f32_inplace", constants); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_scale_f32, "scale_f32", constants); + webgpu_ctx->scale_pipelines[1] = ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_scale_f32_inplace, + "scale_f32_inplace", constants); } static void ggml_webgpu_init_soft_max_pipeline(webgpu_context & webgpu_ctx) { @@ -2788,56 +2849,243 @@ static void ggml_webgpu_init_soft_max_pipeline(webgpu_context & webgpu_ctx) { // f32 (no mask) webgpu_ctx->soft_max_pipelines[2][0][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_soft_max_f32, "soft_max_f32", constants); - webgpu_ctx->soft_max_pipelines[2][0][1] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_soft_max_f32_inplace, "soft_max_f32_inplace", constants); - webgpu_ctx->soft_max_pipelines[2][1][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_soft_max_f32_sink, "soft_max_f32_sink", constants); + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_soft_max_f32, "soft_max_f32", constants); + webgpu_ctx->soft_max_pipelines[2][0][1] = ggml_webgpu_create_pipeline( + webgpu_ctx->global_ctx->device, wgsl_soft_max_f32_inplace, "soft_max_f32_inplace", constants); + webgpu_ctx->soft_max_pipelines[2][1][0] = ggml_webgpu_create_pipeline( + webgpu_ctx->global_ctx->device, wgsl_soft_max_f32_sink, "soft_max_f32_sink", constants); webgpu_ctx->soft_max_pipelines[2][1][1] = ggml_webgpu_create_pipeline( - webgpu_ctx->device, wgsl_soft_max_f32_sink_inplace, "soft_max_f32_sink_inplace", constants); + webgpu_ctx->global_ctx->device, wgsl_soft_max_f32_sink_inplace, "soft_max_f32_sink_inplace", constants); // f32 mask (mask_type = 0) - webgpu_ctx->soft_max_pipelines[0][0][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_soft_max_f32_mask_f32, "soft_max_f32_mask_f32", constants); + webgpu_ctx->soft_max_pipelines[0][0][0] = ggml_webgpu_create_pipeline( + webgpu_ctx->global_ctx->device, wgsl_soft_max_f32_mask_f32, "soft_max_f32_mask_f32", constants); webgpu_ctx->soft_max_pipelines[0][0][1] = ggml_webgpu_create_pipeline( - webgpu_ctx->device, wgsl_soft_max_f32_mask_f32_inplace, "soft_max_f32_mask_f32_inplace", constants); + webgpu_ctx->global_ctx->device, wgsl_soft_max_f32_mask_f32_inplace, "soft_max_f32_mask_f32_inplace", constants); webgpu_ctx->soft_max_pipelines[0][1][0] = ggml_webgpu_create_pipeline( - webgpu_ctx->device, wgsl_soft_max_f32_mask_f32_sink, "soft_max_f32_mask_f32_sink", constants); - webgpu_ctx->soft_max_pipelines[0][1][1] = ggml_webgpu_create_pipeline( - webgpu_ctx->device, wgsl_soft_max_f32_mask_f32_sink_inplace, "soft_max_f32_mask_f32_sink_inplace", constants); + webgpu_ctx->global_ctx->device, wgsl_soft_max_f32_mask_f32_sink, "soft_max_f32_mask_f32_sink", constants); + webgpu_ctx->soft_max_pipelines[0][1][1] = + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_soft_max_f32_mask_f32_sink_inplace, + "soft_max_f32_mask_f32_sink_inplace", constants); // f16 mask (mask_type = 1) - webgpu_ctx->soft_max_pipelines[1][0][0] = - ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_soft_max_f32_mask_f16, "soft_max_f32_mask_f16", constants); + webgpu_ctx->soft_max_pipelines[1][0][0] = ggml_webgpu_create_pipeline( + webgpu_ctx->global_ctx->device, wgsl_soft_max_f32_mask_f16, "soft_max_f32_mask_f16", constants); webgpu_ctx->soft_max_pipelines[1][0][1] = ggml_webgpu_create_pipeline( - webgpu_ctx->device, wgsl_soft_max_f32_mask_f16_inplace, "soft_max_f32_mask_f16_inplace", constants); + webgpu_ctx->global_ctx->device, wgsl_soft_max_f32_mask_f16_inplace, "soft_max_f32_mask_f16_inplace", constants); webgpu_ctx->soft_max_pipelines[1][1][0] = ggml_webgpu_create_pipeline( - webgpu_ctx->device, wgsl_soft_max_f32_mask_f16_sink, "soft_max_f32_mask_f16_sink", constants); - webgpu_ctx->soft_max_pipelines[1][1][1] = ggml_webgpu_create_pipeline( - webgpu_ctx->device, wgsl_soft_max_f32_mask_f16_sink_inplace, "soft_max_f32_mask_f16_sink_inplace", constants); + webgpu_ctx->global_ctx->device, wgsl_soft_max_f32_mask_f16_sink, "soft_max_f32_mask_f16_sink", constants); + webgpu_ctx->soft_max_pipelines[1][1][1] = + ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_soft_max_f32_mask_f16_sink_inplace, + "soft_max_f32_mask_f16_sink_inplace", constants); } -// TODO: move most initialization logic here -static ggml_backend_t ggml_backend_webgpu_device_init(ggml_backend_dev_t dev, const char * params) { +static bool create_webgpu_device(ggml_backend_webgpu_reg_context * ctx) { + wgpu::RequestAdapterOptions options = {}; + +#ifndef __EMSCRIPTEN__ + // TODO: track need for these toggles: https://issues.chromium.org/issues/42251215 + const char * const adapterEnabledToggles[] = { "vulkan_enable_f16_on_nvidia", "use_vulkan_memory_model" }; + wgpu::DawnTogglesDescriptor adapterTogglesDesc; + adapterTogglesDesc.enabledToggles = adapterEnabledToggles; + adapterTogglesDesc.enabledToggleCount = 2; + options.nextInChain = &adapterTogglesDesc; +#endif + + ctx->webgpu_global_ctx->instance.WaitAny( + ctx->webgpu_global_ctx->instance.RequestAdapter( + &options, wgpu::CallbackMode::AllowSpontaneous, + [&ctx](wgpu::RequestAdapterStatus status, wgpu::Adapter adapter, const char * message) { + if (status != wgpu::RequestAdapterStatus::Success) { + GGML_LOG_ERROR("ggml_webgpu: Failed to get an adapter: %s\n", message); + return; + } + ctx->webgpu_global_ctx->adapter = std::move(adapter); + }), + UINT64_MAX); + GGML_ASSERT(ctx->webgpu_global_ctx->adapter != nullptr); + + ctx->webgpu_global_ctx->adapter.GetLimits(&ctx->webgpu_global_ctx->capabilities.limits); + + wgpu::AdapterInfo info{}; +#ifndef __EMSCRIPTEN__ + wgpu::AdapterPropertiesSubgroupMatrixConfigs subgroup_matrix_configs{}; + if (ctx->webgpu_global_ctx->adapter.HasFeature(wgpu::FeatureName::ChromiumExperimentalSubgroupMatrix)) { + info.nextInChain = &subgroup_matrix_configs; + } +#endif + ctx->webgpu_global_ctx->adapter.GetInfo(&info); + wgpu::SupportedFeatures features; + ctx->webgpu_global_ctx->adapter.GetFeatures(&features); + // we require f16 support + GGML_ASSERT(ctx->webgpu_global_ctx->adapter.HasFeature(wgpu::FeatureName::ShaderF16)); + +#ifndef __EMSCRIPTEN__ + // Only support square f16 matrices of size 8 or 16 for now + bool valid_subgroup_matrix_config = false; + if (ctx->webgpu_global_ctx->adapter.HasFeature(wgpu::FeatureName::ChromiumExperimentalSubgroupMatrix)) { + for (size_t i = 0; i < subgroup_matrix_configs.configCount; i++) { + const wgpu::SubgroupMatrixConfig config = subgroup_matrix_configs.configs[i]; + if (config.M == config.N && config.N == config.K && (config.K == 8 || config.K == 16) && + config.componentType == wgpu::SubgroupMatrixComponentType::F16 && + config.resultComponentType == wgpu::SubgroupMatrixComponentType::F16) { + ctx->webgpu_global_ctx->capabilities.sg_mat_m = config.M; + ctx->webgpu_global_ctx->capabilities.sg_mat_n = config.N; + ctx->webgpu_global_ctx->capabilities.sg_mat_k = config.K; + valid_subgroup_matrix_config = true; + break; + } + } + } + ctx->webgpu_global_ctx->capabilities.supports_subgroup_matrix = valid_subgroup_matrix_config; +#endif + + // For subgroup matrix code to be the most efficient, we would like the subgroup size to be consistent and accurate. + // Unfortunately, that is not possible, so we use the maximum subgroup size reported by the adapter. + ctx->webgpu_global_ctx->capabilities.max_subgroup_size = info.subgroupMaxSize; + // Initialize device + std::vector required_features = { wgpu::FeatureName::ShaderF16 }; + +#ifndef __EMSCRIPTEN__ + required_features.push_back(wgpu::FeatureName::ImplicitDeviceSynchronization); + if (ctx->webgpu_global_ctx->capabilities.supports_subgroup_matrix) { + required_features.push_back(wgpu::FeatureName::Subgroups); + required_features.push_back(wgpu::FeatureName::ChromiumExperimentalSubgroupMatrix); + } +#endif + +#ifdef GGML_WEBGPU_GPU_PROFILE + required_features.push_back(wgpu::FeatureName::TimestampQuery); +#endif + + wgpu::DeviceDescriptor dev_desc; + dev_desc.requiredLimits = &ctx->webgpu_global_ctx->capabilities.limits; + dev_desc.requiredFeatures = required_features.data(); + dev_desc.requiredFeatureCount = required_features.size(); + dev_desc.SetDeviceLostCallback( + wgpu::CallbackMode::AllowSpontaneous, + [](const wgpu::Device & device, wgpu::DeviceLostReason reason, wgpu::StringView message) { + GGML_UNUSED(device); + GGML_UNUSED(reason); + GGML_UNUSED(message); + //TODO: uncomment once proper free logic is in place + //GGML_LOG_ERROR("ggml_webgpu: Device lost! Reason: %d, Message: %s\n", static_cast(reason), + //std::string(message).c_str()); + }); + dev_desc.SetUncapturedErrorCallback( + [](const wgpu::Device & device, wgpu::ErrorType reason, wgpu::StringView message) { + GGML_UNUSED(device); + GGML_ABORT("ggml_webgpu: Device error! Reason: %d, Message: %s\n", static_cast(reason), + std::string(message).c_str()); + }); + +#ifndef __EMSCRIPTEN__ + // Enable Dawn-specific toggles to increase native performance + // TODO: Maybe WebGPU needs a "fast" mode where you can request compilers skip adding checks like these, + // only for native performance? + const char * const deviceEnabledToggles[] = { "skip_validation", "disable_robustness", "disable_workgroup_init", + "disable_polyfills_on_integer_div_and_mod" }; + const char * const deviceDisabledToggles[] = { "timestamp_quantization" }; + wgpu::DawnTogglesDescriptor deviceTogglesDesc; + deviceTogglesDesc.enabledToggles = deviceEnabledToggles; + deviceTogglesDesc.enabledToggleCount = 4; + deviceTogglesDesc.disabledToggles = deviceDisabledToggles; + deviceTogglesDesc.disabledToggleCount = 1; + + dev_desc.nextInChain = &deviceTogglesDesc; +#endif + + ctx->webgpu_global_ctx->instance.WaitAny( + ctx->webgpu_global_ctx->adapter.RequestDevice( + &dev_desc, wgpu::CallbackMode::AllowSpontaneous, + [ctx](wgpu::RequestDeviceStatus status, wgpu::Device device, wgpu::StringView message) { + if (status != wgpu::RequestDeviceStatus::Success) { + GGML_LOG_ERROR("ggml_webgpu: Failed to get a device: %s\n", std::string(message).c_str()); + return; + } + ctx->webgpu_global_ctx->device = std::move(device); + }), + UINT64_MAX); + GGML_ASSERT(ctx->webgpu_global_ctx->device != nullptr); + + ggml_webgpu_init_memset_pipeline(ctx->webgpu_global_ctx); + ctx->webgpu_global_ctx->memset_buf_pool.init(ctx->webgpu_global_ctx->device, 1, WEBGPU_PARAMS_BUF_SIZE_BYTES, + wgpu::BufferUsage::CopyDst | wgpu::BufferUsage::Uniform, + wgpu::BufferUsage::CopySrc | wgpu::BufferUsage::MapWrite); + ctx->webgpu_global_ctx->queue = ctx->webgpu_global_ctx->device.GetQueue(); + +#ifdef GGML_WEBGPU_GPU_PROFILE + // Initialize buffer pool for timestamp queries, used for profiling + ctx->webgpu_global_ctx->timestamp_query_buf_pool.init(ctx->webgpu_global_ctx->device, WEBGPU_NUM_TIMESTAMP_QUERY_BUFS, + WEBGPU_TIMESTAMP_QUERY_BUF_SIZE_BYTES, + wgpu::BufferUsage::QueryResolve | wgpu::BufferUsage::CopySrc, + wgpu::BufferUsage::MapRead | wgpu::BufferUsage::CopyDst); +#endif + + GGML_LOG_INFO( + "ggml_webgpu: adapter_info: vendor_id: %u | vendor: %s | architecture: %s | device_id: %u | name: %s | " + "device_desc: %s\n", + info.vendorID, std::string(info.vendor).c_str(), std::string(info.architecture).c_str(), info.deviceID, + std::string(info.device).c_str(), std::string(info.description).c_str()); + return true; +} + +static webgpu_context initialize_webgpu_context(ggml_backend_dev_t dev) { + ggml_backend_webgpu_device_context * dev_ctx = (ggml_backend_webgpu_device_context *) dev->context; + webgpu_context webgpu_ctx = std::make_shared(); + webgpu_ctx->global_ctx = dev_ctx->webgpu_global_ctx; + webgpu_ctx->param_buf_pool.init(webgpu_ctx->global_ctx->device, WEBGPU_NUM_PARAM_BUFS, WEBGPU_PARAMS_BUF_SIZE_BYTES, + wgpu::BufferUsage::CopyDst | wgpu::BufferUsage::Uniform, + wgpu::BufferUsage::CopySrc | wgpu::BufferUsage::MapWrite); + webgpu_ctx->set_rows_error_buf_pool.init(webgpu_ctx->global_ctx->device, WEBGPU_NUM_SET_ROWS_ERROR_BUFS, + WEBGPU_SET_ROWS_ERROR_BUF_SIZE_BYTES, + wgpu::BufferUsage::CopySrc | wgpu::BufferUsage::Storage, + wgpu::BufferUsage::CopyDst | wgpu::BufferUsage::MapRead); + + ggml_webgpu_init_mul_mat_pipeline(webgpu_ctx); + ggml_webgpu_init_get_rows_pipeline(webgpu_ctx); + ggml_webgpu_init_cpy_pipeline(webgpu_ctx); + ggml_webgpu_init_add_pipeline(webgpu_ctx); + ggml_webgpu_init_sub_pipeline(webgpu_ctx); + ggml_webgpu_init_mul_pipeline(webgpu_ctx); + ggml_webgpu_init_div_pipeline(webgpu_ctx); + ggml_webgpu_init_rms_norm_pipeline(webgpu_ctx); + ggml_webgpu_init_rope_pipeline(webgpu_ctx); + ggml_webgpu_init_glu_pipeline(webgpu_ctx); + ggml_webgpu_init_scale_pipeline(webgpu_ctx); + ggml_webgpu_init_soft_max_pipeline(webgpu_ctx); +#ifdef GGML_WEBGPU_DEBUG + // Initialize debug buffers + ggml_webgpu_create_buffer(webgpu_ctx->global_ctx->device, webgpu_ctx->global_ctx->debug_host_buf, + WEBGPU_DEBUG_BUF_ELEMS * sizeof(uint32_t), + wgpu::BufferUsage::CopyDst | wgpu::BufferUsage::MapRead, "debug_host_buf"); + ggml_webgpu_create_buffer(webgpu_ctx->global_ctx->device, webgpu_ctx->global_ctx->debug_dev_buf, + WEBGPU_DEBUG_BUF_ELEMS * sizeof(uint32_t), + wgpu::BufferUsage::Storage | wgpu::BufferUsage::CopySrc, "debug_dev_buf"); +#endif + return webgpu_ctx; +} + +static ggml_backend_t ggml_backend_webgpu_backend_init(ggml_backend_dev_t dev, const char * params) { GGML_UNUSED(params); - WEBGPU_LOG_DEBUG("ggml_backend_webgpu_device_init()"); + WEBGPU_LOG_DEBUG("ggml_backend_webgpu_backend_init()"); - ggml_backend_webgpu_device_context * dev_ctx = static_cast(dev->context); - webgpu_context webgpu_ctx = dev_ctx->webgpu_ctx; + ggml_backend_webgpu_device_context * dev_ctx = static_cast(dev->context); - static ggml_backend_webgpu_context backend_ctx; - backend_ctx.name = GGML_WEBGPU_NAME + std::string(": ") + dev_ctx->device_name; - backend_ctx.webgpu_ctx = webgpu_ctx; + auto * backend_ctx = new ggml_backend_webgpu_context(); + backend_ctx->name = GGML_WEBGPU_NAME + std::string(": ") + dev_ctx->device_name; + backend_ctx->webgpu_ctx = initialize_webgpu_context(dev); // See GGML Backend Interface section - static ggml_backend backend = { + auto * backend = new ggml_backend(); + *backend = { /* .guid = */ ggml_backend_webgpu_guid(), /* .interface = */ ggml_backend_webgpu_i, /* .device = */ dev, - /* .context = */ &backend_ctx, + /* .context = */ backend_ctx, }; - return &backend; + return backend; } static ggml_backend_buffer_type_t ggml_backend_webgpu_device_get_buffer_type(ggml_backend_dev_t dev) { @@ -2854,7 +3102,8 @@ static ggml_backend_buffer_type_t ggml_backend_webgpu_device_get_buffer_type(ggm }, /* .device = */ dev, - /* .context = */ NULL, + /* .context = */ + NULL }; return &ggml_backend_webgpu_buffer_type; @@ -2895,16 +3144,16 @@ static bool ggml_webgpu_supported_qtype(ggml_type type) { static bool ggml_backend_webgpu_device_supports_op(ggml_backend_dev_t dev, const ggml_tensor * op) { ggml_backend_webgpu_device_context * ctx = static_cast(dev->context); - webgpu_context webgpu_ctx = ctx->webgpu_ctx; - ggml_tensor * src0 = op->src[0]; ggml_tensor * src1 = op->src[1]; ggml_tensor * src2 = op->src[2]; // on smaller devices (or CI), tensors may be larger than the max storage buffer size - if (ggml_nbytes(op) > webgpu_ctx->limits.maxStorageBufferBindingSize || - (src0 != nullptr && ggml_nbytes(src0) > webgpu_ctx->limits.maxStorageBufferBindingSize) || - (src1 != nullptr && ggml_nbytes(src1) > webgpu_ctx->limits.maxStorageBufferBindingSize)) { + if (ggml_nbytes(op) > ctx->webgpu_global_ctx->capabilities.limits.maxStorageBufferBindingSize || + (src0 != nullptr && + ggml_nbytes(src0) > ctx->webgpu_global_ctx->capabilities.limits.maxStorageBufferBindingSize) || + (src1 != nullptr && + ggml_nbytes(src1) > ctx->webgpu_global_ctx->capabilities.limits.maxStorageBufferBindingSize)) { return false; } @@ -2984,17 +3233,19 @@ static bool ggml_backend_webgpu_device_supports_op(ggml_backend_dev_t dev, const } case GGML_OP_FLASH_ATTN_EXT: { - if (!webgpu_ctx->supports_subgroup_matrix) { +#ifndef __EMSCRIPTEN__ + if (!ctx->webgpu_global_ctx->capabilities.supports_subgroup_matrix) { break; } // Head dimensions must fit in workgroup memory with minimum tile sizes - size_t limit_bytes = webgpu_ctx->limits.maxComputeWorkgroupStorageSize; + size_t limit_bytes = ctx->webgpu_global_ctx->capabilities.limits.maxComputeWorkgroupStorageSize; const bool has_mask = op->src[3] != nullptr; - const bool kv_direct = src1->type == GGML_TYPE_F16 && (src0->ne[0] % webgpu_ctx->sg_mat_k) == 0 && + const bool kv_direct = src1->type == GGML_TYPE_F16 && + (src0->ne[0] % ctx->webgpu_global_ctx->capabilities.sg_mat_k) == 0 && (src1->ne[1] % GGML_WEBGPU_KV_SEQ_PAD) == 0; const size_t min_bytes = ggml_webgpu_flash_attn_wg_mem_bytes( - webgpu_ctx->sg_mat_m, webgpu_ctx->sg_mat_n, (uint32_t) src0->ne[0], (uint32_t) src2->ne[0], - has_mask, kv_direct); + ctx->webgpu_global_ctx->capabilities.sg_mat_m, ctx->webgpu_global_ctx->capabilities.sg_mat_n, + (uint32_t) src0->ne[0], (uint32_t) src2->ne[0], has_mask, kv_direct); if (min_bytes > limit_bytes) { break; } @@ -3003,6 +3254,7 @@ static bool ggml_backend_webgpu_device_supports_op(ggml_backend_dev_t dev, const (src1->type == GGML_TYPE_F32 || src1->type == GGML_TYPE_F16 || src1->type == GGML_TYPE_Q4_0 || src1->type == GGML_TYPE_Q8_0) && src2->type == src1->type && op->type == GGML_TYPE_F32; +#endif break; } case GGML_OP_RMS_NORM: @@ -3099,10 +3351,13 @@ static bool ggml_backend_webgpu_device_supports_op(ggml_backend_dev_t dev, const default: break; } - if (ggml_nbytes(op) > webgpu_ctx->limits.maxStorageBufferBindingSize || - (src0 != nullptr && ggml_nbytes(src0) > webgpu_ctx->limits.maxStorageBufferBindingSize) || - (src1 != nullptr && ggml_nbytes(src1) > webgpu_ctx->limits.maxStorageBufferBindingSize) || - (src2 != nullptr && ggml_nbytes(src2) > webgpu_ctx->limits.maxStorageBufferBindingSize)) { + if (ggml_nbytes(op) > ctx->webgpu_global_ctx->capabilities.limits.maxStorageBufferBindingSize || + (src0 != nullptr && + ggml_nbytes(src0) > ctx->webgpu_global_ctx->capabilities.limits.maxStorageBufferBindingSize) || + (src1 != nullptr && + ggml_nbytes(src1) > ctx->webgpu_global_ctx->capabilities.limits.maxStorageBufferBindingSize) || + (src2 != nullptr && + ggml_nbytes(src2) > ctx->webgpu_global_ctx->capabilities.limits.maxStorageBufferBindingSize)) { supports_op = false; WEBGPU_LOG_DEBUG("ggml_webgpu op not supported due to size: "); } @@ -3127,7 +3382,7 @@ static struct ggml_backend_device_i ggml_backend_webgpu_device_i = { /* .get_memory = */ ggml_backend_webgpu_device_get_memory, /* .get_type = */ ggml_backend_webgpu_device_get_type, /* .get_props = */ ggml_backend_webgpu_device_get_props, - /* .init_backend = */ ggml_backend_webgpu_device_init, + /* .init_backend = */ ggml_backend_webgpu_backend_init, /* .get_buffer_type = */ ggml_backend_webgpu_device_get_buffer_type, /* .get_host_buffer_type = */ NULL, /* .buffer_from_host_ptr = */ NULL, @@ -3156,6 +3411,7 @@ static size_t ggml_backend_webgpu_reg_get_device_count(ggml_backend_reg_t reg) { // TODO: Does this need to be thread safe? Is it only called once? // TODO: move most logic to device_init function so backend can be freed/initialized properly // Only one device is supported for now + static ggml_backend_dev_t ggml_backend_webgpu_reg_get_device(ggml_backend_reg_t reg, size_t index) { GGML_ASSERT(index == 0); WEBGPU_LOG_DEBUG("ggml_backend_reg_get_device()"); @@ -3164,189 +3420,12 @@ static ggml_backend_dev_t ggml_backend_webgpu_reg_get_device(ggml_backend_reg_t ggml_backend_webgpu_reg_context * reg_ctx = static_cast(reg->context); - webgpu_context ctx = reg_ctx->webgpu_ctx; - - wgpu::RequestAdapterOptions options = {}; - -#ifndef __EMSCRIPTEN__ - // TODO: track need for these toggles: https://issues.chromium.org/issues/42251215 - const char * const adapterEnabledToggles[] = { "vulkan_enable_f16_on_nvidia", "use_vulkan_memory_model" }; - wgpu::DawnTogglesDescriptor adapterTogglesDesc; - adapterTogglesDesc.enabledToggles = adapterEnabledToggles; - adapterTogglesDesc.enabledToggleCount = 2; - options.nextInChain = &adapterTogglesDesc; -#endif - - ctx->instance.WaitAny(ctx->instance.RequestAdapter( - &options, wgpu::CallbackMode::AllowSpontaneous, - [&ctx](wgpu::RequestAdapterStatus status, wgpu::Adapter adapter, const char * message) { - if (status != wgpu::RequestAdapterStatus::Success) { - GGML_LOG_ERROR("ggml_webgpu: Failed to get an adapter: %s\n", message); - return; - } - ctx->adapter = std::move(adapter); - }), - UINT64_MAX); - GGML_ASSERT(ctx->adapter != nullptr); - - ctx->adapter.GetLimits(&ctx->limits); - - wgpu::AdapterInfo info{}; -#ifndef __EMSCRIPTEN__ - wgpu::AdapterPropertiesSubgroupMatrixConfigs subgroup_matrix_configs{}; - if (ctx->adapter.HasFeature(wgpu::FeatureName::ChromiumExperimentalSubgroupMatrix)) { - info.nextInChain = &subgroup_matrix_configs; - } -#endif - ctx->adapter.GetInfo(&info); - - wgpu::SupportedFeatures features; - ctx->adapter.GetFeatures(&features); - // we require f16 support - GGML_ASSERT(ctx->adapter.HasFeature(wgpu::FeatureName::ShaderF16)); - -#ifndef __EMSCRIPTEN__ - // Only support square f16 matrices of size 8 or 16 for now - bool valid_subgroup_matrix_config = false; - if (ctx->adapter.HasFeature(wgpu::FeatureName::ChromiumExperimentalSubgroupMatrix)) { - for (size_t i = 0; i < subgroup_matrix_configs.configCount; i++) { - const wgpu::SubgroupMatrixConfig config = subgroup_matrix_configs.configs[i]; - if (config.M == config.N && config.N == config.K && (config.K == 8 || config.K == 16) && - config.componentType == wgpu::SubgroupMatrixComponentType::F16 && - config.resultComponentType == wgpu::SubgroupMatrixComponentType::F16) { - ctx->sg_mat_m = config.M; - ctx->sg_mat_n = config.N; - ctx->sg_mat_k = config.K; - valid_subgroup_matrix_config = true; - break; - } - } - } - - ctx->supports_subgroup_matrix = valid_subgroup_matrix_config; -#endif - // For subgroup matrix code to be the most efficient, we would like the subgroup size to be consistent and accurate. - // Unfortunately, that is not possible, so we use the maximum subgroup size reported by the adapter. - ctx->max_subgroup_size = info.subgroupMaxSize; - - // Initialize device - std::vector required_features = { wgpu::FeatureName::ShaderF16 }; - -#ifndef __EMSCRIPTEN__ - required_features.push_back(wgpu::FeatureName::ImplicitDeviceSynchronization); - if (ctx->supports_subgroup_matrix) { - required_features.push_back(wgpu::FeatureName::Subgroups); - required_features.push_back(wgpu::FeatureName::ChromiumExperimentalSubgroupMatrix); - } -#endif - -#ifdef GGML_WEBGPU_GPU_PROFILE - required_features.push_back(wgpu::FeatureName::TimestampQuery); -#endif - - wgpu::DeviceDescriptor dev_desc; - dev_desc.requiredLimits = &ctx->limits; - dev_desc.requiredFeatures = required_features.data(); - dev_desc.requiredFeatureCount = required_features.size(); - dev_desc.SetDeviceLostCallback( - wgpu::CallbackMode::AllowSpontaneous, - [](const wgpu::Device & device, wgpu::DeviceLostReason reason, wgpu::StringView message) { - GGML_UNUSED(device); - GGML_UNUSED(reason); - GGML_UNUSED(message); - //TODO: uncomment once proper free logic is in place - //GGML_LOG_ERROR("ggml_webgpu: Device lost! Reason: %d, Message: %s\n", static_cast(reason), - //std::string(message).c_str()); - }); - dev_desc.SetUncapturedErrorCallback( - [](const wgpu::Device & device, wgpu::ErrorType reason, wgpu::StringView message) { - GGML_UNUSED(device); - GGML_ABORT("ggml_webgpu: Device error! Reason: %d, Message: %s\n", static_cast(reason), - std::string(message).c_str()); - }); - -#ifndef __EMSCRIPTEN__ - // Enable Dawn-specific toggles to increase native performance - // TODO: Maybe WebGPU needs a "fast" mode where you can request compilers skip adding checks like these, - // only for native performance? - const char * const deviceEnabledToggles[] = { "skip_validation", "disable_robustness", "disable_workgroup_init", - "disable_polyfills_on_integer_div_and_mod" }; - const char * const deviceDisabledToggles[] = { "timestamp_quantization" }; - wgpu::DawnTogglesDescriptor deviceTogglesDesc; - deviceTogglesDesc.enabledToggles = deviceEnabledToggles; - deviceTogglesDesc.enabledToggleCount = 4; - deviceTogglesDesc.disabledToggles = deviceDisabledToggles; - deviceTogglesDesc.disabledToggleCount = 1; - - dev_desc.nextInChain = &deviceTogglesDesc; -#endif - - ctx->instance.WaitAny(ctx->adapter.RequestDevice( - &dev_desc, wgpu::CallbackMode::AllowSpontaneous, - [ctx](wgpu::RequestDeviceStatus status, wgpu::Device device, wgpu::StringView message) { - if (status != wgpu::RequestDeviceStatus::Success) { - GGML_LOG_ERROR("ggml_webgpu: Failed to get a device: %s\n", - std::string(message).c_str()); - return; - } - ctx->device = std::move(device); - }), - UINT64_MAX); - GGML_ASSERT(ctx->device != nullptr); - - // Initialize (compute) queue - ctx->queue = ctx->device.GetQueue(); - - // Create buffer pool for shader parameters - ctx->param_buf_pool.init(ctx->device, WEBGPU_NUM_PARAM_BUFS, WEBGPU_PARAMS_BUF_SIZE_BYTES, - wgpu::BufferUsage::CopyDst | wgpu::BufferUsage::Uniform, - wgpu::BufferUsage::CopySrc | wgpu::BufferUsage::MapWrite); - -#ifdef GGML_WEBGPU_GPU_PROFILE - // Initialize buffer pool for timestamp queries (profiling) - ctx->timestamp_query_buf_pool.init(ctx->device, WEBGPU_NUM_TIMESTAMP_QUERY_BUFS, - WEBGPU_TIMESTAMP_QUERY_BUF_SIZE_BYTES, - wgpu::BufferUsage::QueryResolve | wgpu::BufferUsage::CopySrc, - wgpu::BufferUsage::MapRead | wgpu::BufferUsage::CopyDst); -#endif - - ctx->set_rows_error_buf_pool.init(ctx->device, WEBGPU_NUM_SET_ROWS_ERROR_BUFS, WEBGPU_SET_ROWS_ERROR_BUF_SIZE_BYTES, - wgpu::BufferUsage::CopySrc | wgpu::BufferUsage::Storage, - wgpu::BufferUsage::CopyDst | wgpu::BufferUsage::MapRead); - - ggml_webgpu_init_memset_pipeline(ctx); - ggml_webgpu_init_mul_mat_pipeline(ctx); - ggml_webgpu_init_get_rows_pipeline(ctx); - ggml_webgpu_init_cpy_pipeline(ctx); - ggml_webgpu_init_add_pipeline(ctx); - ggml_webgpu_init_sub_pipeline(ctx); - ggml_webgpu_init_mul_pipeline(ctx); - ggml_webgpu_init_div_pipeline(ctx); - ggml_webgpu_init_rms_norm_pipeline(ctx); - ggml_webgpu_init_rope_pipeline(ctx); - ggml_webgpu_init_glu_pipeline(ctx); - ggml_webgpu_init_scale_pipeline(ctx); - ggml_webgpu_init_soft_max_pipeline(ctx); - -#ifdef GGML_WEBGPU_DEBUG - // Initialize debug buffers - ggml_webgpu_create_buffer(ctx->device, ctx->debug_host_buf, WEBGPU_DEBUG_BUF_ELEMS * sizeof(uint32_t), - wgpu::BufferUsage::CopyDst | wgpu::BufferUsage::MapRead, "debug_host_buf"); - ggml_webgpu_create_buffer(ctx->device, ctx->debug_dev_buf, WEBGPU_DEBUG_BUF_ELEMS * sizeof(uint32_t), - wgpu::BufferUsage::Storage | wgpu::BufferUsage::CopySrc, "debug_dev_buf"); -#endif + create_webgpu_device(reg_ctx); static ggml_backend_webgpu_device_context device_ctx; - device_ctx.webgpu_ctx = ctx; - device_ctx.device_name = GGML_WEBGPU_NAME; - device_ctx.device_desc = info.description; - - GGML_LOG_INFO( - "ggml_webgpu: adapter_info: vendor_id: %u | vendor: %s | architecture: %s | device_id: %u | name: %s | " - "device_desc: %s\n", - info.vendorID, std::string(info.vendor).c_str(), std::string(info.architecture).c_str(), info.deviceID, - std::string(info.device).c_str(), std::string(info.description).c_str()); - + device_ctx.device_name = GGML_WEBGPU_NAME; + device_ctx.device_desc = GGML_WEBGPU_NAME; + device_ctx.webgpu_global_ctx = reg_ctx->webgpu_global_ctx; // See GGML Backend Device Interface section static ggml_backend_device device = { /* .iface = */ ggml_backend_webgpu_device_i, @@ -3354,7 +3433,7 @@ static ggml_backend_dev_t ggml_backend_webgpu_reg_get_device(ggml_backend_reg_t /* .context = */ &device_ctx, }; - WEBGPU_CPU_PROFILE_TOTAL_END(reg_get_device, ctx); + WEBGPU_CPU_PROFILE_TOTAL_END(reg_get_device, reg_ctx->webgpu_global_ctx); return &device; } @@ -3370,10 +3449,7 @@ static const struct ggml_backend_reg_i ggml_backend_webgpu_reg_i = { ggml_backend_reg_t ggml_backend_webgpu_reg() { WEBGPU_LOG_DEBUG("ggml_backend_webgpu_reg()"); - webgpu_context webgpu_ctx = std::make_shared(); - static ggml_backend_webgpu_reg_context ctx; - ctx.webgpu_ctx = webgpu_ctx; ctx.name = GGML_WEBGPU_NAME; ctx.device_count = 1; @@ -3390,15 +3466,17 @@ ggml_backend_reg_t ggml_backend_webgpu_reg() { instance_descriptor.nextInChain = &instanceTogglesDesc; #endif - webgpu_ctx->instance = wgpu::CreateInstance(&instance_descriptor); + wgpu::Instance inst = wgpu::CreateInstance(&instance_descriptor); + ctx.webgpu_global_ctx = webgpu_global_context(new webgpu_global_context_struct()); + ctx.webgpu_global_ctx->instance = std::move(inst); #ifdef __EMSCRIPTEN__ - if (webgpu_ctx->instance == nullptr) { + if (ctx.webgpu_global_ctx->instance == nullptr) { GGML_LOG_ERROR("ggml_webgpu: Failed to create WebGPU instance. Make sure either -sASYNCIFY or -sJSPI is set\n"); return nullptr; } #endif - GGML_ASSERT(webgpu_ctx->instance != nullptr); + GGML_ASSERT(ctx.webgpu_global_ctx->instance != nullptr); static ggml_backend_reg reg = { /* .api_version = */ GGML_BACKEND_API_VERSION, @@ -3411,7 +3489,7 @@ ggml_backend_reg_t ggml_backend_webgpu_reg() { ggml_backend_t ggml_backend_webgpu_init(void) { ggml_backend_dev_t dev = ggml_backend_reg_dev_get(ggml_backend_webgpu_reg(), 0); - return ggml_backend_webgpu_device_init(dev, nullptr); + return ggml_backend_webgpu_backend_init(dev, nullptr); } GGML_BACKEND_DL_IMPL(ggml_backend_webgpu_reg) diff --git a/ggml/src/ggml-zendnn/CMakeLists.txt b/ggml/src/ggml-zendnn/CMakeLists.txt index bdbfc74369..f5cf6eedd3 100644 --- a/ggml/src/ggml-zendnn/CMakeLists.txt +++ b/ggml/src/ggml-zendnn/CMakeLists.txt @@ -21,7 +21,7 @@ if (NOT ZENDNN_ROOT OR ZENDNN_ROOT STREQUAL "" OR ZENDNN_ROOT STREQUAL "OFF") ExternalProject_Add( zendnn GIT_REPOSITORY https://github.com/amd/ZenDNN.git - GIT_TAG zendnnl + GIT_TAG 21ce8f7879c86bf3637f707fae6f29e0951db5fe PREFIX ${ZENDNN_PREFIX} SOURCE_DIR ${ZENDNN_SOURCE_DIR} BINARY_DIR ${ZENDNN_BUILD_DIR} diff --git a/ggml/src/ggml-zendnn/ggml-zendnn.cpp b/ggml/src/ggml-zendnn/ggml-zendnn.cpp index afbecde7a5..551c15bb4a 100644 --- a/ggml/src/ggml-zendnn/ggml-zendnn.cpp +++ b/ggml/src/ggml-zendnn/ggml-zendnn.cpp @@ -2,7 +2,6 @@ #include "ggml-backend-impl.h" #include "ggml-impl.h" -#include "ggml-cpu.h" #include "zendnnl.hpp" #include @@ -122,8 +121,8 @@ static void ggml_zendnn_compute_forward_mul_mat( GGML_TENSOR_BINARY_OP_LOCALS - ggml_type const vec_dot_type = ggml_get_type_traits_cpu(src0->type)->vec_dot_type; - ggml_from_float_t const from_float = ggml_get_type_traits_cpu(vec_dot_type)->from_float; + ggml_type const vec_dot_type = src0->type; + ggml_from_float_t const from_float = ggml_get_type_traits(vec_dot_type)->from_float_ref; GGML_ASSERT(ne0 == ne01); GGML_ASSERT(ne1 == ne11); diff --git a/ggml/src/gguf.cpp b/ggml/src/gguf.cpp index bfab5c4d60..ed0d7f2cae 100644 --- a/ggml/src/gguf.cpp +++ b/ggml/src/gguf.cpp @@ -585,6 +585,14 @@ struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_par break; } + // check that the size of the tensor in bytes is representable + if (ok && uint64_t(ggml_nelements(&info.t)/ggml_blck_size(info.t.type)) > SIZE_MAX/ggml_type_size(info.t.type)) { + GGML_LOG_ERROR("%s: tensor '%s' with shape (%" PRIi64 ", %" PRIi64 ", %" PRIi64 ", %" PRIi64 ") has a size in bytes > %zu\n", + __func__, info.t.name, info.t.ne[0], info.t.ne[1], info.t.ne[2], info.t.ne[3], SIZE_MAX); + ok = false; + break; + } + // calculate byte offsets given the tensor shape and type info.t.nb[0] = type_size; info.t.nb[1] = info.t.nb[0]*(info.t.ne[0]/blck_size); diff --git a/include/llama.h b/include/llama.h index 280745713e..bf4e28a8be 100644 --- a/include/llama.h +++ b/include/llama.h @@ -309,7 +309,7 @@ extern "C" { // Keep the booleans together to avoid misalignment during copy-by-value. bool vocab_only; // only load the vocabulary, no weights bool use_mmap; // use mmap if possible - bool use_direct_io; // use direct io, takes precedence over use_mmap + bool use_direct_io; // use direct io, takes precedence over use_mmap when supported bool use_mlock; // force system to keep model in RAM bool check_tensors; // validate model tensor data bool use_extra_bufts; // use extra buffer types (used for weight repacking) @@ -489,6 +489,7 @@ extern "C" { // - returns true if the parameters could be successfully modified to fit device memory // - this function is NOT thread safe because it modifies the global llama logger state // - only parameters that have the same value as in llama_default_model_params are modified + // with the exception of the context size which is modified if and only if equal to 0 LLAMA_API enum llama_params_fit_status llama_params_fit( const char * path_model, struct llama_model_params * mparams, @@ -1475,12 +1476,12 @@ extern "C" { /// @details Build a split GGUF final path for this chunk. /// llama_split_path(split_path, sizeof(split_path), "/models/ggml-model-q4_0", 2, 4) => split_path = "/models/ggml-model-q4_0-00002-of-00004.gguf" // Returns the split_path length. - LLAMA_API int llama_split_path(char * split_path, size_t maxlen, const char * path_prefix, int split_no, int split_count); + LLAMA_API int32_t llama_split_path(char * split_path, size_t maxlen, const char * path_prefix, int32_t split_no, int32_t split_count); /// @details Extract the path prefix from the split_path if and only if the split_no and split_count match. /// llama_split_prefix(split_prefix, 64, "/models/ggml-model-q4_0-00002-of-00004.gguf", 2, 4) => split_prefix = "/models/ggml-model-q4_0" // Returns the split_prefix length. - LLAMA_API int llama_split_prefix(char * split_prefix, size_t maxlen, const char * split_path, int split_no, int split_count); + LLAMA_API int32_t llama_split_prefix(char * split_prefix, size_t maxlen, const char * split_path, int32_t split_no, int32_t split_count); // Print system information LLAMA_API const char * llama_print_system_info(void); diff --git a/src/llama-context.cpp b/src/llama-context.cpp index a35cf5a94b..10b306a853 100644 --- a/src/llama-context.cpp +++ b/src/llama-context.cpp @@ -253,11 +253,7 @@ llama_context::llama_context( // graph outputs buffer { - // resized during inference when a batch uses more outputs - // Create a dummy batch for initialization. - llama_batch dummy_batch = {}; - dummy_batch.n_tokens = 0; - if (output_reserve(params.n_seq_max, dummy_batch) < params.n_seq_max) { + if (output_reserve(params.n_seq_max) < params.n_seq_max) { throw std::runtime_error("failed to reserve initial output buffer"); } @@ -793,7 +789,7 @@ float * llama_context::get_embeddings_ith(int32_t i) { throw std::runtime_error(format("corrupt output buffer (j=%" PRId64 ", n_outputs=%d)", j, n_outputs)); } - const uint32_t n_embd_out = model.hparams.get_n_embd_out(); + const uint32_t n_embd_out = model.hparams.n_embd_out(); return embd + j*n_embd_out; } catch (const std::exception & err) { LLAMA_LOG_ERROR("%s: invalid embeddings id %d, reason: %s\n", __func__, i, err.what()); @@ -1225,7 +1221,7 @@ int llama_context::encode(const llama_batch & batch_inp) { n_queued_tokens += n_tokens; // reserve output buffer - if (output_reserve(n_tokens, batch_inp) < n_tokens) { + if (output_reserve(n_tokens) < n_tokens) { LLAMA_LOG_ERROR("%s: could not reserve space for batch with %u outputs\n", __func__, n_tokens); return -2; }; @@ -1279,7 +1275,7 @@ int llama_context::encode(const llama_batch & batch_inp) { { // extract token embeddings GGML_ASSERT(embd != nullptr); - const uint32_t n_embd_out = hparams.get_n_embd_out(); + const uint32_t n_embd_out = hparams.n_embd_out(); GGML_ASSERT(n_tokens*n_embd_out <= (int64_t) embd_size); ggml_backend_tensor_get_async(backend_embd, t_embd, embd, 0, n_tokens*n_embd_out*sizeof(float)); @@ -1456,6 +1452,23 @@ static void copy_tensor_async_candidates( } } +static bool needs_raw_logits(const llama_ubatch & ubatch, const std::map & samplers) { + for (uint32_t i = 0; i < ubatch.n_tokens; i++) { + if (!ubatch.output[i]) { + continue; + } + + // Check if the output token has at least one sequence without a backend sampler. + for (int32_t j = 0; j < ubatch.n_seq_id[i]; ++j) { + llama_seq_id seq_id = ubatch.seq_id[i][j]; + if (samplers.find(seq_id) == samplers.end()) { + return true; + } + } + } + return false; // all sequences use backend sampling +} + int llama_context::decode(const llama_batch & batch_inp) { GGML_ASSERT((!batch_inp.token && batch_inp.embd) || (batch_inp.token && !batch_inp.embd)); // NOLINT @@ -1588,7 +1601,7 @@ int llama_context::decode(const llama_batch & batch_inp) { } // reserve output buffer - if (output_reserve(n_outputs_all, balloc->get_batch()) < n_outputs_all) { + if (output_reserve(n_outputs_all) < n_outputs_all) { LLAMA_LOG_ERROR("%s: could not reserve space for batch with %d outputs\n", __func__, n_outputs_all); return -2; }; @@ -1661,10 +1674,7 @@ int llama_context::decode(const llama_batch & batch_inp) { } // extract logits - // For multi-sequence batches that mix backend samplers and CPU sampler - // this is currently inefficient as we copy all logits even for the - // backend sampled tokens. - if (logits && t_logits && n_outputs > 0) { + if (logits && t_logits && n_outputs > 0 && needs_raw_logits(ubatch, sampling.samplers)) { ggml_backend_t backend_res = ggml_backend_sched_get_tensor_backend(sched.get(), t_logits); GGML_ASSERT(backend_res != nullptr); GGML_ASSERT(logits != nullptr); @@ -1688,7 +1698,7 @@ int llama_context::decode(const llama_batch & batch_inp) { { // extract token embeddings GGML_ASSERT(embd != nullptr); - const uint32_t n_embd_out = hparams.get_n_embd_out(); + const uint32_t n_embd_out = hparams.n_embd_out(); float * embd_out = embd + n_outputs_prev*n_embd_out; if (n_outputs) { @@ -1734,11 +1744,8 @@ int llama_context::decode(const llama_batch & batch_inp) { } } - // This flag indicates whether a backend sampler has actually sampled a specific - // token, or if it has produced probabilites. If true, we can skip the normal copying of logits and embeddings. - const bool has_sampled = !res->t_sampled.empty() || !res->t_sampled_probs.empty() || !res->t_sampled_logits.empty(); - - if (has_samplers && has_sampled) { + // Copy backend sampling output if this ubatch produced any sampling tensors. + if (has_samplers && (!res->t_sampled.empty() || !res->t_sampled_probs.empty() || !res->t_sampled_logits.empty())) { const auto seq_to_output_row = build_seq_to_output_row(ubatch, n_outputs_prev); const auto stride = n_vocab; @@ -1813,7 +1820,8 @@ int llama_context::decode(const llama_batch & batch_inp) { // output // -uint32_t llama_context::output_reserve(int32_t n_outputs, const llama_batch & batch) { +uint32_t llama_context::output_reserve(int32_t n_outputs) { + const auto & hparams = model.hparams; const auto & vocab = model.vocab; @@ -1821,7 +1829,7 @@ uint32_t llama_context::output_reserve(int32_t n_outputs, const llama_batch & ba const auto n_batch = cparams.n_batch; const auto n_vocab = vocab.n_tokens(); - const auto n_embd_out = hparams.get_n_embd_out(); + const auto n_embd_out = hparams.n_embd_out(); bool has_logits = true; bool has_embd = cparams.embeddings; @@ -1832,45 +1840,16 @@ uint32_t llama_context::output_reserve(int32_t n_outputs, const llama_batch & ba has_embd = true; } - // Check which sampling modes are needed for the current batch. - // TODO: avoid this branching by working with the worst-case - bool has_sampling = false; - bool cpu_logits = false; - - if (batch.logits) { - for (int32_t i = 0; i < batch.n_tokens; i++) { - if (!batch.logits[i]) { - continue; - } - for (int32_t j = 0; j < batch.n_seq_id[i]; j++) { - llama_seq_id seq_id = batch.seq_id[i][j]; - if (sampling.samplers.find(seq_id) != sampling.samplers.end()) { - has_sampling = true; - } else { - cpu_logits = true; - } - } - } - } else { - // When batch.logits is nullptr (when loading state with a dummy batch), - // allocate CPU logits. - cpu_logits = true; - } size_t backend_float_count = 0; size_t backend_token_count = 0; - // Allocate CPU logits buffer only if needed by sequences in this batch - logits_size = (has_logits && cpu_logits) ? n_vocab*n_outputs_max : 0; + logits_size = has_logits ? n_vocab*n_outputs_max : 0; embd_size = has_embd ? n_embd_out*n_outputs_max : 0; - // TODO: avoid this branching by working with the worst-case - if (!has_sampling) { - sampling.logits_size = 0; - sampling.probs_size = 0; - sampling.sampled_size = 0; - sampling.candidates_size = 0; - } else { + // Allocate backend sampling output buffers if there are backend samplers configured. + const bool has_sampling = !sampling.samplers.empty(); + if (has_sampling) { sampling.logits_size = n_vocab*n_outputs_max; sampling.probs_size = n_vocab*n_outputs_max; sampling.sampled_size = n_outputs_max; @@ -1928,7 +1907,7 @@ uint32_t llama_context::output_reserve(int32_t n_outputs, const llama_batch & ba size_t offset = 0; uint8_t * base = (uint8_t *) output_base; - logits = (has_logits && cpu_logits) ? output_base : nullptr; + logits = has_logits ? output_base : nullptr; offset += logits_size * sizeof(float); embd = has_embd ? (float *) (base + offset) : nullptr; @@ -2173,13 +2152,6 @@ llm_graph_cb llama_context::graph_get_cb() const { ggml_set_name(cur, name); } - if (!cparams.offload_kqv) { - if (strcmp(name, "kqv_merged_cont") == 0) { - // all nodes between the KV store and the attention output are run on the CPU - ggml_backend_sched_set_tensor_backend(sched.get(), cur, backend_cpu); - } - } - // norm may be automatically assigned to the backend of the previous layer, increasing data transfer between backends // FIXME: fix in ggml_backend_sched const bool full_offload = model.n_gpu_layers() > model.hparams.n_layer; @@ -2559,6 +2531,7 @@ size_t llama_context::state_write_data(llama_io_write_i & io) { } } + // [TAG_CONTEXT_STATE_LOGITS] // write logits { LLAMA_LOG_DEBUG("%s: - writing logits\n", __func__); @@ -2620,10 +2593,7 @@ size_t llama_context::state_read_data(llama_io_read_i & io) { auto n_outputs = this->n_outputs; io.read_to(&n_outputs, sizeof(n_outputs)); - // Create a dummy batch for state loading. - llama_batch dummy_batch = {}; - dummy_batch.n_tokens = 0; - if (n_outputs > output_reserve(n_outputs, dummy_batch)) { + if (n_outputs > output_reserve(n_outputs)) { throw std::runtime_error("could not reserve outputs"); } @@ -2868,7 +2838,7 @@ void llama_context::opt_epoch_iter( } // reserve output buffer - if (output_reserve(n_outputs_all, balloc->get_batch()) < n_outputs_all) { + if (output_reserve(n_outputs_all) < n_outputs_all) { LLAMA_LOG_ERROR("%s: could not reserve space for batch with %d outputs\n", __func__, n_outputs_all); GGML_ABORT("TODO: handle this error"); }; diff --git a/src/llama-context.h b/src/llama-context.h index 86decc05fb..8e71cdd1dc 100644 --- a/src/llama-context.h +++ b/src/llama-context.h @@ -212,7 +212,7 @@ private: // Make sure enough space is available for outputs. // Returns max number of outputs for which space was reserved. - uint32_t output_reserve(int32_t n_outputs, const llama_batch & batch); + uint32_t output_reserve(int32_t n_outputs); void output_reorder(); diff --git a/src/llama-graph.cpp b/src/llama-graph.cpp index f9ed87cce1..b3198b7e3a 100644 --- a/src/llama-graph.cpp +++ b/src/llama-graph.cpp @@ -407,6 +407,27 @@ bool llm_graph_input_attn_kv::can_reuse(const llm_graph_params & params) { return res; } +void llm_graph_input_attn_k::set_input(const llama_ubatch * ubatch) { + mctx->set_input_k_idxs(self_k_idxs, ubatch); + + mctx->set_input_kq_mask(self_kq_mask, ubatch, cparams.causal_attn); +} + +bool llm_graph_input_attn_k::can_reuse(const llm_graph_params & params) { + const auto * mctx = static_cast(params.mctx); + + this->mctx = mctx; + + bool res = true; + + res &= self_k_idxs->ne[0] == params.ubatch.n_tokens; + + res &= self_kq_mask->ne[0] == mctx->get_n_kv(); + res &= self_kq_mask->ne[1] == params.ubatch.n_tokens; + + return res; +} + void llm_graph_input_attn_kv_iswa::set_input(const llama_ubatch * ubatch) { mctx->get_base()->set_input_k_idxs(self_k_idxs, ubatch); mctx->get_base()->set_input_v_idxs(self_v_idxs, ubatch); @@ -1596,11 +1617,6 @@ ggml_tensor * llm_graph_context::build_attn_mha( v = ggml_transpose(ctx0, v); } - // TODO: update llama_kv_cache to not store V cache in the MLA case and automatically return a view of K - if (v_mla) { - v = ggml_view_4d(ctx0, k, v->ne[0], v->ne[1], v->ne[2], v->ne[3], k->nb[1], k->nb[2], k->nb[3], 0); - } - // this can happen when KV cache is not used (e.g. an embedding model with non-causal attn) if (k->type == GGML_TYPE_F32) { k = ggml_cast(ctx0, k, GGML_TYPE_F16); @@ -1614,6 +1630,11 @@ ggml_tensor * llm_graph_context::build_attn_mha( hparams.attn_soft_cap ? hparams.f_attn_logit_softcapping : 0.0f); cb(cur, LLAMA_TENSOR_NAME_FATTN, il); + if (!cparams.offload_kqv) { + // all nodes between the KV store and the attention output are run on the CPU + ggml_backend_sched_set_tensor_backend(sched, cur, backend_cpu); + } + ggml_flash_attn_ext_add_sinks(cur, sinks); ggml_flash_attn_ext_set_prec (cur, GGML_PREC_F32); @@ -1823,9 +1844,11 @@ ggml_tensor * llm_graph_context::build_attn( ggml_tensor * v_cur, ggml_tensor * kq_b, ggml_tensor * sinks, - ggml_tensor * v_mla, + ggml_tensor * v_mla, // TODO: remove float kq_scale, int il) const { + GGML_ASSERT(v_mla == nullptr); + // these nodes are added to the graph together so that they are not reordered // by doing so, the number of splits in the graph is reduced // expand k later to enable rope fusion which directly writes into k-v cache @@ -1868,6 +1891,93 @@ ggml_tensor * llm_graph_context::build_attn( return cur; } +static std::unique_ptr build_attn_inp_k_impl( + ggml_context * ctx0, + const llama_ubatch & ubatch, + const llama_hparams & hparams, + const llama_cparams & cparams, + const llama_kv_cache_context * mctx_cur) { + + auto inp = std::make_unique(hparams, cparams, mctx_cur); + + { + GGML_ASSERT(hparams.swa_type == LLAMA_SWA_TYPE_NONE && "Use llama_kv_cache_iswa for SWA"); + + const auto n_kv = mctx_cur->get_n_kv(); + const auto n_tokens = ubatch.n_tokens; + const auto n_stream = cparams.kv_unified ? 1 : ubatch.n_seqs_unq; + + inp->self_k_idxs = mctx_cur->build_input_k_idxs(ctx0, ubatch); + + inp->self_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, n_tokens/n_stream, 1, n_stream); + ggml_set_input(inp->self_kq_mask); + + inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask, GGML_TYPE_F16) : inp->self_kq_mask; + } + + return inp; +} + +llm_graph_input_attn_k * llm_graph_context::build_attn_inp_k() const { + const auto * mctx_cur = static_cast(mctx); + + auto inp = build_attn_inp_k_impl(ctx0, ubatch, hparams, cparams, mctx_cur); + + return (llm_graph_input_attn_k *) res->add_input(std::move(inp)); +} + +ggml_tensor * llm_graph_context::build_attn( + llm_graph_input_attn_k * inp, + ggml_tensor * wo, + ggml_tensor * wo_b, + ggml_tensor * q_cur, + ggml_tensor * k_cur, + ggml_tensor * v_cur, + ggml_tensor * kq_b, + ggml_tensor * sinks, + ggml_tensor * v_mla, + float kq_scale, + int il) const { + // these nodes are added to the graph together so that they are not reordered + // by doing so, the number of splits in the graph is reduced + // expand k later to enable rope fusion which directly writes into k-v cache + ggml_build_forward_expand(gf, q_cur); + ggml_build_forward_expand(gf, v_cur); + ggml_build_forward_expand(gf, k_cur); + + const auto * mctx_cur = inp->mctx; + + // store to KV cache + { + const auto & k_idxs = inp->get_k_idxs(); + + ggml_build_forward_expand(gf, mctx_cur->cpy_k(ctx0, k_cur, k_idxs, il)); + } + + const auto & kq_mask = inp->get_kq_mask(); + + ggml_tensor * q = q_cur; + ggml_tensor * k = mctx_cur->get_k(ctx0, il); + ggml_tensor * v = ggml_view_4d(ctx0, k, v_cur->ne[0], k->ne[1], k->ne[2], k->ne[3], k->nb[1], k->nb[2], k->nb[3], 0); + + ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, sinks, v_mla, kq_scale, il); + cb(cur, "kqv_out", il); + + if (wo) { + cur = build_lora_mm(wo, cur); + if (arch == LLM_ARCH_GLM4 || arch == LLM_ARCH_GLM4_MOE) { + // GLM4 and GLM4_MOE seem to have numerical issues with half-precision accumulators + ggml_mul_mat_set_prec(cur, GGML_PREC_F32); + } + } + + if (wo_b) { + cur = ggml_add(ctx0, cur, wo_b); + } + + return cur; +} + ggml_tensor * llm_graph_context::build_attn( llm_graph_input_attn_kv_iswa * inp, ggml_tensor * wo, diff --git a/src/llama-graph.h b/src/llama-graph.h index 242a046d56..4090d8116c 100644 --- a/src/llama-graph.h +++ b/src/llama-graph.h @@ -317,6 +317,39 @@ public: const llama_kv_cache_context * mctx; }; +// V-less input for the KV cache +// ref: https://github.com/ggml-org/llama.cpp/pull/19067 +class llm_graph_input_attn_k : public llm_graph_input_i { +public: + llm_graph_input_attn_k( + const llama_hparams & hparams, + const llama_cparams & cparams, + const llama_kv_cache_context * mctx) : + hparams(hparams), + cparams(cparams), + mctx(mctx) { + } + ~llm_graph_input_attn_k() = default; + + void set_input(const llama_ubatch * ubatch) override; + + bool can_reuse(const llm_graph_params & params) override; + + ggml_tensor * get_k_idxs() const { return self_k_idxs; } + + ggml_tensor * get_kq_mask() const { return self_kq_mask_cnv; } + + ggml_tensor * self_k_idxs = nullptr; // I64 [n_batch] + + ggml_tensor * self_kq_mask = nullptr; // F32 [n_kv, n_batch/n_stream, 1, n_stream] + ggml_tensor * self_kq_mask_cnv = nullptr; // [n_kv, n_batch/n_stream, 1, n_stream] + + const llama_hparams hparams; + const llama_cparams cparams; + + const llama_kv_cache_context * mctx; +}; + class llm_graph_input_attn_kv_iswa : public llm_graph_input_i { public: llm_graph_input_attn_kv_iswa( @@ -833,6 +866,21 @@ struct llm_graph_context { ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens] ggml_tensor * kq_b, ggml_tensor * sinks, // [n_head_q] + ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v] // TODO: remove + float kq_scale, + int il) const; + + llm_graph_input_attn_k * build_attn_inp_k() const; + + ggml_tensor * build_attn( + llm_graph_input_attn_k * inp, + ggml_tensor * wo, + ggml_tensor * wo_b, + ggml_tensor * q_cur, // [n_embd_head_q, n_head_q, n_tokens] + ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens] + ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens] + ggml_tensor * kq_b, + ggml_tensor * sinks, // [n_head_q] ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v] float kq_scale, int il) const; diff --git a/src/llama-hparams.cpp b/src/llama-hparams.cpp index 5f1df995f3..392f9160ce 100644 --- a/src/llama-hparams.cpp +++ b/src/llama-hparams.cpp @@ -72,8 +72,8 @@ uint32_t llama_hparams::n_embd_inp() const { return n_embd_inp; } -uint32_t llama_hparams::get_n_embd_out() const { - return n_embd_out > 0 ? n_embd_out : n_embd; +uint32_t llama_hparams::n_embd_out() const { + return n_embd_out_impl > 0 ? n_embd_out_impl : n_embd; } uint32_t llama_hparams::n_embd_k_gqa(uint32_t il) const { @@ -175,6 +175,21 @@ bool llama_hparams::is_swa(uint32_t il) const { GGML_ABORT("fatal error"); } +bool llama_hparams::is_mla() const { + assert((n_embd_head_k_mla_impl == 0 && n_embd_head_v_mla_impl == 0) || + (n_embd_head_k_mla_impl != 0 && n_embd_head_v_mla_impl != 0)); + + return n_embd_head_k_mla_impl != 0 && n_embd_head_v_mla_impl != 0; +} + +uint32_t llama_hparams::n_embd_head_k_mla() const { + return is_mla() ? n_embd_head_k_mla_impl : n_embd_head_k; +} + +uint32_t llama_hparams::n_embd_head_v_mla() const { + return is_mla() ? n_embd_head_v_mla_impl : n_embd_head_v; +} + bool llama_hparams::has_kv(uint32_t il) const { if (n_layer_kv_from_start >= 0) { if (il < (uint32_t) n_layer_kv_from_start) { diff --git a/src/llama-hparams.h b/src/llama-hparams.h index 2bf8665520..caed0ec1b7 100644 --- a/src/llama-hparams.h +++ b/src/llama-hparams.h @@ -53,8 +53,8 @@ struct llama_hparams { uint32_t n_rel_attn_bkts = 0; // note: deepseek2 using MLA converts into MQA with larger heads, then decompresses to MHA - uint32_t n_embd_head_k_mla = 0; - uint32_t n_embd_head_v_mla = 0; + uint32_t n_embd_head_k_mla_impl = 0; + uint32_t n_embd_head_v_mla_impl = 0; // for WavTokenizer struct llama_hparams_posnet posnet; @@ -164,7 +164,7 @@ struct llama_hparams { uint32_t n_cls_out = 1; // output embedding dimension (0 = use n_embd) - uint32_t n_embd_out = 0; + uint32_t n_embd_out_impl = 0; // llama4 smallthinker uint32_t n_moe_layer_step = 0; @@ -239,7 +239,7 @@ struct llama_hparams { uint32_t n_embd_inp() const; // dimension of output embeddings - uint32_t get_n_embd_out() const; + uint32_t n_embd_out() const; // dimension of key embeddings across all k-v heads uint32_t n_embd_k_gqa(uint32_t il = 0) const; @@ -269,6 +269,12 @@ struct llama_hparams { bool is_swa(uint32_t il) const; + // note: currently only support if either all or none of the layers are MLA + bool is_mla() const; + + uint32_t n_embd_head_k_mla() const; + uint32_t n_embd_head_v_mla() const; + bool has_kv(uint32_t il) const; // number of layers for which has_kv() returns true diff --git a/src/llama-kv-cache.cpp b/src/llama-kv-cache.cpp index a7327c4987..f3c9b49f30 100644 --- a/src/llama-kv-cache.cpp +++ b/src/llama-kv-cache.cpp @@ -97,6 +97,8 @@ llama_kv_cache::llama_kv_cache( __func__, hparams.n_embd_v_gqa_max()); } + const bool is_mla = hparams.is_mla(); + for (uint32_t il = 0; il < hparams.n_layer; il++) { if (!hparams.has_kv(il)) { LLAMA_LOG_DEBUG("%s: layer %3d: does not have KV cache\n", __func__, il); @@ -130,18 +132,21 @@ llama_kv_cache::llama_kv_cache( throw std::runtime_error("failed to create ggml context for kv cache"); } - ggml_tensor * k = ggml_new_tensor_3d(ctx, type_k, n_embd_k_gqa, kv_size, n_stream); - ggml_tensor * v = ggml_new_tensor_3d(ctx, type_v, n_embd_v_gqa, kv_size, n_stream); + const bool has_k = true; + const bool has_v = !is_mla; - ggml_format_name(k, "cache_k_l%d", il); - ggml_format_name(v, "cache_v_l%d", il); + ggml_tensor * k = has_k ? ggml_new_tensor_3d(ctx, type_k, n_embd_k_gqa, kv_size, n_stream) : nullptr; + ggml_tensor * v = has_v ? ggml_new_tensor_3d(ctx, type_v, n_embd_v_gqa, kv_size, n_stream) : nullptr; + + has_k && ggml_format_name(k, "cache_k_l%d", il); + has_v && ggml_format_name(v, "cache_v_l%d", il); std::vector k_stream; std::vector v_stream; for (uint32_t s = 0; s < n_stream; ++s) { - k_stream.push_back(ggml_view_2d(ctx, k, n_embd_k_gqa, kv_size, k->nb[1], s*k->nb[2])); - v_stream.push_back(ggml_view_2d(ctx, v, n_embd_v_gqa, kv_size, v->nb[1], s*v->nb[2])); + k_stream.push_back(has_k ? ggml_view_2d(ctx, k, n_embd_k_gqa, kv_size, k->nb[1], s*k->nb[2]) : nullptr); + v_stream.push_back(has_v ? ggml_view_2d(ctx, v, n_embd_v_gqa, kv_size, v->nb[1], s*v->nb[2]) : nullptr); } map_layer_ids[il] = layers.size(); @@ -647,7 +652,10 @@ bool llama_kv_cache::update(llama_context * lctx, bool do_shift, const stream_co const auto & layer = layers[il]; ggml_backend_tensor_copy(layer.k_stream[ssrc], layer.k_stream[sdst]); - ggml_backend_tensor_copy(layer.v_stream[ssrc], layer.v_stream[sdst]); + + if (layer.v_stream[ssrc]) { + ggml_backend_tensor_copy(layer.v_stream[ssrc], layer.v_stream[sdst]); + } } } } @@ -1516,7 +1524,7 @@ size_t llama_kv_cache::size_v_bytes() const { size_t size_v_bytes = 0; for (const auto & layer : layers) { - size_v_bytes += ggml_nbytes(layer.v); + size_v_bytes += layer.v ? ggml_nbytes(layer.v) : 0; } return size_v_bytes; @@ -1798,6 +1806,9 @@ void llama_kv_cache::state_write_data(llama_io_write_i & io, const cell_ranges_t const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il); auto * v = layer.v_stream[cr.strm]; + if (!v) { + continue; + } // Write value type const int32_t v_type_i = (int32_t) v->type; @@ -1824,6 +1835,9 @@ void llama_kv_cache::state_write_data(llama_io_write_i & io, const cell_ranges_t const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il); auto * v = layer.v_stream[cr.strm]; + if (!v) { + continue; + } // Write value type const int32_t v_type_i = (int32_t) v->type; @@ -2027,6 +2041,9 @@ bool llama_kv_cache::state_read_data(llama_io_read_i & io, uint32_t strm, uint32 const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il); auto * v = layer.v_stream[strm]; + if (!v) { + continue; + } // Read type of value int32_t v_type_i_ref; @@ -2068,6 +2085,9 @@ bool llama_kv_cache::state_read_data(llama_io_read_i & io, uint32_t strm, uint32 const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il); auto * v = layer.v_stream[strm]; + if (!v) { + continue; + } // Read type of value int32_t v_type_i_ref; diff --git a/src/llama-model-loader.cpp b/src/llama-model-loader.cpp index 383b8dc761..1501e392ca 100644 --- a/src/llama-model-loader.cpp +++ b/src/llama-model-loader.cpp @@ -541,15 +541,15 @@ llama_model_loader::llama_model_loader( if (use_mmap && use_direct_io) { if (files.back()->has_direct_io()) { - // Disable mmap, as DirectIO is available - use_mmap = false; LLAMA_LOG_WARN("%s: direct I/O is enabled, disabling mmap\n", __func__); + use_mmap = false; } else { - // Disable DirectIO and reopen file using std::fopen for mmap + LLAMA_LOG_WARN("%s: direct I/O is not available, using mmap\n", __func__); use_direct_io = false; + + // reopen file using std::fopen for mmap files.pop_back(); files.emplace_back(new llama_file(fname.c_str(), "rb", false)); - LLAMA_LOG_WARN("%s: direct I/O is not available, using mmap\n", __func__); } } diff --git a/src/llama-model-saver.cpp b/src/llama-model-saver.cpp index ae27c71ce2..36e353074e 100644 --- a/src/llama-model-saver.cpp +++ b/src/llama-model-saver.cpp @@ -146,8 +146,8 @@ void llama_model_saver::add_kv_from_model() { add_kv(LLM_KV_VOCAB_SIZE, vocab.n_tokens()); add_kv(LLM_KV_CONTEXT_LENGTH, hparams.n_ctx_train); add_kv(LLM_KV_EMBEDDING_LENGTH, hparams.n_embd); - if (hparams.n_embd_out > 0) { - add_kv(LLM_KV_EMBEDDING_LENGTH_OUT, hparams.n_embd_out); + if (hparams.n_embd_out_impl > 0) { + add_kv(LLM_KV_EMBEDDING_LENGTH_OUT, hparams.n_embd_out_impl); } add_kv(LLM_KV_BLOCK_COUNT, hparams.n_layer); add_kv(LLM_KV_LEADING_DENSE_BLOCK_COUNT, hparams.n_layer_dense_lead); diff --git a/src/llama-model.cpp b/src/llama-model.cpp index b58b35a426..72490a89b5 100644 --- a/src/llama-model.cpp +++ b/src/llama-model.cpp @@ -512,7 +512,7 @@ void llama_model::load_hparams(llama_model_loader & ml) { ml.get_key(LLM_KV_CONTEXT_LENGTH, hparams.n_ctx_train); ml.get_key(LLM_KV_EMBEDDING_LENGTH, hparams.n_embd); - ml.get_key(LLM_KV_EMBEDDING_LENGTH_OUT, hparams.n_embd_out, false); + ml.get_key(LLM_KV_EMBEDDING_LENGTH_OUT, hparams.n_embd_out_impl, false); ml.get_key(LLM_KV_BLOCK_COUNT, hparams.n_layer); ml.get_key(LLM_KV_EXPERT_COUNT, hparams.n_expert, false); ml.get_key(LLM_KV_EXPERT_USED_COUNT, hparams.n_expert_used, false); @@ -1697,15 +1697,16 @@ void llama_model::load_hparams(llama_model_loader & ml) { case LLM_ARCH_DEEPSEEK2: { // lite variants include DeepSeek-V2-Lite, GigaChat3-10B-A1.8B - bool is_lite = (hparams.n_layer == 27 || hparams.n_layer == 26); + const bool is_lite = (hparams.n_layer == 27 || hparams.n_layer == 26); + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT, hparams.n_layer_dense_lead); if (!is_lite) { ml.get_key(LLM_KV_ATTENTION_Q_LORA_RANK, hparams.n_lora_q); } ml.get_key(LLM_KV_ATTENTION_KV_LORA_RANK, hparams.n_lora_kv); - ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH_MLA, hparams.n_embd_head_k_mla, false); - ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH_MLA, hparams.n_embd_head_v_mla, false); + ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH_MLA, hparams.n_embd_head_k_mla_impl, false); + ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH_MLA, hparams.n_embd_head_v_mla_impl, false); ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp); ml.get_key(LLM_KV_EXPERT_SHARED_COUNT, hparams.n_expert_shared); ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE, hparams.expert_weights_scale, false); @@ -1736,6 +1737,7 @@ void llama_model::load_hparams(llama_model_loader & ml) { switch (hparams.n_layer) { case 27: type = LLM_TYPE_16B; break; + case 47: type = LLM_TYPE_30B_A3B; break; case 60: type = LLM_TYPE_236B; break; case 61: type = LLM_TYPE_671B; break; default: type = LLM_TYPE_UNKNOWN; @@ -4909,14 +4911,11 @@ bool llama_model::load_tensors(llama_model_loader & ml) { } break; case LLM_ARCH_DEEPSEEK2: { - // lite variants include DeepSeek-V2-Lite, GigaChat3-10B-A1.8B - const bool is_lite = (hparams.n_layer == 27 || hparams.n_layer == 26); - - const bool is_mla = (hparams.n_embd_head_k_mla != 0 && hparams.n_embd_head_v_mla != 0); + const bool is_mla = hparams.is_mla(); // note: these are the actual head sizes you get when treating as MHA or after "decompression" using wv_b for MLA - const int64_t n_embd_head_k_mla = is_mla ? hparams.n_embd_head_k_mla : hparams.n_embd_head_k; - const int64_t n_embd_head_v_mla = is_mla ? hparams.n_embd_head_v_mla : hparams.n_embd_head_v; + const int64_t n_embd_head_k_mla = hparams.n_embd_head_k_mla(); + const int64_t n_embd_head_v_mla = hparams.n_embd_head_v_mla(); const int64_t n_embd_head_qk_rope = hparams.n_rot; const int64_t n_embd_head_qk_nope = n_embd_head_k_mla - n_embd_head_qk_rope; @@ -4941,13 +4940,13 @@ bool llama_model::load_tensors(llama_model_loader & ml) { auto & layer = layers[i]; layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0); - if (!is_lite) { + if (q_lora_rank > 0) { layer.attn_q_a_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_A_NORM, "weight", i), {q_lora_rank}, 0); } layer.attn_kv_a_norm = create_tensor(tn(LLM_TENSOR_ATTN_KV_A_NORM, "weight", i), {kv_lora_rank}, 0); - if (!is_lite) { + if (q_lora_rank > 0) { layer.wq_a = create_tensor(tn(LLM_TENSOR_ATTN_Q_A, "weight", i), {n_embd, q_lora_rank}, 0); layer.wq_b = create_tensor(tn(LLM_TENSOR_ATTN_Q_B, "weight", i), {q_lora_rank, n_head * n_embd_head_k_mla}, 0); } else { @@ -6597,7 +6596,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) { } // for LFM2-ColBert-350M - dense_2_out_layers = create_tensor(tn(LLM_TENSOR_DENSE_2_OUT, "weight"), {n_embd, hparams.get_n_embd_out()}, TENSOR_NOT_REQUIRED); + dense_2_out_layers = create_tensor(tn(LLM_TENSOR_DENSE_2_OUT, "weight"), {n_embd, hparams.n_embd_out()}, TENSOR_NOT_REQUIRED); } break; case LLM_ARCH_SMALLTHINKER: { @@ -7316,8 +7315,8 @@ void llama_model::print_info() const { LLAMA_LOG_INFO("%s: n_layer_dense_lead = %d\n", __func__, hparams.n_layer_dense_lead); LLAMA_LOG_INFO("%s: n_lora_q = %d\n", __func__, hparams.n_lora_q); LLAMA_LOG_INFO("%s: n_lora_kv = %d\n", __func__, hparams.n_lora_kv); - LLAMA_LOG_INFO("%s: n_embd_head_k_mla = %d\n", __func__, hparams.n_embd_head_k_mla); - LLAMA_LOG_INFO("%s: n_embd_head_v_mla = %d\n", __func__, hparams.n_embd_head_v_mla); + LLAMA_LOG_INFO("%s: n_embd_head_k_mla = %d\n", __func__, hparams.n_embd_head_k_mla()); + LLAMA_LOG_INFO("%s: n_embd_head_v_mla = %d\n", __func__, hparams.n_embd_head_v_mla()); LLAMA_LOG_INFO("%s: n_ff_exp = %d\n", __func__, hparams.n_ff_exp); LLAMA_LOG_INFO("%s: n_expert_shared = %d\n", __func__, hparams.n_expert_shared); LLAMA_LOG_INFO("%s: expert_weights_scale = %.1f\n", __func__, hparams.expert_weights_scale); @@ -8126,7 +8125,7 @@ llama_model_params llama_model_default_params() { /*.kv_overrides =*/ nullptr, /*.vocab_only =*/ false, /*.use_mmap =*/ true, - /*.use_direct_io =*/ true, + /*.use_direct_io =*/ false, /*.use_mlock =*/ false, /*.check_tensors =*/ false, /*.use_extra_bufts =*/ true, @@ -8162,7 +8161,7 @@ int32_t llama_model_n_embd_inp(const llama_model * model) { } int32_t llama_model_n_embd_out(const llama_model * model) { - return model->hparams.get_n_embd_out(); + return model->hparams.n_embd_out(); } int32_t llama_model_n_layer(const llama_model * model) { diff --git a/src/llama-quant.cpp b/src/llama-quant.cpp index a2b8d4e56c..776222cb6f 100644 --- a/src/llama-quant.cpp +++ b/src/llama-quant.cpp @@ -545,7 +545,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std:: } std::vector splits = {}; - llama_model_loader ml(fname_inp, splits, use_mmap, /*use_direct_io*/ true, /*check_tensors*/ true, /*no_alloc*/ false, kv_overrides, nullptr); + llama_model_loader ml(fname_inp, splits, use_mmap, /*use_direct_io*/ false, /*check_tensors*/ true, /*no_alloc*/ false, kv_overrides, nullptr); ml.init_mappings(false); // no prefetching llama_model model(llama_model_default_params()); diff --git a/src/llama.cpp b/src/llama.cpp index f1096d960e..6da90d6f1f 100644 --- a/src/llama.cpp +++ b/src/llama.cpp @@ -311,8 +311,12 @@ static void llama_params_fit_impl( __func__, hp_nct, cparams->n_ctx, memory_reduction/MiB); } } else { - LLAMA_LOG_INFO("%s: default model context size is %" PRIu32 " which is <= the min. context size of %" PRIu32 " -> no change\n", - __func__, hp_nct, n_ctx_min); + if (n_ctx_min == UINT32_MAX) { + LLAMA_LOG_INFO("%s: user has requested full context size of %" PRIu32 " -> no change\n", __func__, hp_nct); + } else { + LLAMA_LOG_INFO("%s: default model context size is %" PRIu32 " which is <= the min. context size of %" PRIu32 " -> no change\n", + __func__, hp_nct, n_ctx_min); + } } } else { LLAMA_LOG_INFO("%s: context size set by user to %" PRIu32 " -> no change\n", __func__, cparams->n_ctx); @@ -1091,25 +1095,55 @@ int32_t llama_chat_apply_template( // model split // -int llama_split_path(char * split_path, size_t maxlen, const char * path_prefix, int split_no, int split_count) { +int32_t llama_split_path( + char * split_path, + size_t maxlen, + const char * path_prefix, + int32_t split_no, + int32_t split_count) { + static const char * const SPLIT_PATH_FORMAT = "%s-%05d-of-%05d.gguf"; - if (snprintf(split_path, maxlen, SPLIT_PATH_FORMAT, path_prefix, split_no + 1, split_count)) { - return strlen(split_path); + + const int written = snprintf( + split_path, + maxlen, + SPLIT_PATH_FORMAT, + path_prefix, + split_no + 1, + split_count + ); + + if (written < 0 || (size_t) written >= maxlen) { + return 0; } - return 0; + + return (int32_t) written; } -int llama_split_prefix(char * split_prefix, size_t maxlen, const char * split_path, int split_no, int split_count) { - std::string str_split_path(split_path); - char postfix[32]; - snprintf(postfix, 32, "-%05d-of-%05d.gguf", split_no + 1, split_count); - std::string str_postfix(postfix); +int32_t llama_split_prefix( + char * split_prefix, + size_t maxlen, + const char * split_path, + int32_t split_no, + int32_t split_count) { - // check if split_prefix ends with postfix - int size_prefix = str_split_path.size() - str_postfix.size(); - if (size_prefix > 0 && str_split_path.find(str_postfix, size_prefix) != std::string::npos) { - snprintf(split_prefix, std::min((size_t) size_prefix + 1, maxlen), "%s", split_path); - return size_prefix; + const std::string str_split_path(split_path); + + char postfix[32]; + snprintf(postfix, sizeof(postfix), "-%05d-of-%05d.gguf", split_no + 1, split_count); + + const std::string str_postfix(postfix); + if (str_split_path.size() <= str_postfix.size()) { + return 0; + } + + const size_t size_prefix = str_split_path.size() - str_postfix.size(); + + if (str_split_path.compare(size_prefix, std::string::npos, str_postfix) == 0) { + const size_t copy_len = std::min(size_prefix + 1, maxlen); + snprintf(split_prefix, copy_len, "%s", split_path); + + return (int32_t) size_prefix; } return 0; diff --git a/src/models/deepseek2.cpp b/src/models/deepseek2.cpp index c404c1946d..297dca5136 100644 --- a/src/models/deepseek2.cpp +++ b/src/models/deepseek2.cpp @@ -2,14 +2,11 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - // lite variants include DeepSeek-V2-Lite, GigaChat3-10B-A1.8B - bool is_lite = (hparams.n_layer == 27 || hparams.n_layer == 26); - - const bool is_mla = (hparams.n_embd_head_k_mla != 0 && hparams.n_embd_head_v_mla != 0); + const bool is_mla = hparams.is_mla(); // note: these are the actual head sizes you get when treating as MHA or after "decompression" using wv_b for MLA - const int64_t n_embd_head_k = is_mla ? hparams.n_embd_head_k_mla : hparams.n_embd_head_k; - const int64_t n_embd_head_v = is_mla ? hparams.n_embd_head_v_mla : hparams.n_embd_head_v; + const int64_t n_embd_head_k = hparams.n_embd_head_k_mla(); + const int64_t n_embd_head_v = hparams.n_embd_head_v_mla(); const int64_t n_embd_head_qk_rope = hparams.n_rot; const int64_t n_embd_head_qk_nope = n_embd_head_k - n_embd_head_qk_rope; @@ -43,7 +40,8 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr // inp_pos - contains the positions ggml_tensor * inp_pos = build_inp_pos(); - auto * inp_attn = build_attn_inp_kv(); + auto * inp_attn_kv = !is_mla ? build_attn_inp_kv() : nullptr; + auto * inp_attn_k = is_mla ? build_attn_inp_k() : nullptr; ggml_tensor * inp_out_ids = build_inp_out_ids(); @@ -57,6 +55,9 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr // self_attention { ggml_tensor * q = NULL; + + const bool is_lite = model.layers[il].wq; + if (!is_lite) { q = ggml_mul_mat(ctx0, model.layers[il].wq_a, cur); cb(q, "q", il); @@ -145,7 +146,7 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr } // note: MLA with the absorption optimzation converts into MQA (ie: GQA with 1 group) - cur = build_attn(inp_attn, + cur = build_attn(inp_attn_k, model.layers[il].wo, NULL, Qcur, Kcur, Vcur, nullptr, nullptr, model.layers[il].wv_b, kq_scale, il); } else { @@ -182,7 +183,7 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr } // note: MLA without the absorption optimization converts into MHA (ie: GQA with full n_head groups) - cur = build_attn(inp_attn, + cur = build_attn(inp_attn_kv, model.layers[il].wo, NULL, Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il); } diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp index 146d05f53b..d4c1f525c6 100644 --- a/tests/test-backend-ops.cpp +++ b/tests/test-backend-ops.cpp @@ -8216,8 +8216,8 @@ static std::vector> make_test_cases_eval() { for (int nh : { 4, }) { for (int nr3 : { 1, 3, }) { if (hsk > 64 && nr3 > 1) continue; // skip broadcast for large head sizes - for (int nr2 : { 1, 4, 16 }) { - if (nr2 == 16 && hsk != 128) continue; + for (int nr2 : { 1, 4, 12 }) { + if (nr2 == 12 && hsk != 128) continue; //for (int kv : { 1, 17, 31, 33, 61, 113, 65, 127, 129, 130, 255, 260, 371, 380, 407, 512, 1024, }) { for (int kv : { 113, 512, 1024, }) { if (nr2 != 1 && kv != 512) continue; diff --git a/tests/test-chat-template.cpp b/tests/test-chat-template.cpp index e142900723..d2a1437ca4 100644 --- a/tests/test-chat-template.cpp +++ b/tests/test-chat-template.cpp @@ -481,7 +481,7 @@ int main_automated_tests(void) { /* .name= */ "Mistral-Large-Instruct-2407 (mistralai 'v3' template; modified to have system prompt at start)", /* .template_str= */ "{%- if messages[0][\"role\"] == \"system\" %}\n {%- set system_message = messages[0][\"content\"] %}\n {%- set loop_messages = messages[1:] %}\n{%- else %}\n {%- set loop_messages = messages %}\n{%- endif %}\n{%- if not tools is defined %}\n {%- set tools = none %}\n{%- endif %}\n{%- set user_messages = loop_messages | selectattr(\"role\", \"equalto\", \"user\") | list %}\n\n{#- This block checks for alternating user/assistant messages, skipping tool calling messages #}\n{%- set ns = namespace() %}\n{%- set ns.index = 0 %}\n{%- for message in loop_messages %}\n {%- if not (message.role == \"tool\" or message.role == \"tool_results\" or (message.tool_calls is defined and message.tool_calls is not none)) %}\n {%- if (message[\"role\"] == \"user\") != (ns.index % 2 == 0) %}\n {{- raise_exception(\"After the optional system message, conversation roles must alternate user/assistant/user/assistant/...\") }}\n {%- endif %}\n {%- set ns.index = ns.index + 1 %}\n {%- endif %}\n{%- endfor %}\n\n{{- bos_token }}\n{%- for message in loop_messages %}\n {%- if message[\"role\"] == \"user\" %}\n {%- if tools is not none and (message == user_messages[-1]) %}\n {{- \"[AVAILABLE_TOOLS] [\" }}\n {%- for tool in tools %}\n {%- set tool = tool.function %}\n {{- '{\"type\": \"function\", \"function\": {' }}\n {%- for key, val in tool.items() if key != \"return\" %}\n {%- if val is string %}\n {{- '\"' + key + '\": \"' + val + '\"' }}\n {%- else %}\n {{- '\"' + key + '\": ' + val|tojson }}\n {%- endif %}\n {%- if not loop.last %}\n {{- \", \" }}\n {%- endif %}\n {%- endfor %}\n {{- \"}}\" }}\n {%- if not loop.last %}\n {{- \", \" }}\n {%- else %}\n {{- \"]\" }}\n {%- endif %}\n {%- endfor %}\n {{- \"[/AVAILABLE_TOOLS]\" }}\n {%- endif %}\n {%- if loop.last and system_message is defined %}\n {{- \"[INST] \" + system_message + \"\\n\\n\" + message[\"content\"] + \"[/INST]\" }}\n {%- else %}\n {{- \"[INST] \" + message[\"content\"] + \"[/INST]\" }}\n {%- endif %}\n {%- elif message.tool_calls is defined and message.tool_calls is not none %}\n {{- \"[TOOL_CALLS] [\" }}\n {%- for tool_call in message.tool_calls %}\n {%- set out = tool_call.function|tojson %}\n {{- out[:-1] }}\n {%- if not tool_call.id is defined or tool_call.id|length != 9 %}\n {{- raise_exception(\"Tool call IDs should be alphanumeric strings with length 9!\") }}\n {%- endif %}\n {{- ', \"id\": \"' + tool_call.id + '\"}' }}\n {%- if not loop.last %}\n {{- \", \" }}\n {%- else %}\n {{- \"]\" + eos_token }}\n {%- endif %}\n {%- endfor %}\n {%- elif message[\"role\"] == \"assistant\" %}\n {{- \" \" + message[\"content\"]|trim + eos_token}}\n {%- elif message[\"role\"] == \"tool_results\" or message[\"role\"] == \"tool\" %}\n {%- if message.content is defined and message.content.content is defined %}\n {%- set content = message.content.content %}\n {%- else %}\n {%- set content = message.content %}\n {%- endif %}\n {{- '[TOOL_RESULTS] {\"content\": ' + content|string + \", \" }}\n {%- if not message.tool_call_id is defined or message.tool_call_id|length != 9 %}\n {{- raise_exception(\"Tool call IDs should be alphanumeric strings with length 9!\") }}\n {%- endif %}\n {{- '\"call_id\": \"' + message.tool_call_id + '\"}[/TOOL_RESULTS]' }}\n {%- else %}\n {{- raise_exception(\"Only user and assistant roles are supported, with the exception of an initial optional system message!\") }}\n {%- endif %}\n{%- endfor %}\n", /* .expected_output= */ "[INST] You are a helpful assistant\n\nHello[/INST] Hi there[INST] Who are you[/INST] I am an assistant[INST] Another question[/INST]", - /* .expected_output_jinja= */ "[INST] Hello[/INST] Hi there[INST] Who are you[/INST] I am an assistant[INST] You are a helpful assistant\n\nAnother question[/INST]", + /* .expected_output_jinja= */ "[INST] Hello[/INST] Hi there[INST] Who are you[/INST] I am an assistant[AVAILABLE_TOOLS] [[/AVAILABLE_TOOLS][INST] You are a helpful assistant\n\nAnother question[/INST]", /* .bos_token= */ "", /* .eos_token= */ "", }, @@ -489,7 +489,7 @@ int main_automated_tests(void) { /* .name= */ "Mistral-Nemo-Instruct-2407 (mistralai 'v3-tekken' template; modified to have system prompt at start)", /* .template_str= */ "{%- if messages[0][\"role\"] == \"system\" %}\n {%- set system_message = messages[0][\"content\"] %}\n {%- set loop_messages = messages[1:] %}\n{%- else %}\n {%- set loop_messages = messages %}\n{%- endif %}\n{%- if not tools is defined %}\n {%- set tools = none %}\n{%- endif %}\n{%- set user_messages = loop_messages | selectattr(\"role\", \"equalto\", \"user\") | list %}\n\n{#- This block checks for alternating user/assistant messages, skipping tool calling messages #}\n{%- set ns = namespace() %}\n{%- set ns.index = 0 %}\n{%- for message in loop_messages %}\n {%- if not (message.role == \"tool\" or message.role == \"tool_results\" or (message.tool_calls is defined and message.tool_calls is not none)) %}\n {%- if (message[\"role\"] == \"user\") != (ns.index % 2 == 0) %}\n {{- raise_exception(\"After the optional system message, conversation roles must alternate user/assistant/user/assistant/...\") }}\n {%- endif %}\n {%- set ns.index = ns.index + 1 %}\n {%- endif %}\n{%- endfor %}\n\n{{- bos_token }}\n{%- for message in loop_messages %}\n {%- if message[\"role\"] == \"user\" %}\n {%- if tools is not none and (message == user_messages[-1]) %}\n {{- \"[AVAILABLE_TOOLS][\" }}\n {%- for tool in tools %}\n {%- set tool = tool.function %}\n {{- '{\"type\": \"function\", \"function\": {' }}\n {%- for key, val in tool.items() if key != \"return\" %}\n {%- if val is string %}\n {{- '\"' + key + '\": \"' + val + '\"' }}\n {%- else %}\n {{- '\"' + key + '\": ' + val|tojson }}\n {%- endif %}\n {%- if not loop.last %}\n {{- \", \" }}\n {%- endif %}\n {%- endfor %}\n {{- \"}}\" }}\n {%- if not loop.last %}\n {{- \", \" }}\n {%- else %}\n {{- \"]\" }}\n {%- endif %}\n {%- endfor %}\n {{- \"[/AVAILABLE_TOOLS]\" }}\n {%- endif %}\n {%- if loop.last and system_message is defined %}\n {{- \"[INST]\" + system_message + \"\\n\\n\" + message[\"content\"] + \"[/INST]\" }}\n {%- else %}\n {{- \"[INST]\" + message[\"content\"] + \"[/INST]\" }}\n {%- endif %}\n {%- elif (message.tool_calls is defined and message.tool_calls is not none) %}\n {{- \"[TOOL_CALLS][\" }}\n {%- for tool_call in message.tool_calls %}\n {%- set out = tool_call.function|tojson %}\n {{- out[:-1] }}\n {%- if not tool_call.id is defined or tool_call.id|length != 9 %}\n {{- raise_exception(\"Tool call IDs should be alphanumeric strings with length 9!\") }}\n {%- endif %}\n {{- ', \"id\": \"' + tool_call.id + '\"}' }}\n {%- if not loop.last %}\n {{- \", \" }}\n {%- else %}\n {{- \"]\" + eos_token }}\n {%- endif %}\n {%- endfor %}\n {%- elif message[\"role\"] == \"assistant\" %}\n {{- message[\"content\"] + eos_token}}\n {%- elif message[\"role\"] == \"tool_results\" or message[\"role\"] == \"tool\" %}\n {%- if message.content is defined and message.content.content is defined %}\n {%- set content = message.content.content %}\n {%- else %}\n {%- set content = message.content %}\n {%- endif %}\n {{- '[TOOL_RESULTS]{\"content\": ' + content|string + \", \" }}\n {%- if not message.tool_call_id is defined or message.tool_call_id|length != 9 %}\n {{- raise_exception(\"Tool call IDs should be alphanumeric strings with length 9!\") }}\n {%- endif %}\n {{- '\"call_id\": \"' + message.tool_call_id + '\"}[/TOOL_RESULTS]' }}\n {%- else %}\n {{- raise_exception(\"Only user and assistant roles are supported, with the exception of an initial optional system message!\") }}\n {%- endif %}\n{%- endfor %}\n", /* .expected_output= */ "[INST]You are a helpful assistant\n\nHello[/INST]Hi there[INST]Who are you[/INST] I am an assistant [INST]Another question[/INST]", - /* .expected_output_jinja= */ "[INST]Hello[/INST]Hi there[INST]Who are you[/INST] I am an assistant [INST]You are a helpful assistant\n\nAnother question[/INST]", + /* .expected_output_jinja= */ "[INST]Hello[/INST]Hi there[INST]Who are you[/INST] I am an assistant [AVAILABLE_TOOLS][[/AVAILABLE_TOOLS][INST]You are a helpful assistant\n\nAnother question[/INST]", /* .bos_token= */ "", /* .eos_token= */ "", }, diff --git a/tests/test-gguf.cpp b/tests/test-gguf.cpp index 3f0c312e2f..84b7f3bc49 100644 --- a/tests/test-gguf.cpp +++ b/tests/test-gguf.cpp @@ -1,9 +1,11 @@ #include "ggml.h" #include "ggml-backend.h" #include "../ggml/src/ggml-impl.h" +#include "gguf.h" #include #include +#include #include #include #include @@ -34,6 +36,7 @@ enum handcrafted_file_type { HANDCRAFTED_TENSORS_BAD_N_DIMS = 20 + offset_has_tensors, HANDCRAFTED_TENSORS_BAD_SHAPE = 30 + offset_has_tensors, HANDCRAFTED_TENSORS_NE_TOO_BIG = 40 + offset_has_tensors, + HANDCRAFTED_TENSORS_NBYTES_TOO_BIG = 45 + offset_has_tensors, HANDCRAFTED_TENSORS_BAD_TYPE = 50 + offset_has_tensors, HANDCRAFTED_TENSORS_BAD_OFFSET = 60 + offset_has_tensors, HANDCRAFTED_TENSORS_DUPLICATE_NAME = 70 + offset_has_tensors, @@ -69,6 +72,7 @@ static std::string handcrafted_file_type_name(const enum handcrafted_file_type h case HANDCRAFTED_TENSORS_BAD_N_DIMS: return "TENSORS_BAD_N_DIMS"; case HANDCRAFTED_TENSORS_BAD_SHAPE: return "TENSORS_BAD_SHAPE"; case HANDCRAFTED_TENSORS_NE_TOO_BIG: return "TENSORS_NE_TOO_BIG"; + case HANDCRAFTED_TENSORS_NBYTES_TOO_BIG: return "TENSORS_NBYTES_TOO_BIG"; case HANDCRAFTED_TENSORS_BAD_TYPE: return "TENSORS_BAD_TYPE"; case HANDCRAFTED_TENSORS_BAD_OFFSET: return "TENSORS_BAD_OFFSET"; case HANDCRAFTED_TENSORS_DUPLICATE_NAME: return "TENSORS_DUPLICATE_NAME"; @@ -326,7 +330,7 @@ static FILE * get_handcrafted_file(const unsigned int seed, const enum handcraft uint64_t offset = 0; for (int i = 0; i < int(tensor_configs.size()); ++i) { - const ggml_type type = tensor_configs[i].first; + const ggml_type type = hft == HANDCRAFTED_TENSORS_NBYTES_TOO_BIG ? GGML_TYPE_I64 : tensor_configs[i].first; const std::array shape = tensor_configs[i].second; std::string name = "my_tensor"; @@ -343,7 +347,7 @@ static FILE * get_handcrafted_file(const unsigned int seed, const enum handcraft } helper_write(file, name.data(), name.length()); - uint32_t n_dims = hft == HANDCRAFTED_TENSORS_NE_TOO_BIG ? 2 : 1; + uint32_t n_dims = (hft == HANDCRAFTED_TENSORS_NE_TOO_BIG || hft == HANDCRAFTED_TENSORS_NBYTES_TOO_BIG) ? 2 : 1; for (int i = GGML_MAX_DIMS-1; i >= 1; --i) { if (shape[i] != 1) { n_dims = i + 1; @@ -358,13 +362,19 @@ static FILE * get_handcrafted_file(const unsigned int seed, const enum handcraft } if (hft == HANDCRAFTED_TENSORS_BAD_SHAPE) { + const int64_t bad_dim = -1; for (uint32_t j = 0; j < n_dims; ++j) { - const int64_t bad_dim = -1; helper_write(file, bad_dim); } } else if (hft == HANDCRAFTED_TENSORS_NE_TOO_BIG){ + const int64_t big_dim = 4*int64_t(INT32_MAX); + for (uint32_t j = 0; j < n_dims; ++j) { + helper_write(file, big_dim); + } + } else if (hft == HANDCRAFTED_TENSORS_NBYTES_TOO_BIG){ + const size_t big_ne = SIZE_MAX/ggml_type_size(type); + const int64_t big_dim = GGML_PAD(int64_t(1.01f*std::pow(big_ne, 1.0f/n_dims)) + 1, ggml_blck_size(type)); for (uint32_t j = 0; j < n_dims; ++j) { - const int64_t big_dim = 4*int64_t(INT32_MAX); helper_write(file, big_dim); } } else { @@ -682,6 +692,7 @@ static std::pair test_handcrafted_file(const unsigned int seed) { HANDCRAFTED_TENSORS_BAD_N_DIMS, HANDCRAFTED_TENSORS_BAD_SHAPE, HANDCRAFTED_TENSORS_NE_TOO_BIG, + HANDCRAFTED_TENSORS_NBYTES_TOO_BIG, HANDCRAFTED_TENSORS_BAD_TYPE, HANDCRAFTED_TENSORS_BAD_OFFSET, HANDCRAFTED_TENSORS_DUPLICATE_NAME, diff --git a/tests/test-jinja.cpp b/tests/test-jinja.cpp index 54d3a0923b..f6114f1e2f 100644 --- a/tests/test-jinja.cpp +++ b/tests/test-jinja.cpp @@ -9,6 +9,7 @@ #include "jinja/runtime.h" #include "jinja/parser.h" #include "jinja/lexer.h" +#include "jinja/utils.h" #include "testing.h" @@ -30,6 +31,7 @@ static void test_tests(testing & t); static void test_string_methods(testing & t); static void test_array_methods(testing & t); static void test_object_methods(testing & t); +static void test_hasher(testing & t); static void test_fuzzing(testing & t); static bool g_python_mode = false; @@ -67,6 +69,7 @@ int main(int argc, char *argv[]) { t.test("array methods", test_array_methods); t.test("object methods", test_object_methods); if (!g_python_mode) { + t.test("hasher", test_hasher); t.test("fuzzing", test_fuzzing); } @@ -156,6 +159,18 @@ static void test_conditionals(testing & t) { "big" ); + test_template(t, "object comparison", + "{% if {0: 1, none: 2, 1.0: 3, '0': 4, true: 5} == {false: 1, none: 2, 1: 5, '0': 4} %}equal{% endif %}", + json::object(), + "equal" + ); + + test_template(t, "array comparison", + "{% if [0, 1.0, false] == [false, 1, 0.0] %}equal{% endif %}", + json::object(), + "equal" + ); + test_template(t, "logical and", "{% if a and b %}both{% endif %}", {{"a", true}, {"b", true}}, @@ -314,6 +329,12 @@ static void test_loops(testing & t) { "empty" ); + test_template(t, "for undefined empty", + "{% for i in items %}{{ i }}{% else %}empty{% endfor %}", + json::object(), + "empty" + ); + test_template(t, "nested for", "{% for i in a %}{% for j in b %}{{ i }}{{ j }}{% endfor %}{% endfor %}", {{"a", json::array({1, 2})}, {"b", json::array({"x", "y"})}}, @@ -358,6 +379,30 @@ static void test_expressions(testing & t) { "b" ); + test_template(t, "array negative access", + "{{ items[-1] }}", + {{"items", json::array({"a", "b", "c"})}}, + "c" + ); + + test_template(t, "array slice", + "{{ items[1:-1]|string }}", + {{"items", json::array({"a", "b", "c"})}}, + "['b']" + ); + + test_template(t, "array slice step", + "{{ items[::2]|string }}", + {{"items", json::array({"a", "b", "c"})}}, + "['a', 'c']" + ); + + test_template(t, "tuple slice", + "{{ ('a', 'b', 'c')[::-1]|string }}", + json::object(), + "('c', 'b', 'a')" + ); + test_template(t, "arithmetic", "{{ (a + b) * c }}", {{"a", 2}, {"b", 3}, {"c", 4}}, @@ -401,6 +446,36 @@ static void test_set_statement(testing & t) { json::object(), "1" ); + + test_template(t, "set dict with mixed type keys", + "{% set d = {0: 1, none: 2, 1.0: 3, '0': 4, (0, 0): 5, false: 6, 1: 7} %}{{ d[(0, 0)] + d[0] + d[none] + d['0'] + d[false] + d[1.0] + d[1] }}", + json::object(), + "37" + ); + + test_template(t, "print dict with mixed type keys", + "{% set d = {0: 1, none: 2, 1.0: 3, '0': 4, (0, 0): 5, true: 6} %}{{ d|string }}", + json::object(), + "{0: 1, None: 2, 1.0: 6, '0': 4, (0, 0): 5}" + ); + + test_template(t, "print array with mixed types", + "{% set d = [0, none, 1.0, '0', true, (0, 0)] %}{{ d|string }}", + json::object(), + "[0, None, 1.0, '0', True, (0, 0)]" + ); + + test_template(t, "object member assignment with mixed key types", + "{% set d = namespace() %}{% set d.a = 123 %}{{ d['a'] == 123 }}", + json::object(), + "True" + ); + + test_template(t, "tuple unpacking", + "{% set t = (1, 2, 3) %}{% set a, b, c = t %}{{ a + b + c }}", + json::object(), + "6" + ); } static void test_filters(testing & t) { @@ -949,6 +1024,18 @@ static void test_tests(testing & t) { {{"x", {{"a", 1}}}}, "yes" ); + + test_template(t, "undefined is sequence", + "{{ 'yes' if x is sequence }}", + json::object(), + "yes" + ); + + test_template(t, "undefined is iterable", + "{{ 'yes' if x is iterable }}", + json::object(), + "yes" + ); } static void test_string_methods(testing & t) { @@ -1053,6 +1140,54 @@ static void test_string_methods(testing & t) { {{"s", "banana"}}, "bXnXna" ); + + test_template(t, "undefined|capitalize", + "{{ arr|capitalize }}", + json::object(), + "" + ); + + test_template(t, "undefined|title", + "{{ arr|title }}", + json::object(), + "" + ); + + test_template(t, "undefined|truncate", + "{{ arr|truncate(9) }}", + json::object(), + "" + ); + + test_template(t, "undefined|upper", + "{{ arr|upper }}", + json::object(), + "" + ); + + test_template(t, "undefined|lower", + "{{ arr|lower }}", + json::object(), + "" + ); + + test_template(t, "undefined|replace", + "{{ arr|replace('a', 'b') }}", + json::object(), + "" + ); + + test_template(t, "undefined|trim", + "{{ arr|trim }}", + json::object(), + "" + ); + + test_template(t, "undefined|wordcount", + "{{ arr|wordcount }}", + json::object(), + "0" + ); } static void test_array_methods(testing & t) { @@ -1220,6 +1355,108 @@ static void test_array_methods(testing & t) { // {{"arr", json::array({"a", "b", "c"})}}, // "a,x,b,c" // ); + + test_template(t, "undefined|select", + "{% for item in items|select('odd') %}{{ item.name }} {% endfor %}", + json::object(), + "" + ); + + test_template(t, "undefined|selectattr", + "{% for item in items|selectattr('active') %}{{ item.name }} {% endfor %}", + json::object(), + "" + ); + + test_template(t, "undefined|reject", + "{% for item in items|reject('even') %}{{ item.name }} {% endfor %}", + json::object(), + "" + ); + + test_template(t, "undefined|rejectattr", + "{% for item in items|rejectattr('active') %}{{ item.name }} {% endfor %}", + json::object(), + "" + ); + + test_template(t, "undefined|list", + "{{ arr|list|string }}", + json::object(), + "[]" + ); + + test_template(t, "undefined|string", + "{{ arr|string }}", + json::object(), + "" + ); + + test_template(t, "undefined|first", + "{{ arr|first }}", + json::object(), + "" + ); + + test_template(t, "undefined|last", + "{{ arr|last }}", + json::object(), + "" + ); + + test_template(t, "undefined|length", + "{{ arr|length }}", + json::object(), + "0" + ); + + test_template(t, "undefined|join", + "{{ arr|join }}", + json::object(), + "" + ); + + test_template(t, "undefined|sort", + "{{ arr|sort|string }}", + json::object(), + "[]" + ); + + test_template(t, "undefined|reverse", + "{{ arr|reverse|join }}", + json::object(), + "" + ); + + test_template(t, "undefined|map", + "{% for v in arr|map(attribute='age') %}{{ v }} {% endfor %}", + json::object(), + "" + ); + + test_template(t, "undefined|min", + "{{ arr|min }}", + json::object(), + "" + ); + + test_template(t, "undefined|max", + "{{ arr|max }}", + json::object(), + "" + ); + + test_template(t, "undefined|unique", + "{{ arr|unique|join }}", + json::object(), + "" + ); + + test_template(t, "undefined|sum", + "{{ arr|sum }}", + json::object(), + "0" + ); } static void test_object_methods(testing & t) { @@ -1312,6 +1549,160 @@ static void test_object_methods(testing & t) { {{"obj", {{"a", "b"}}}}, "True True" ); + + test_template(t, "expression as object key", + "{% set d = {'ab': 123} %}{{ d['a' + 'b'] == 123 }}", + json::object(), + "True" + ); + + test_template(t, "numeric as object key (template: Seed-OSS)", + "{% set d = {1: 'a', 2: 'b'} %}{{ d[1] == 'a' and d[2] == 'b' }}", + json::object(), + "True" + ); + + test_template(t, "undefined|items", + "{{ arr|items|join }}", + json::object(), + "" + ); +} + +static void test_hasher(testing & t) { + static const std::vector> chunk_sizes = { + {1, 2}, + {1, 16}, + {8, 1}, + {1, 1024}, + {5, 512}, + {16, 256}, + {45, 122}, + {70, 634}, + }; + + static auto random_bytes = [](size_t length) -> std::string { + std::string data; + data.resize(length); + for (size_t i = 0; i < length; ++i) { + data[i] = static_cast(rand() % 256); + } + return data; + }; + + t.test("state unchanged with empty input", [](testing & t) { + jinja::hasher hasher; + hasher.update("some data"); + size_t initial_state = hasher.digest(); + hasher.update("", 0); + size_t final_state = hasher.digest(); + t.assert_true("Hasher state should remain unchanged", initial_state == final_state); + }); + + t.test("different inputs produce different hashes", [](testing & t) { + jinja::hasher hasher1; + hasher1.update("data one"); + size_t hash1 = hasher1.digest(); + + jinja::hasher hasher2; + hasher2.update("data two"); + size_t hash2 = hasher2.digest(); + + t.assert_true("Different inputs should produce different hashes", hash1 != hash2); + }); + + t.test("same inputs produce same hashes", [](testing & t) { + jinja::hasher hasher1; + hasher1.update("consistent data"); + size_t hash1 = hasher1.digest(); + + jinja::hasher hasher2; + hasher2.update("consistent data"); + size_t hash2 = hasher2.digest(); + + t.assert_true("Same inputs should produce same hashes", hash1 == hash2); + }); + + t.test("property: update(a ~ b) == update(a).update(b)", [](testing & t) { + for (const auto & [size1, size2] : chunk_sizes) { + std::string data1 = random_bytes(size1); + std::string data2 = random_bytes(size2); + + jinja::hasher hasher1; + hasher1.update(data1); + hasher1.update(data2); + size_t hash1 = hasher1.digest(); + + jinja::hasher hasher2; + hasher2.update(data1 + data2); + size_t hash2 = hasher2.digest(); + + t.assert_true( + "Hashing in multiple updates should match single update (" + std::to_string(size1) + ", " + std::to_string(size2) + ")", + hash1 == hash2); + } + }); + + t.test("property: update(a ~ b) == update(a).update(b) with more update passes", [](testing & t) { + static const std::vector sizes = {3, 732, 131, 13, 17, 256, 436, 99, 4}; + + jinja::hasher hasher1; + jinja::hasher hasher2; + + std::string combined_data; + for (size_t size : sizes) { + std::string data = random_bytes(size); + hasher1.update(data); + combined_data += data; + } + + hasher2.update(combined_data); + size_t hash1 = hasher1.digest(); + size_t hash2 = hasher2.digest(); + t.assert_true( + "Hashing in multiple updates should match single update with many chunks", + hash1 == hash2); + }); + + t.test("property: non associativity of update", [](testing & t) { + for (const auto & [size1, size2] : chunk_sizes) { + std::string data1 = random_bytes(size1); + std::string data2 = random_bytes(size2); + + jinja::hasher hasher1; + hasher1.update(data1); + hasher1.update(data2); + size_t hash1 = hasher1.digest(); + + jinja::hasher hasher2; + hasher2.update(data2); + hasher2.update(data1); + size_t hash2 = hasher2.digest(); + + t.assert_true( + "Hashing order should matter (" + std::to_string(size1) + ", " + std::to_string(size2) + ")", + hash1 != hash2); + } + }); + + t.test("property: different lengths produce different hashes (padding block size)", [](testing & t) { + std::string random_data = random_bytes(64); + + jinja::hasher hasher1; + hasher1.update(random_data); + size_t hash1 = hasher1.digest(); + + for (int i = 0; i < 16; ++i) { + random_data.push_back('A'); // change length + jinja::hasher hasher2; + hasher2.update(random_data); + size_t hash2 = hasher2.digest(); + + t.assert_true("Different lengths should produce different hashes (length " + std::to_string(random_data.size()) + ")", hash1 != hash2); + + hash1 = hash2; + } + }); } static void test_template_cpp(testing & t, const std::string & name, const std::string & tmpl, const json & vars, const std::string & expect) { diff --git a/tools/cli/README.md b/tools/cli/README.md index 3b6f0708ed..4a15cbad9d 100644 --- a/tools/cli/README.md +++ b/tools/cli/README.md @@ -45,10 +45,10 @@ | `--rope-freq-base N` | RoPE base frequency, used by NTK-aware scaling (default: loaded from model)
(env: LLAMA_ARG_ROPE_FREQ_BASE) | | `--rope-freq-scale N` | RoPE frequency scaling factor, expands context by a factor of 1/N
(env: LLAMA_ARG_ROPE_FREQ_SCALE) | | `--yarn-orig-ctx N` | YaRN: original context size of model (default: 0 = model training context size)
(env: LLAMA_ARG_YARN_ORIG_CTX) | -| `--yarn-ext-factor N` | YaRN: extrapolation mix factor (default: -1.0, 0.0 = full interpolation)
(env: LLAMA_ARG_YARN_EXT_FACTOR) | -| `--yarn-attn-factor N` | YaRN: scale sqrt(t) or attention magnitude (default: -1.0)
(env: LLAMA_ARG_YARN_ATTN_FACTOR) | -| `--yarn-beta-slow N` | YaRN: high correction dim or alpha (default: -1.0)
(env: LLAMA_ARG_YARN_BETA_SLOW) | -| `--yarn-beta-fast N` | YaRN: low correction dim or beta (default: -1.0)
(env: LLAMA_ARG_YARN_BETA_FAST) | +| `--yarn-ext-factor N` | YaRN: extrapolation mix factor (default: -1.00, 0.0 = full interpolation)
(env: LLAMA_ARG_YARN_EXT_FACTOR) | +| `--yarn-attn-factor N` | YaRN: scale sqrt(t) or attention magnitude (default: -1.00)
(env: LLAMA_ARG_YARN_ATTN_FACTOR) | +| `--yarn-beta-slow N` | YaRN: high correction dim or alpha (default: -1.00)
(env: LLAMA_ARG_YARN_BETA_SLOW) | +| `--yarn-beta-fast N` | YaRN: low correction dim or beta (default: -1.00)
(env: LLAMA_ARG_YARN_BETA_FAST) | | `-kvo, --kv-offload, -nkvo, --no-kv-offload` | whether to enable KV cache offloading (default: enabled)
(env: LLAMA_ARG_KV_OFFLOAD) | | `--repack, -nr, --no-repack` | whether to enable weight repacking (default: enabled)
(env: LLAMA_ARG_REPACK) | | `--no-host` | bypass host buffer allowing extra buffers to be used
(env: LLAMA_ARG_NO_HOST) | @@ -109,30 +109,30 @@ | `-s, --seed SEED` | RNG seed (default: -1, use random seed for -1) | | `--sampler-seq, --sampling-seq SEQUENCE` | simplified sequence for samplers that will be used (default: edskypmxt) | | `--ignore-eos` | ignore end of stream token and continue generating (implies --logit-bias EOS-inf) | -| `--temp N` | temperature (default: 0.8) | +| `--temp N` | temperature (default: 0.80) | | `--top-k N` | top-k sampling (default: 40, 0 = disabled)
(env: LLAMA_ARG_TOP_K) | -| `--top-p N` | top-p sampling (default: 0.9, 1.0 = disabled) | -| `--min-p N` | min-p sampling (default: 0.1, 0.0 = disabled) | -| `--adaptive-target N` | adaptive-p: select tokens near this probability (valid range 0.0 to 1.0; negative = disabled) | -| `--adaptive-decay N` | adaptive-p: EMA decay for adaptation; effective history length ≈ 1/(1-decay) tokens (valid range 0.0 - 0.99) | -| `--top-nsigma N` | top-n-sigma sampling (default: -1.0, -1.0 = disabled) | -| `--xtc-probability N` | xtc probability (default: 0.0, 0.0 = disabled) | -| `--xtc-threshold N` | xtc threshold (default: 0.1, 1.0 = disabled) | -| `--typical N` | locally typical sampling, parameter p (default: 1.0, 1.0 = disabled) | +| `--top-p N` | top-p sampling (default: 0.95, 1.0 = disabled) | +| `--min-p N` | min-p sampling (default: 0.05, 0.0 = disabled) | +| `--top-nsigma N` | top-n-sigma sampling (default: -1.00, -1.0 = disabled) | +| `--xtc-probability N` | xtc probability (default: 0.00, 0.0 = disabled) | +| `--xtc-threshold N` | xtc threshold (default: 0.10, 1.0 = disabled) | +| `--typical N` | locally typical sampling, parameter p (default: 1.00, 1.0 = disabled) | | `--repeat-last-n N` | last n tokens to consider for penalize (default: 64, 0 = disabled, -1 = ctx_size) | -| `--repeat-penalty N` | penalize repeat sequence of tokens (default: 1.0, 1.0 = disabled) | -| `--presence-penalty N` | repeat alpha presence penalty (default: 0.0, 0.0 = disabled) | -| `--frequency-penalty N` | repeat alpha frequency penalty (default: 0.0, 0.0 = disabled) | -| `--dry-multiplier N` | set DRY sampling multiplier (default: 0.0, 0.0 = disabled) | +| `--repeat-penalty N` | penalize repeat sequence of tokens (default: 1.00, 1.0 = disabled) | +| `--presence-penalty N` | repeat alpha presence penalty (default: 0.00, 0.0 = disabled) | +| `--frequency-penalty N` | repeat alpha frequency penalty (default: 0.00, 0.0 = disabled) | +| `--dry-multiplier N` | set DRY sampling multiplier (default: 0.00, 0.0 = disabled) | | `--dry-base N` | set DRY sampling base value (default: 1.75) | | `--dry-allowed-length N` | set allowed length for DRY sampling (default: 2) | | `--dry-penalty-last-n N` | set DRY penalty for the last n tokens (default: -1, 0 = disable, -1 = context size) | | `--dry-sequence-breaker STRING` | add sequence breaker for DRY sampling, clearing out default breakers ('\n', ':', '"', '*') in the process; use "none" to not use any sequence breakers | -| `--dynatemp-range N` | dynamic temperature range (default: 0.0, 0.0 = disabled) | -| `--dynatemp-exp N` | dynamic temperature exponent (default: 1.0) | +| `--adaptive-target N` | adaptive-p: select tokens near this probability (valid range 0.0 to 1.0; negative = disabled) (default: -1.00)
[(more info)](https://github.com/ggml-org/llama.cpp/pull/17927) | +| `--adaptive-decay N` | adaptive-p: decay rate for target adaptation over time. lower values are more reactive, higher values are more stable.
(valid range 0.0 to 0.99) (default: 0.90) | +| `--dynatemp-range N` | dynamic temperature range (default: 0.00, 0.0 = disabled) | +| `--dynatemp-exp N` | dynamic temperature exponent (default: 1.00) | | `--mirostat N` | use Mirostat sampling.
Top K, Nucleus and Locally Typical samplers are ignored if used.
(default: 0, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0) | -| `--mirostat-lr N` | Mirostat learning rate, parameter eta (default: 0.1) | -| `--mirostat-ent N` | Mirostat target entropy, parameter tau (default: 5.0) | +| `--mirostat-lr N` | Mirostat learning rate, parameter eta (default: 0.10) | +| `--mirostat-ent N` | Mirostat target entropy, parameter tau (default: 5.00) | | `-l, --logit-bias TOKEN_ID(+/-)BIAS` | modifies the likelihood of token appearing in the completion,
i.e. `--logit-bias 15043+1` to increase likelihood of token ' Hello',
or `--logit-bias 15043-1` to decrease likelihood of token ' Hello' | | `--grammar GRAMMAR` | BNF-like grammar to constrain generations (see samples in grammars/ dir) (default: '') | | `--grammar-file FNAME` | file to read grammar from | @@ -173,12 +173,12 @@ | `--jinja, --no-jinja` | whether to use jinja template engine for chat (default: enabled)
(env: LLAMA_ARG_JINJA) | | `--reasoning-format FORMAT` | controls whether thought tags are allowed and/or extracted from the response, and in which format they're returned; one of:
- none: leaves thoughts unparsed in `message.content`
- deepseek: puts thoughts in `message.reasoning_content`
- deepseek-legacy: keeps `` tags in `message.content` while also populating `message.reasoning_content`
(default: auto)
(env: LLAMA_ARG_THINK) | | `--reasoning-budget N` | controls the amount of thinking allowed; currently only one of: -1 for unrestricted thinking budget, or 0 to disable thinking (default: -1)
(env: LLAMA_ARG_THINK_BUDGET) | -| `--chat-template JINJA_TEMPLATE` | set custom jinja chat template (default: template taken from model's metadata)
if suffix/prefix are specified, template will be disabled
only commonly used templates are accepted (unless --jinja is set before this flag):
list of built-in templates:
bailing, bailing-think, bailing2, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, grok-2, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, pangu-embedded, phi3, phi4, rwkv-world, seed_oss, smolvlm, solar-open, vicuna, vicuna-orca, yandex, zephyr
(env: LLAMA_ARG_CHAT_TEMPLATE) | -| `--chat-template-file JINJA_TEMPLATE_FILE` | set custom jinja chat template file (default: template taken from model's metadata)
if suffix/prefix are specified, template will be disabled
only commonly used templates are accepted (unless --jinja is set before this flag):
list of built-in templates:
bailing, bailing-think, bailing2, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, grok-2, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, pangu-embedded, phi3, phi4, rwkv-world, seed_oss, smolvlm, solar-open, vicuna, vicuna-orca, yandex, zephyr
(env: LLAMA_ARG_CHAT_TEMPLATE_FILE) | +| `--chat-template JINJA_TEMPLATE` | set custom jinja chat template (default: template taken from model's metadata)
if suffix/prefix are specified, template will be disabled
only commonly used templates are accepted (unless --jinja is set before this flag):
list of built-in templates:
bailing, bailing-think, bailing2, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone-moe, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, grok-2, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, pangu-embedded, phi3, phi4, rwkv-world, seed_oss, smolvlm, solar-open, vicuna, vicuna-orca, yandex, zephyr
(env: LLAMA_ARG_CHAT_TEMPLATE) | +| `--chat-template-file JINJA_TEMPLATE_FILE` | set custom jinja chat template file (default: template taken from model's metadata)
if suffix/prefix are specified, template will be disabled
only commonly used templates are accepted (unless --jinja is set before this flag):
list of built-in templates:
bailing, bailing-think, bailing2, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone-moe, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, grok-2, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, pangu-embedded, phi3, phi4, rwkv-world, seed_oss, smolvlm, solar-open, vicuna, vicuna-orca, yandex, zephyr
(env: LLAMA_ARG_CHAT_TEMPLATE_FILE) | | `--simple-io` | use basic IO for better compatibility in subprocesses and limited consoles | | `--draft, --draft-n, --draft-max N` | number of tokens to draft for speculative decoding (default: 16)
(env: LLAMA_ARG_DRAFT_MAX) | | `--draft-min, --draft-n-min N` | minimum number of draft tokens to use for speculative decoding (default: 0)
(env: LLAMA_ARG_DRAFT_MIN) | -| `--draft-p-min P` | minimum speculative decoding probability (greedy) (default: 0.8)
(env: LLAMA_ARG_DRAFT_P_MIN) | +| `--draft-p-min P` | minimum speculative decoding probability (greedy) (default: 0.75)
(env: LLAMA_ARG_DRAFT_P_MIN) | | `-cd, --ctx-size-draft N` | size of the prompt context for the draft model (default: 0, 0 = loaded from model)
(env: LLAMA_ARG_CTX_SIZE_DRAFT) | | `-devd, --device-draft ` | comma-separated list of devices to use for offloading the draft model (none = don't offload)
use --list-devices to see a list of available devices | | `-ngld, --gpu-layers-draft, --n-gpu-layers-draft N` | max. number of draft model layers to store in VRAM, either an exact number, 'auto', or 'all' (default: auto)
(env: LLAMA_ARG_N_GPU_LAYERS_DRAFT) | diff --git a/tools/completion/README.md b/tools/completion/README.md index a16be3f684..3ca3e68454 100644 --- a/tools/completion/README.md +++ b/tools/completion/README.md @@ -128,10 +128,10 @@ llama-completion.exe -m models\gemma-1.1-7b-it.Q4_K_M.gguf --ignore-eos -n -1 | `--rope-freq-base N` | RoPE base frequency, used by NTK-aware scaling (default: loaded from model)
(env: LLAMA_ARG_ROPE_FREQ_BASE) | | `--rope-freq-scale N` | RoPE frequency scaling factor, expands context by a factor of 1/N
(env: LLAMA_ARG_ROPE_FREQ_SCALE) | | `--yarn-orig-ctx N` | YaRN: original context size of model (default: 0 = model training context size)
(env: LLAMA_ARG_YARN_ORIG_CTX) | -| `--yarn-ext-factor N` | YaRN: extrapolation mix factor (default: -1.0, 0.0 = full interpolation)
(env: LLAMA_ARG_YARN_EXT_FACTOR) | -| `--yarn-attn-factor N` | YaRN: scale sqrt(t) or attention magnitude (default: -1.0)
(env: LLAMA_ARG_YARN_ATTN_FACTOR) | -| `--yarn-beta-slow N` | YaRN: high correction dim or alpha (default: -1.0)
(env: LLAMA_ARG_YARN_BETA_SLOW) | -| `--yarn-beta-fast N` | YaRN: low correction dim or beta (default: -1.0)
(env: LLAMA_ARG_YARN_BETA_FAST) | +| `--yarn-ext-factor N` | YaRN: extrapolation mix factor (default: -1.00, 0.0 = full interpolation)
(env: LLAMA_ARG_YARN_EXT_FACTOR) | +| `--yarn-attn-factor N` | YaRN: scale sqrt(t) or attention magnitude (default: -1.00)
(env: LLAMA_ARG_YARN_ATTN_FACTOR) | +| `--yarn-beta-slow N` | YaRN: high correction dim or alpha (default: -1.00)
(env: LLAMA_ARG_YARN_BETA_SLOW) | +| `--yarn-beta-fast N` | YaRN: low correction dim or beta (default: -1.00)
(env: LLAMA_ARG_YARN_BETA_FAST) | | `-kvo, --kv-offload, -nkvo, --no-kv-offload` | whether to enable KV cache offloading (default: enabled)
(env: LLAMA_ARG_KV_OFFLOAD) | | `--repack, -nr, --no-repack` | whether to enable weight repacking (default: enabled)
(env: LLAMA_ARG_REPACK) | | `--no-host` | bypass host buffer allowing extra buffers to be used
(env: LLAMA_ARG_NO_HOST) | @@ -192,28 +192,30 @@ llama-completion.exe -m models\gemma-1.1-7b-it.Q4_K_M.gguf --ignore-eos -n -1 | `-s, --seed SEED` | RNG seed (default: -1, use random seed for -1) | | `--sampler-seq, --sampling-seq SEQUENCE` | simplified sequence for samplers that will be used (default: edskypmxt) | | `--ignore-eos` | ignore end of stream token and continue generating (implies --logit-bias EOS-inf) | -| `--temp N` | temperature (default: 0.8) | +| `--temp N` | temperature (default: 0.80) | | `--top-k N` | top-k sampling (default: 40, 0 = disabled)
(env: LLAMA_ARG_TOP_K) | -| `--top-p N` | top-p sampling (default: 0.9, 1.0 = disabled) | -| `--min-p N` | min-p sampling (default: 0.1, 0.0 = disabled) | -| `--top-nsigma N` | top-n-sigma sampling (default: -1.0, -1.0 = disabled) | -| `--xtc-probability N` | xtc probability (default: 0.0, 0.0 = disabled) | -| `--xtc-threshold N` | xtc threshold (default: 0.1, 1.0 = disabled) | -| `--typical N` | locally typical sampling, parameter p (default: 1.0, 1.0 = disabled) | +| `--top-p N` | top-p sampling (default: 0.95, 1.0 = disabled) | +| `--min-p N` | min-p sampling (default: 0.05, 0.0 = disabled) | +| `--top-nsigma N` | top-n-sigma sampling (default: -1.00, -1.0 = disabled) | +| `--xtc-probability N` | xtc probability (default: 0.00, 0.0 = disabled) | +| `--xtc-threshold N` | xtc threshold (default: 0.10, 1.0 = disabled) | +| `--typical N` | locally typical sampling, parameter p (default: 1.00, 1.0 = disabled) | | `--repeat-last-n N` | last n tokens to consider for penalize (default: 64, 0 = disabled, -1 = ctx_size) | -| `--repeat-penalty N` | penalize repeat sequence of tokens (default: 1.0, 1.0 = disabled) | -| `--presence-penalty N` | repeat alpha presence penalty (default: 0.0, 0.0 = disabled) | -| `--frequency-penalty N` | repeat alpha frequency penalty (default: 0.0, 0.0 = disabled) | -| `--dry-multiplier N` | set DRY sampling multiplier (default: 0.0, 0.0 = disabled) | +| `--repeat-penalty N` | penalize repeat sequence of tokens (default: 1.00, 1.0 = disabled) | +| `--presence-penalty N` | repeat alpha presence penalty (default: 0.00, 0.0 = disabled) | +| `--frequency-penalty N` | repeat alpha frequency penalty (default: 0.00, 0.0 = disabled) | +| `--dry-multiplier N` | set DRY sampling multiplier (default: 0.00, 0.0 = disabled) | | `--dry-base N` | set DRY sampling base value (default: 1.75) | | `--dry-allowed-length N` | set allowed length for DRY sampling (default: 2) | | `--dry-penalty-last-n N` | set DRY penalty for the last n tokens (default: -1, 0 = disable, -1 = context size) | | `--dry-sequence-breaker STRING` | add sequence breaker for DRY sampling, clearing out default breakers ('\n', ':', '"', '*') in the process; use "none" to not use any sequence breakers | -| `--dynatemp-range N` | dynamic temperature range (default: 0.0, 0.0 = disabled) | -| `--dynatemp-exp N` | dynamic temperature exponent (default: 1.0) | +| `--adaptive-target N` | adaptive-p: select tokens near this probability (valid range 0.0 to 1.0; negative = disabled) (default: -1.00)
[(more info)](https://github.com/ggml-org/llama.cpp/pull/17927) | +| `--adaptive-decay N` | adaptive-p: decay rate for target adaptation over time. lower values are more reactive, higher values are more stable.
(valid range 0.0 to 0.99) (default: 0.90) | +| `--dynatemp-range N` | dynamic temperature range (default: 0.00, 0.0 = disabled) | +| `--dynatemp-exp N` | dynamic temperature exponent (default: 1.00) | | `--mirostat N` | use Mirostat sampling.
Top K, Nucleus and Locally Typical samplers are ignored if used.
(default: 0, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0) | -| `--mirostat-lr N` | Mirostat learning rate, parameter eta (default: 0.1) | -| `--mirostat-ent N` | Mirostat target entropy, parameter tau (default: 5.0) | +| `--mirostat-lr N` | Mirostat learning rate, parameter eta (default: 0.10) | +| `--mirostat-ent N` | Mirostat target entropy, parameter tau (default: 5.00) | | `-l, --logit-bias TOKEN_ID(+/-)BIAS` | modifies the likelihood of token appearing in the completion,
i.e. `--logit-bias 15043+1` to increase likelihood of token ' Hello',
or `--logit-bias 15043-1` to decrease likelihood of token ' Hello' | | `--grammar GRAMMAR` | BNF-like grammar to constrain generations (see samples in grammars/ dir) (default: '') | | `--grammar-file FNAME` | file to read grammar from | @@ -251,8 +253,8 @@ llama-completion.exe -m models\gemma-1.1-7b-it.Q4_K_M.gguf --ignore-eos -n -1 | `--jinja, --no-jinja` | whether to use jinja template engine for chat (default: disabled)
(env: LLAMA_ARG_JINJA) | | `--reasoning-format FORMAT` | controls whether thought tags are allowed and/or extracted from the response, and in which format they're returned; one of:
- none: leaves thoughts unparsed in `message.content`
- deepseek: puts thoughts in `message.reasoning_content`
- deepseek-legacy: keeps `` tags in `message.content` while also populating `message.reasoning_content`
(default: auto)
(env: LLAMA_ARG_THINK) | | `--reasoning-budget N` | controls the amount of thinking allowed; currently only one of: -1 for unrestricted thinking budget, or 0 to disable thinking (default: -1)
(env: LLAMA_ARG_THINK_BUDGET) | -| `--chat-template JINJA_TEMPLATE` | set custom jinja chat template (default: template taken from model's metadata)
if suffix/prefix are specified, template will be disabled
only commonly used templates are accepted (unless --jinja is set before this flag):
list of built-in templates:
bailing, bailing-think, bailing2, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, grok-2, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, pangu-embedded, phi3, phi4, rwkv-world, seed_oss, smolvlm, solar-open, vicuna, vicuna-orca, yandex, zephyr
(env: LLAMA_ARG_CHAT_TEMPLATE) | -| `--chat-template-file JINJA_TEMPLATE_FILE` | set custom jinja chat template file (default: template taken from model's metadata)
if suffix/prefix are specified, template will be disabled
only commonly used templates are accepted (unless --jinja is set before this flag):
list of built-in templates:
bailing, bailing-think, bailing2, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, grok-2, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, pangu-embedded, phi3, phi4, rwkv-world, seed_oss, smolvlm, solar-open, vicuna, vicuna-orca, yandex, zephyr
(env: LLAMA_ARG_CHAT_TEMPLATE_FILE) | +| `--chat-template JINJA_TEMPLATE` | set custom jinja chat template (default: template taken from model's metadata)
if suffix/prefix are specified, template will be disabled
only commonly used templates are accepted (unless --jinja is set before this flag):
list of built-in templates:
bailing, bailing-think, bailing2, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone-moe, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, grok-2, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, pangu-embedded, phi3, phi4, rwkv-world, seed_oss, smolvlm, solar-open, vicuna, vicuna-orca, yandex, zephyr
(env: LLAMA_ARG_CHAT_TEMPLATE) | +| `--chat-template-file JINJA_TEMPLATE_FILE` | set custom jinja chat template file (default: template taken from model's metadata)
if suffix/prefix are specified, template will be disabled
only commonly used templates are accepted (unless --jinja is set before this flag):
list of built-in templates:
bailing, bailing-think, bailing2, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone-moe, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, grok-2, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, pangu-embedded, phi3, phi4, rwkv-world, seed_oss, smolvlm, solar-open, vicuna, vicuna-orca, yandex, zephyr
(env: LLAMA_ARG_CHAT_TEMPLATE_FILE) | | `--simple-io` | use basic IO for better compatibility in subprocesses and limited consoles | diff --git a/tools/completion/completion.cpp b/tools/completion/completion.cpp index a9eda119d7..f368a2f4c6 100644 --- a/tools/completion/completion.cpp +++ b/tools/completion/completion.cpp @@ -342,44 +342,51 @@ int main(int argc, char ** argv) { return 1; } - // debug message about similarity of saved session, if applicable - size_t n_matching_session_tokens = 0; - if (!session_tokens.empty()) { - for (llama_token id : session_tokens) { - if (n_matching_session_tokens >= embd_inp.size() || id != embd_inp[n_matching_session_tokens]) { - break; + bool session_do_save = false; + + { + size_t n_match = 0; + + if (!session_tokens.empty()) { + for (llama_token id : session_tokens) { + if (n_match >= embd_inp.size() || id != embd_inp[n_match]) { + break; + } + n_match++; + } + if (params.prompt.empty() && n_match == embd_inp.size()) { + LOG_INF("%s: using full prompt from session file\n", __func__); + } else if (n_match >= embd_inp.size()) { + LOG_INF("%s: session file has exact match for prompt!\n", __func__); + } else if (n_match < (embd_inp.size() / 2)) { + LOG_WRN("%s: session file has low similarity to prompt (%zu / %zu tokens); will mostly be reevaluated\n", + __func__, n_match, embd_inp.size()); + } else { + LOG_INF("%s: session file matches %zu / %zu tokens of prompt\n", + __func__, n_match, embd_inp.size()); + } + + if (session_tokens.size() == n_match) { + // [TAG_CONTEXT_STATE_LOGITS] + // in this case, we are going to reuse the logits from the session + // if we ever decide to remove the logits from the session, we need to handle this somehow + // ref: https://github.com/ggml-org/llama.cpp/pull/18862#issuecomment-3756330941 + } + + // remove any "future" tokens that we might have inherited from the previous session + if (session_tokens.size() > n_match) { + if (!llama_memory_seq_rm(mem, -1, n_match, -1)) { + LOG_WRN("%s: unable to resuse common prefix (for example, when the memory is recurrent)\n", __func__); + llama_memory_clear(mem, true); + session_tokens.clear(); + n_match = 0; + } else { + session_tokens.resize(n_match); + } } - n_matching_session_tokens++; - } - if (params.prompt.empty() && n_matching_session_tokens == embd_inp.size()) { - LOG_INF("%s: using full prompt from session file\n", __func__); - } else if (n_matching_session_tokens >= embd_inp.size()) { - LOG_INF("%s: session file has exact match for prompt!\n", __func__); - } else if (n_matching_session_tokens < (embd_inp.size() / 2)) { - LOG_WRN("%s: session file has low similarity to prompt (%zu / %zu tokens); will mostly be reevaluated\n", - __func__, n_matching_session_tokens, embd_inp.size()); - } else { - LOG_INF("%s: session file matches %zu / %zu tokens of prompt\n", - __func__, n_matching_session_tokens, embd_inp.size()); } - // remove any "future" tokens that we might have inherited from the previous session - if (!llama_memory_seq_rm(mem, -1, n_matching_session_tokens, -1)) { - LOG_INF("%s: unable to resuse common prefix\n", __func__); - n_matching_session_tokens = 0; - llama_memory_seq_rm(mem, -1, -1, -1); - } - } - - LOG_DBG("recalculate the cached logits (check): embd_inp.size() %zu, n_matching_session_tokens %zu, embd_inp.size() %zu, session_tokens.size() %zu\n", - embd_inp.size(), n_matching_session_tokens, embd_inp.size(), session_tokens.size()); - - // if we will use the cache for the full prompt without reaching the end of the cache, force - // reevaluation of the last token to recalculate the cached logits - if (!embd_inp.empty() && n_matching_session_tokens == embd_inp.size() && session_tokens.size() > embd_inp.size()) { - LOG_DBG("recalculate the cached logits (do): session_tokens.resize( %zu )\n", embd_inp.size() - 1); - - session_tokens.resize(embd_inp.size() - 1); + session_do_save = !path_session.empty() && n_match < embd_inp.size() && !params.prompt_cache_ro; } // number of tokens to keep when resetting context @@ -521,10 +528,9 @@ int main(int argc, char ** argv) { is_interacting = params.interactive_first; } - bool is_antiprompt = false; - bool input_echo = true; - bool display = true; - bool need_to_save_session = !path_session.empty() && n_matching_session_tokens < embd_inp.size(); + bool is_antiprompt = false; + bool input_echo = true; + bool display = true; int n_past = 0; int n_remain = params.n_predict; @@ -700,8 +706,8 @@ int main(int argc, char ** argv) { if ((int) embd_inp.size() <= n_consumed && !is_interacting) { // optionally save the session on first sample (for faster prompt loading next time) - if (!path_session.empty() && need_to_save_session && !params.prompt_cache_ro) { - need_to_save_session = false; + if (session_do_save) { + session_do_save = false; llama_state_save_file(ctx, path_session.c_str(), session_tokens.data(), session_tokens.size()); LOG_DBG("saved session to %s\n", path_session.c_str()); diff --git a/tools/fit-params/fit-params.cpp b/tools/fit-params/fit-params.cpp index f9d9cb34c7..0176be06e7 100644 --- a/tools/fit-params/fit-params.cpp +++ b/tools/fit-params/fit-params.cpp @@ -36,7 +36,7 @@ int main(int argc, char ** argv) { LOG_INF("%s: printing fitted CLI arguments to stdout...\n", __func__); common_log_flush(common_log_main()); - printf("-c %" PRIu32 " -ngl %" PRIu32, cparams.n_ctx, mparams.n_gpu_layers); + printf("-c %" PRIu32 " -ngl %" PRIi32, cparams.n_ctx, mparams.n_gpu_layers); size_t nd = llama_max_devices(); while (nd > 1 && mparams.tensor_split[nd - 1] == 0.0f) { diff --git a/tools/server/README.md b/tools/server/README.md index f113f9cb75..d132830171 100644 --- a/tools/server/README.md +++ b/tools/server/README.md @@ -63,10 +63,10 @@ For the ful list of features, please refer to [server's changelog](https://githu | `--rope-freq-base N` | RoPE base frequency, used by NTK-aware scaling (default: loaded from model)
(env: LLAMA_ARG_ROPE_FREQ_BASE) | | `--rope-freq-scale N` | RoPE frequency scaling factor, expands context by a factor of 1/N
(env: LLAMA_ARG_ROPE_FREQ_SCALE) | | `--yarn-orig-ctx N` | YaRN: original context size of model (default: 0 = model training context size)
(env: LLAMA_ARG_YARN_ORIG_CTX) | -| `--yarn-ext-factor N` | YaRN: extrapolation mix factor (default: -1.0, 0.0 = full interpolation)
(env: LLAMA_ARG_YARN_EXT_FACTOR) | -| `--yarn-attn-factor N` | YaRN: scale sqrt(t) or attention magnitude (default: -1.0)
(env: LLAMA_ARG_YARN_ATTN_FACTOR) | -| `--yarn-beta-slow N` | YaRN: high correction dim or alpha (default: -1.0)
(env: LLAMA_ARG_YARN_BETA_SLOW) | -| `--yarn-beta-fast N` | YaRN: low correction dim or beta (default: -1.0)
(env: LLAMA_ARG_YARN_BETA_FAST) | +| `--yarn-ext-factor N` | YaRN: extrapolation mix factor (default: -1.00, 0.0 = full interpolation)
(env: LLAMA_ARG_YARN_EXT_FACTOR) | +| `--yarn-attn-factor N` | YaRN: scale sqrt(t) or attention magnitude (default: -1.00)
(env: LLAMA_ARG_YARN_ATTN_FACTOR) | +| `--yarn-beta-slow N` | YaRN: high correction dim or alpha (default: -1.00)
(env: LLAMA_ARG_YARN_BETA_SLOW) | +| `--yarn-beta-fast N` | YaRN: low correction dim or beta (default: -1.00)
(env: LLAMA_ARG_YARN_BETA_FAST) | | `-kvo, --kv-offload, -nkvo, --no-kv-offload` | whether to enable KV cache offloading (default: enabled)
(env: LLAMA_ARG_KV_OFFLOAD) | | `--repack, -nr, --no-repack` | whether to enable weight repacking (default: enabled)
(env: LLAMA_ARG_REPACK) | | `--no-host` | bypass host buffer allowing extra buffers to be used
(env: LLAMA_ARG_NO_HOST) | @@ -126,30 +126,30 @@ For the ful list of features, please refer to [server's changelog](https://githu | `-s, --seed SEED` | RNG seed (default: -1, use random seed for -1) | | `--sampler-seq, --sampling-seq SEQUENCE` | simplified sequence for samplers that will be used (default: edskypmxt) | | `--ignore-eos` | ignore end of stream token and continue generating (implies --logit-bias EOS-inf) | -| `--temp N` | temperature (default: 0.8) | +| `--temp N` | temperature (default: 0.80) | | `--top-k N` | top-k sampling (default: 40, 0 = disabled)
(env: LLAMA_ARG_TOP_K) | -| `--top-p N` | top-p sampling (default: 0.9, 1.0 = disabled) | -| `--min-p N` | min-p sampling (default: 0.1, 0.0 = disabled) | -| `--adaptive-target N` | adaptive-p: select tokens near this probability (valid range 0.0 to 1.0; negative = disabled) | -| `--adaptive-decay N` | adaptive-p: EMA decay for adaptation; effective history length ≈ 1/(1-decay) tokens (valid range 0.0 - 0.99) | -| `--top-nsigma N` | top-n-sigma sampling (default: -1.0, -1.0 = disabled) | -| `--xtc-probability N` | xtc probability (default: 0.0, 0.0 = disabled) | -| `--xtc-threshold N` | xtc threshold (default: 0.1, 1.0 = disabled) | -| `--typical N` | locally typical sampling, parameter p (default: 1.0, 1.0 = disabled) | +| `--top-p N` | top-p sampling (default: 0.95, 1.0 = disabled) | +| `--min-p N` | min-p sampling (default: 0.05, 0.0 = disabled) | +| `--top-nsigma N` | top-n-sigma sampling (default: -1.00, -1.0 = disabled) | +| `--xtc-probability N` | xtc probability (default: 0.00, 0.0 = disabled) | +| `--xtc-threshold N` | xtc threshold (default: 0.10, 1.0 = disabled) | +| `--typical N` | locally typical sampling, parameter p (default: 1.00, 1.0 = disabled) | | `--repeat-last-n N` | last n tokens to consider for penalize (default: 64, 0 = disabled, -1 = ctx_size) | -| `--repeat-penalty N` | penalize repeat sequence of tokens (default: 1.0, 1.0 = disabled) | -| `--presence-penalty N` | repeat alpha presence penalty (default: 0.0, 0.0 = disabled) | -| `--frequency-penalty N` | repeat alpha frequency penalty (default: 0.0, 0.0 = disabled) | -| `--dry-multiplier N` | set DRY sampling multiplier (default: 0.0, 0.0 = disabled) | +| `--repeat-penalty N` | penalize repeat sequence of tokens (default: 1.00, 1.0 = disabled) | +| `--presence-penalty N` | repeat alpha presence penalty (default: 0.00, 0.0 = disabled) | +| `--frequency-penalty N` | repeat alpha frequency penalty (default: 0.00, 0.0 = disabled) | +| `--dry-multiplier N` | set DRY sampling multiplier (default: 0.00, 0.0 = disabled) | | `--dry-base N` | set DRY sampling base value (default: 1.75) | | `--dry-allowed-length N` | set allowed length for DRY sampling (default: 2) | | `--dry-penalty-last-n N` | set DRY penalty for the last n tokens (default: -1, 0 = disable, -1 = context size) | | `--dry-sequence-breaker STRING` | add sequence breaker for DRY sampling, clearing out default breakers ('\n', ':', '"', '*') in the process; use "none" to not use any sequence breakers | -| `--dynatemp-range N` | dynamic temperature range (default: 0.0, 0.0 = disabled) | -| `--dynatemp-exp N` | dynamic temperature exponent (default: 1.0) | +| `--adaptive-target N` | adaptive-p: select tokens near this probability (valid range 0.0 to 1.0; negative = disabled) (default: -1.00)
[(more info)](https://github.com/ggml-org/llama.cpp/pull/17927) | +| `--adaptive-decay N` | adaptive-p: decay rate for target adaptation over time. lower values are more reactive, higher values are more stable.
(valid range 0.0 to 0.99) (default: 0.90) | +| `--dynatemp-range N` | dynamic temperature range (default: 0.00, 0.0 = disabled) | +| `--dynatemp-exp N` | dynamic temperature exponent (default: 1.00) | | `--mirostat N` | use Mirostat sampling.
Top K, Nucleus and Locally Typical samplers are ignored if used.
(default: 0, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0) | -| `--mirostat-lr N` | Mirostat learning rate, parameter eta (default: 0.1) | -| `--mirostat-ent N` | Mirostat target entropy, parameter tau (default: 5.0) | +| `--mirostat-lr N` | Mirostat learning rate, parameter eta (default: 0.10) | +| `--mirostat-ent N` | Mirostat target entropy, parameter tau (default: 5.00) | | `-l, --logit-bias TOKEN_ID(+/-)BIAS` | modifies the likelihood of token appearing in the completion,
i.e. `--logit-bias 15043+1` to increase likelihood of token ' Hello',
or `--logit-bias 15043-1` to decrease likelihood of token ' Hello' | | `--grammar GRAMMAR` | BNF-like grammar to constrain generations (see samples in grammars/ dir) (default: '') | | `--grammar-file FNAME` | file to read grammar from | @@ -199,7 +199,8 @@ For the ful list of features, please refer to [server's changelog](https://githu | `--chat-template-kwargs STRING` | sets additional params for the json template parser, must be a valid json object string, e.g. '{"key1":"value1","key2":"value2"}'
(env: LLAMA_CHAT_TEMPLATE_KWARGS) | | `-to, --timeout N` | server read/write timeout in seconds (default: 600)
(env: LLAMA_ARG_TIMEOUT) | | `--threads-http N` | number of threads used to process HTTP requests (default: -1)
(env: LLAMA_ARG_THREADS_HTTP) | -| `--cache-reuse N` | min chunk size to attempt reusing from the cache via KV shifting (default: 0)
[(card)](https://ggml.ai/f0.png)
(env: LLAMA_ARG_CACHE_REUSE) | +| `--cache-prompt, --no-cache-prompt` | whether to enable prompt caching (default: enabled)
(env: LLAMA_ARG_CACHE_PROMPT) | +| `--cache-reuse N` | min chunk size to attempt reusing from the cache via KV shifting, requires prompt caching to be enabled (default: 0)
[(card)](https://ggml.ai/f0.png)
(env: LLAMA_ARG_CACHE_REUSE) | | `--metrics` | enable prometheus compatible metrics endpoint (default: disabled)
(env: LLAMA_ARG_ENDPOINT_METRICS) | | `--props` | enable changing global properties via POST /props (default: disabled)
(env: LLAMA_ARG_ENDPOINT_PROPS) | | `--slots, --no-slots` | expose slots monitoring endpoint (default: enabled)
(env: LLAMA_ARG_ENDPOINT_SLOTS) | @@ -212,8 +213,8 @@ For the ful list of features, please refer to [server's changelog](https://githu | `--jinja, --no-jinja` | whether to use jinja template engine for chat (default: enabled)
(env: LLAMA_ARG_JINJA) | | `--reasoning-format FORMAT` | controls whether thought tags are allowed and/or extracted from the response, and in which format they're returned; one of:
- none: leaves thoughts unparsed in `message.content`
- deepseek: puts thoughts in `message.reasoning_content`
- deepseek-legacy: keeps `` tags in `message.content` while also populating `message.reasoning_content`
(default: auto)
(env: LLAMA_ARG_THINK) | | `--reasoning-budget N` | controls the amount of thinking allowed; currently only one of: -1 for unrestricted thinking budget, or 0 to disable thinking (default: -1)
(env: LLAMA_ARG_THINK_BUDGET) | -| `--chat-template JINJA_TEMPLATE` | set custom jinja chat template (default: template taken from model's metadata)
if suffix/prefix are specified, template will be disabled
only commonly used templates are accepted (unless --jinja is set before this flag):
list of built-in templates:
bailing, bailing-think, bailing2, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, grok-2, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, pangu-embedded, phi3, phi4, rwkv-world, seed_oss, smolvlm, solar-open, vicuna, vicuna-orca, yandex, zephyr
(env: LLAMA_ARG_CHAT_TEMPLATE) | -| `--chat-template-file JINJA_TEMPLATE_FILE` | set custom jinja chat template file (default: template taken from model's metadata)
if suffix/prefix are specified, template will be disabled
only commonly used templates are accepted (unless --jinja is set before this flag):
list of built-in templates:
bailing, bailing-think, bailing2, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, grok-2, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, pangu-embedded, phi3, phi4, rwkv-world, seed_oss, smolvlm, solar-open, vicuna, vicuna-orca, yandex, zephyr
(env: LLAMA_ARG_CHAT_TEMPLATE_FILE) | +| `--chat-template JINJA_TEMPLATE` | set custom jinja chat template (default: template taken from model's metadata)
if suffix/prefix are specified, template will be disabled
only commonly used templates are accepted (unless --jinja is set before this flag):
list of built-in templates:
bailing, bailing-think, bailing2, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone-moe, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, grok-2, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, pangu-embedded, phi3, phi4, rwkv-world, seed_oss, smolvlm, solar-open, vicuna, vicuna-orca, yandex, zephyr
(env: LLAMA_ARG_CHAT_TEMPLATE) | +| `--chat-template-file JINJA_TEMPLATE_FILE` | set custom jinja chat template file (default: template taken from model's metadata)
if suffix/prefix are specified, template will be disabled
only commonly used templates are accepted (unless --jinja is set before this flag):
list of built-in templates:
bailing, bailing-think, bailing2, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone-moe, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, grok-2, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, pangu-embedded, phi3, phi4, rwkv-world, seed_oss, smolvlm, solar-open, vicuna, vicuna-orca, yandex, zephyr
(env: LLAMA_ARG_CHAT_TEMPLATE_FILE) | | `--prefill-assistant, --no-prefill-assistant` | whether to prefill the assistant's response if the last message is an assistant message (default: prefill enabled)
when this flag is set, if the last message is an assistant message then it will be treated as a full message and not prefilled

(env: LLAMA_ARG_PREFILL_ASSISTANT) | | `-sps, --slot-prompt-similarity SIMILARITY` | how much the prompt of a request must match the prompt of a slot in order to use that slot (default: 0.10, 0.0 = disabled) | | `--lora-init-without-apply` | load LoRA adapters without applying them (apply later via POST /lora-adapters) (default: disabled) | @@ -222,7 +223,7 @@ For the ful list of features, please refer to [server's changelog](https://githu | `-tbd, --threads-batch-draft N` | number of threads to use during batch and prompt processing (default: same as --threads-draft) | | `--draft, --draft-n, --draft-max N` | number of tokens to draft for speculative decoding (default: 16)
(env: LLAMA_ARG_DRAFT_MAX) | | `--draft-min, --draft-n-min N` | minimum number of draft tokens to use for speculative decoding (default: 0)
(env: LLAMA_ARG_DRAFT_MIN) | -| `--draft-p-min P` | minimum speculative decoding probability (greedy) (default: 0.8)
(env: LLAMA_ARG_DRAFT_P_MIN) | +| `--draft-p-min P` | minimum speculative decoding probability (greedy) (default: 0.75)
(env: LLAMA_ARG_DRAFT_P_MIN) | | `-cd, --ctx-size-draft N` | size of the prompt context for the draft model (default: 0, 0 = loaded from model)
(env: LLAMA_ARG_CTX_SIZE_DRAFT) | | `-devd, --device-draft ` | comma-separated list of devices to use for offloading the draft model (none = don't offload)
use --list-devices to see a list of available devices | | `-ngld, --gpu-layers-draft, --n-gpu-layers-draft N` | max. number of draft model layers to store in VRAM, either an exact number, 'auto', or 'all' (default: auto)
(env: LLAMA_ARG_N_GPU_LAYERS_DRAFT) | diff --git a/tools/server/server-context.cpp b/tools/server/server-context.cpp index 73cb4c75b3..1ca4e3cc0e 100644 --- a/tools/server/server-context.cpp +++ b/tools/server/server-context.cpp @@ -48,11 +48,8 @@ enum server_state { struct server_slot { int id; - llama_batch batch_spec = {}; - // TODO: change to unique_ptrs for consistency: llama_context * ctx = nullptr; - llama_context * ctx_dft = nullptr; // multimodal mtmd_context * mctx = nullptr; @@ -259,7 +256,7 @@ struct server_slot { } bool can_speculate() const { - return ctx_dft; + return !!spec; } void add_token(const completion_token_output & token) { @@ -295,6 +292,7 @@ struct server_slot { SLT_DBG(*this, "the max possible draft is too small: %d < %d - skipping speculative decoding\n", n_draft_max, task->params.speculative.n_min); n_draft_max = 0; } + return n_draft_max; } @@ -397,6 +395,8 @@ struct server_slot { draft_ratio, n_draft_accepted, n_draft_total ); } + + common_speculative_print_stats(spec); } json to_json(bool only_metrics = false) const { @@ -553,18 +553,13 @@ private: // note: keep these alive - they determine the lifetime of the model, context, etc. common_init_result_ptr llama_init; - common_init_result_ptr llama_init_dft; llama_context * ctx = nullptr; - bool vocab_dft_compatible = true; - - llama_model * model_dft = nullptr; - - llama_context_params cparams_dft; - llama_batch batch {}; + llama_model_ptr model_dft; + bool add_bos_token = true; int32_t n_ctx; // total context for all clients / slots @@ -597,13 +592,8 @@ private: // Clear any sampling context for (server_slot & slot : slots) { - llama_free(slot.ctx_dft); - slot.ctx_dft = nullptr; - common_speculative_free(slot.spec); slot.spec = nullptr; - - llama_batch_free(slot.batch_spec); } llama_batch_free(batch); @@ -648,44 +638,39 @@ private: add_bos_token = llama_vocab_get_add_bos(vocab); - if (params_base.has_speculative()) { - SRV_INF("loading draft model '%s'\n", params_base.speculative.model.path.c_str()); + if (params_base.speculative.has_dft()) { + SRV_INF("loading draft model '%s'\n", params_base.speculative.mparams_dft.path.c_str()); + + const auto & params_spec = params_base.speculative; auto params_dft = params_base; - params_dft.devices = params_base.speculative.devices; - params_dft.model = params_base.speculative.model; - params_dft.n_ctx = params_base.speculative.n_ctx == 0 ? llama_n_ctx_seq(ctx) : params_base.speculative.n_ctx; - params_dft.n_gpu_layers = params_base.speculative.n_gpu_layers; params_dft.n_parallel = 1; - params_dft.cache_type_k = params_base.speculative.cache_type_k; - params_dft.cache_type_v = params_base.speculative.cache_type_v; + params_dft.n_ctx = params_spec.n_ctx == 0 ? llama_n_ctx_seq(ctx) : params_spec.n_ctx; + params_dft.n_batch = llama_n_ctx_seq(ctx); + params_dft.devices = params_spec.devices; + params_dft.model = params_spec.mparams_dft; + params_dft.n_gpu_layers = params_spec.n_gpu_layers; + params_dft.cache_type_k = params_spec.cache_type_k; + params_dft.cache_type_v = params_spec.cache_type_v; - params_dft.cpuparams.n_threads = params_base.speculative.cpuparams.n_threads; - params_dft.cpuparams_batch.n_threads = params_base.speculative.cpuparams_batch.n_threads; - params_dft.tensor_buft_overrides = params_base.speculative.tensor_buft_overrides; + if (params_spec.cpuparams.n_threads > 0) { + params_dft.cpuparams.n_threads = params_spec.cpuparams.n_threads; + params_dft.cpuparams_batch.n_threads = params_spec.cpuparams_batch.n_threads; + } - llama_init_dft = common_init_from_params(params_dft); + params_dft.tensor_buft_overrides = params_spec.tensor_buft_overrides; - model_dft = llama_init_dft->model(); + auto mparams_dft = common_model_params_to_llama(params_dft); + model_dft.reset(llama_model_load_from_file(params_dft.model.path.c_str(), mparams_dft)); if (model_dft == nullptr) { - SRV_ERR("failed to load draft model, '%s'\n", params_base.speculative.model.path.c_str()); + SRV_ERR("failed to load draft model, '%s'\n", params_dft.model.path.c_str()); return false; } - vocab_dft_compatible = common_speculative_are_compatible(ctx, llama_init_dft->context()); - if (!vocab_dft_compatible) { - SRV_INF("the draft model '%s' is not compatible with the target model '%s'. tokens will be translated between the draft and target models.\n", params_base.speculative.model.path.c_str(), params_base.model.path.c_str()); - } - - const int n_ctx_dft = llama_n_ctx(llama_init_dft->context()); - - cparams_dft = common_context_params_to_llama(params_dft); - cparams_dft.n_batch = n_ctx_dft; - - // the context is not needed - we will create one for each slot - llama_init_dft->free_context(); + params_base.speculative.model_dft = model_dft.get(); + params_base.speculative.cparams_dft = common_context_params_to_llama(params_dft); } std::string & mmproj_path = params_base.mmproj.path; @@ -695,6 +680,7 @@ private: } mtmd_context_params mparams = mtmd_context_params_default(); + mparams.use_gpu = params_base.mmproj_use_gpu; mparams.print_timings = false; mparams.n_threads = params_base.cpuparams.n_threads; @@ -702,6 +688,7 @@ private: mparams.warmup = params_base.warmup; mparams.image_min_tokens = params_base.image_min_tokens; mparams.image_max_tokens = params_base.image_max_tokens; + mctx = mtmd_init_from_file(mmproj_path.c_str(), model, mparams); if (mctx == nullptr) { SRV_ERR("failed to load multimodal model, '%s'\n", mmproj_path.c_str()); @@ -718,11 +705,6 @@ private: params_base.n_cache_reuse = 0; SRV_WRN("%s\n", "cache_reuse is not supported by multimodal, it will be disabled"); } - - if (params_base.has_speculative()) { - SRV_ERR("%s\n", "err: speculative decode is not supported by multimodal"); - return false; - } } if (!llama_memory_can_shift(llama_get_memory(ctx))) { @@ -757,29 +739,24 @@ private: for (int i = 0; i < params_base.n_parallel; i++) { server_slot slot; - slot.id = i; - slot.ctx = ctx; + slot.id = i; + slot.ctx = ctx; slot.n_ctx = n_ctx_slot; - slot.mctx = mctx; + + slot.mctx = mctx; slot.prompt.tokens.has_mtmd = mctx != nullptr; - if (model_dft) { - slot.batch_spec = llama_batch_init(params_base.speculative.n_max + 1, 0, 1); - - // TODO: rework speculative decoding [TAG_SERVER_SPEC_REWORK] - slot.ctx_dft = llama_init_from_model(model_dft, cparams_dft); - if (slot.ctx_dft == nullptr) { - SRV_ERR("%s", "failed to create draft context\n"); - return false; - } - - slot.spec = common_speculative_init(slot.ctx, slot.ctx_dft); - if (slot.spec == nullptr) { - SRV_ERR("%s", "failed to create speculator\n"); - return false; - } - for (auto & pair : params_base.speculative.replacements) { - common_speculative_add_replacement_tgt_dft(slot.spec, pair.first.c_str(), pair.second.c_str()); + // try speculative decoding + { + slot.spec = common_speculative_init(params_base.speculative, slot.ctx); + if (slot.spec) { + if (mctx) { + SRV_ERR("%s\n", "speculative decoding is not supported with multimodal"); + return false; + } + SRV_WRN("%s", "speculative decoding context initialized\n"); + } else { + SRV_WRN("%s", "speculative decoding context not initialized\n"); } } @@ -1059,7 +1036,7 @@ private: return res; } - std::vector construct_lora_list(const std::map & config) { + std::vector construct_lora_list(const std::map & config) const { std::vector output = params_base.lora_adapters; // copy for (size_t i = 0; i < output.size(); ++i) { auto it = config.find(i); @@ -1162,7 +1139,7 @@ private: backend_sampling &= task.params.sampling.backend_sampling; // TODO: speculative decoding requires multiple samples per batch - not supported yet - backend_sampling &= !(slot.ctx_dft && task.params.speculative.n_max > 0); + backend_sampling &= !(slot.spec && task.params.speculative.n_max > 0); // TODO: getting post/pre sampling logits is not yet supported with backend sampling backend_sampling &= !need_logits; @@ -1179,14 +1156,6 @@ private: slot.smpl.reset(); } - // initialize draft batch - // TODO: rework speculative decoding [TAG_SERVER_SPEC_REWORK] - if (slot.ctx_dft) { - llama_batch_free(slot.batch_spec); - - slot.batch_spec = llama_batch_init(task.params.speculative.n_max + 1, 0, 1); - } - slot.task = std::make_unique(std::move(task)); slot.state = slot.task->is_child() @@ -2059,19 +2028,23 @@ private: // generate draft tokens in speculative decoding mode // TODO: rework to have a single draft llama_context shared across all slots [TAG_SERVER_SPEC_REWORK] // perform the speculative drafting for all sequences at the same time in a single batch - int n_draft_max = slot.get_n_draft_max(); + const int n_draft_max = slot.get_n_draft_max(); if (n_draft_max > 0) { if (mctx) { // we should never reach this, as speculative is automatically disabled if mmproj is loaded GGML_ABORT("not supported by multimodal"); } - struct common_speculative_params params_spec; - params_spec.n_draft = n_draft_max; - params_spec.n_reuse = llama_n_ctx(slot.ctx_dft) - slot.task->params.speculative.n_max; - params_spec.p_min = slot.task->params.speculative.p_min; const llama_tokens & cached_text_tokens = slot.prompt.tokens.get_text_tokens(); - llama_tokens draft = common_speculative_gen_draft(slot.spec, params_spec, cached_text_tokens, slot.sampled); + + const auto & params_spec = slot.task->params.speculative; + + llama_tokens draft = common_speculative_draft(slot.spec, params_spec, cached_text_tokens, slot.sampled); + + if (draft.size() > (size_t) n_draft_max) { + SLT_WRN(slot, "draft size %d exceeds max %d, truncating\n", (int) draft.size(), n_draft_max); + draft.resize(n_draft_max); + } // add the sampled token to the batch slot.i_batch_dft.push_back(batch.n_tokens); @@ -2742,6 +2715,10 @@ private: // prompt evaluated for next-token prediction slot.state = SLOT_STATE_GENERATING; + + if (slot.can_speculate()) { + common_speculative_begin(slot.spec, slot.prompt.tokens.get_text_tokens()); + } } else if (slot.state != SLOT_STATE_GENERATING) { continue; // continue loop of slots } @@ -2813,6 +2790,9 @@ private: // update how many tokens out of those tested were accepted slot.n_draft_accepted += ids.size() - 1; + // inform the speculative decoding about the number of accepted tokens + common_speculative_accept(slot.spec, ids.size() - 1); + // rollback to the state before sampling the draft tokens slot.prompt.tokens.keep_first(slot.prompt.n_tokens() - n_draft); diff --git a/tools/server/server-task.cpp b/tools/server/server-task.cpp index 799e341d37..2d25db63b7 100644 --- a/tools/server/server-task.cpp +++ b/tools/server/server-task.cpp @@ -5,6 +5,7 @@ #include "llama.h" #include "chat.h" #include "sampling.h" +#include "speculative.h" #include "json-schema-to-grammar.h" using json = nlohmann::ordered_json; @@ -76,6 +77,11 @@ json task_params::to_json(bool only_metrics) const { {"speculative.n_max", speculative.n_max}, {"speculative.n_min", speculative.n_min}, {"speculative.p_min", speculative.p_min}, + {"speculative.type", common_speculative_type_to_str(speculative.type)}, + {"speculative.ngram_size_n", speculative.ngram_size_n}, + {"speculative.ngram_size_m", speculative.ngram_size_m}, + {"speculative.ngram_c_rate", speculative.ngram_check_rate}, + {"speculative.ngram_m_hits", speculative.ngram_min_hits}, {"timings_per_token", timings_per_token}, {"post_sampling_probs", post_sampling_probs}, {"backend_sampling", sampling.backend_sampling}, @@ -135,6 +141,11 @@ json task_params::to_json(bool only_metrics) const { {"speculative.n_max", speculative.n_max}, {"speculative.n_min", speculative.n_min}, {"speculative.p_min", speculative.p_min}, + {"speculative.type", common_speculative_type_to_str(speculative.type)}, + {"speculative.ngram_size_n", speculative.ngram_size_n}, + {"speculative.ngram_size_m", speculative.ngram_size_m}, + {"speculative.ngram_c_rate", speculative.ngram_check_rate}, + {"speculative.ngram_m_hits", speculative.ngram_min_hits}, {"timings_per_token", timings_per_token}, {"post_sampling_probs", post_sampling_probs}, {"backend_sampling", sampling.backend_sampling}, @@ -242,6 +253,18 @@ task_params server_task::params_from_json_cmpl( params.speculative.n_min = std::max(params.speculative.n_min, 0); params.speculative.n_max = std::max(params.speculative.n_max, 0); + params.speculative.type = common_speculative_type_from_name(json_value(data, "speculative.type", common_speculative_type_to_str(defaults.speculative.type))); + + params.speculative.ngram_size_n = json_value(data, "speculative.ngram_size_n", defaults.speculative.ngram_size_n); + params.speculative.ngram_size_m = json_value(data, "speculative.ngram_size_m", defaults.speculative.ngram_size_m); + params.speculative.ngram_check_rate = json_value(data, "speculative.ngram_c_rate", defaults.speculative.ngram_check_rate); + params.speculative.ngram_min_hits = json_value(data, "speculative.ngram_m_hits", defaults.speculative.ngram_min_hits); + + params.speculative.ngram_size_n = std::max(std::min(1, (int) params.speculative.ngram_size_n), 1024); + params.speculative.ngram_size_m = std::max(std::min(1, (int) params.speculative.ngram_size_m), 1024); + params.speculative.ngram_check_rate = std::max(std::min(1, (int) params.speculative.ngram_check_rate), 1024); + params.speculative.ngram_min_hits = std::max(std::min(1, (int) params.speculative.ngram_min_hits), 1024); + // Use OpenAI API logprobs only if n_probs wasn't provided if (data.contains("logprobs") && params.sampling.n_probs == defaults.sampling.n_probs){ params.sampling.n_probs = json_value(data, "logprobs", defaults.sampling.n_probs); diff --git a/tools/server/tests/unit/test_speculative.py b/tools/server/tests/unit/test_speculative.py index d2f3fba5fe..eebd3cc8fa 100644 --- a/tools/server/tests/unit/test_speculative.py +++ b/tools/server/tests/unit/test_speculative.py @@ -30,6 +30,7 @@ def test_with_and_without_draft(): "prompt": "I believe the meaning of life is", "temperature": 0.0, "top_k": 1, + "n_predict": 16, }) assert res.status_code == 200 content_no_draft = res.body["content"] @@ -42,6 +43,7 @@ def test_with_and_without_draft(): "prompt": "I believe the meaning of life is", "temperature": 0.0, "top_k": 1, + "n_predict": 16, }) assert res.status_code == 200 content_draft = res.body["content"] @@ -68,6 +70,7 @@ def test_different_draft_min_draft_max(): "prompt": "I believe the meaning of life is", "temperature": 0.0, "top_k": 1, + "n_predict": 16, }) assert res.status_code == 200 if last_content is not None: