happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'master' into pr/19440

This commit is contained in:
inforithmics 2026-02-12 17:18:10 +01:00
parent 5a800fc88d 57487a64c8
commit f5f2203ed4
53 changed files with 4038 additions and 569 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

73
.github/workflows/server-metal.yml vendored Normal file
View File

@ -0,0 +1,73 @@
name: Server-Metal
on:
workflow_dispatch: # allows manual triggering
inputs:
sha:
description: 'Commit SHA1 to build'
required: false
type: string
slow_tests:
description: 'Run slow tests'
required: true
type: boolean
push:
branches:
- master
paths: ['.github/workflows/server-metal.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'tools/server/**.*']
env:
LLAMA_LOG_COLORS: 1
LLAMA_LOG_PREFIX: 1
LLAMA_LOG_TIMESTAMPS: 1
LLAMA_LOG_VERBOSITY: 10
concurrency:
group: ${{ github.workflow }}-${{ github.ref }}-${{ github.head_ref || github.run_id }}
cancel-in-progress: true
jobs:
server-metal:
runs-on: [self-hosted, macOS, ARM64]
name: server-metal (${{ matrix.wf_name }})
strategy:
matrix:
build_type: [Release]
wf_name: ["GPUx1"]
include:
- build_type: Release
extra_args: "LLAMA_ARG_BACKEND_SAMPLING=1"
wf_name: "GPUx1, backend-sampling"
- build_type: Release
extra_args: "GGML_METAL_DEVICES=2"
wf_name: "GPUx2"
- build_type: Release
extra_args: "GGML_METAL_DEVICES=2 LLAMA_ARG_BACKEND_SAMPLING=1"
wf_name: "GPUx2, backend-sampling"
fail-fast: false
steps:
- name: Clone
id: checkout
uses: actions/checkout@v6
with:
fetch-depth: 0
ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}
- name: Build
id: cmake_build
run: |
cmake -B build -DGGML_SCHED_NO_REALLOC=ON
cmake --build build --config ${{ matrix.build_type }} -j $(sysctl -n hw.logicalcpu) --target llama-server
- name: Tests
id: server_integration_tests
if: ${{ (!matrix.disabled_on_pr || !github.event.pull_request) }}
run: |
cd tools/server/tests
python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt
export ${{ matrix.extra_args }}
pytest -v -x -m "not slow"

1
CMakeLists.txt
View File

@ -109,6 +109,7 @@ option(LLAMA_BUILD_TOOLS "llama: build tools" ${LLAMA_STANDALONE})
option(LLAMA_BUILD_EXAMPLES "llama: build examples" ${LLAMA_STANDALONE})
option(LLAMA_BUILD_SERVER "llama: build server example" ${LLAMA_STANDALONE})
option(LLAMA_TOOLS_INSTALL "llama: install tools" ${LLAMA_TOOLS_INSTALL_DEFAULT})
option(LLAMA_TESTS_INSTALL "llama: install tests" ON)
# 3rd party libs
option(LLAMA_HTTPLIB "llama: httplib for downloading functionality" ON)

1
README.md
View File

@ -288,6 +288,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
| [WebGPU [In Progress]](docs/build.md#webgpu) | All |
| [RPC](https://github.com/ggml-org/llama.cpp/tree/master/tools/rpc) | All |
| [Hexagon [In Progress]](docs/backend/hexagon/README.md) | Snapdragon |
| [VirtGPU](docs/backend/VirtGPU.md) | VirtGPU APIR |
## Obtaining and quantizing models

10
common/arg.cpp
View File

@ -3437,16 +3437,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
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),

39
common/chat.cpp
View File

@ -380,15 +380,46 @@ std::vector<common_chat_msg> common_chat_msgs_parse_oaicompat(const json & messa
return msgs;
}
json common_chat_msgs_to_json_oaicompat(const std::vector<common_chat_msg> & msgs, bool concat_typed_text) {
static json render_message_to_json(const std::vector<common_chat_msg> & msgs, const jinja::caps & c) {
if (!c.supports_string_content && !c.supports_typed_content) {
LOG_WRN("%s: Neither string content nor typed content is supported by the template. This is unexpected and may lead to issues.\n", __func__);
}
bool only_string_accepted = c.supports_string_content && !c.supports_typed_content;
bool only_typed_accepted = !c.supports_string_content && c.supports_typed_content;
json messages = json::array();
for (const auto & msg : msgs) {
json jmsg = msg.to_json_oaicompat(concat_typed_text);
messages.push_back(jmsg);
if (only_string_accepted) {
json jmsg = msg.to_json_oaicompat(/* concat_typed_text= */ true);
messages.push_back(jmsg);
} else if (only_typed_accepted) {
json jmsg = msg.to_json_oaicompat(/* concat_typed_text= */ false);
if (jmsg.at("content").is_string()) {
jmsg["content"] = json::array({
json{
{"type", "text"},
{"text", jmsg.at("content").get<std::string>()},
}
});
}
messages.push_back(jmsg);
} else {
json jmsg = msg.to_json_oaicompat(/* concat_typed_text= */ false);
messages.push_back(jmsg);
}
}
return messages;
}
// DEPRECATED: only used in tests
json common_chat_msgs_to_json_oaicompat(const std::vector<common_chat_msg> & msgs, bool concat_typed_text) {
jinja::caps c;
c.supports_string_content = true;
c.supports_typed_content = !concat_typed_text;
return render_message_to_json(msgs, c);
}
std::vector<common_chat_tool> common_chat_tools_parse_oaicompat(const json & tools) {
std::vector<common_chat_tool> result;
@ -3020,7 +3051,7 @@ static common_chat_params common_chat_templates_apply_jinja(
: *tmpls->template_default;
const auto & src = tmpl.source();
const auto & caps = tmpl.original_caps();
params.messages = common_chat_msgs_to_json_oaicompat(inputs.messages, /* concat_text= */ !tmpl.original_caps().requires_typed_content);
params.messages = render_message_to_json(inputs.messages, tmpl.original_caps());
params.add_generation_prompt = inputs.add_generation_prompt;
params.tool_choice = inputs.tool_choice;
params.reasoning_format = inputs.reasoning_format;

2
common/chat.h
View File

@ -240,6 +240,8 @@ bool common_chat_templates_support_enable_thinking(const common_chat_templates *
// Parses a JSON array of messages in OpenAI's chat completion API format.
std::vector<common_chat_msg> common_chat_msgs_parse_oaicompat(const nlohmann::ordered_json & messages);
// DEPRECATED: only used in tests
nlohmann::ordered_json common_chat_msgs_to_json_oaicompat(const std::vector<common_chat_msg> & msgs, bool concat_typed_text = false);
std::vector<common_chat_tool> common_chat_tools_parse_oaicompat(const nlohmann::ordered_json & tools);

1
common/common.h
View File

@ -269,7 +269,6 @@ struct common_params_speculative {
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::shared_ptr<common_ngram_mod> ngram_mod;

13
common/jinja/caps.cpp
View File

@ -63,7 +63,8 @@ static void caps_print_stats(value & v, const std::string & path) {
std::map<std::string, bool> caps::to_map() const {
return {
{"requires_typed_content", requires_typed_content},
{"supports_string_content", supports_string_content},
{"supports_typed_content", supports_typed_content},
{"supports_tools", supports_tools},
{"supports_tool_calls", supports_tool_calls},
{"supports_parallel_tool_calls", supports_parallel_tool_calls},
@ -89,7 +90,7 @@ caps caps_get(jinja::program & prog) {
return v->stats.ops.find(op_name) != v->stats.ops.end();
};
// case: typed content requirement
// case: typed content support
caps_try_execute(
prog,
[&]() {
@ -105,12 +106,16 @@ caps caps_get(jinja::program & prog) {
// tools
return json{nullptr};
},
[&](bool, value & messages, value &) {
[&](bool success, value & messages, value &) {
auto & content = messages->at(0)->at("content");
caps_print_stats(content, "messages[0].content");
if (has_op(content, "selectattr") || has_op(content, "array_access")) {
// accessed as an array
result.requires_typed_content = true;
result.supports_typed_content = true;
}
if (!success) {
// failed to execute with content as string
result.supports_string_content = false;
}
}
);

4
common/jinja/caps.h
View File

@ -14,7 +14,9 @@ struct caps {
bool supports_parallel_tool_calls = true;
bool supports_preserve_reasoning = false; // support assistant message with reasoning_content
bool requires_typed_content = false; // default: use string content
// one of the 2 content capabilities must be true
bool supports_string_content = true;
bool supports_typed_content = false;
// for reporting on server
std::map<std::string, bool> to_map() const;

6
common/jinja/runtime.cpp
View File

@ -446,6 +446,12 @@ value for_statement::execute_impl(context & ctx) {
value iterable_val = iter_expr->execute(scope);
// mark the variable being iterated as used for stats
if (ctx.is_get_stats) {
iterable_val->stats.used = true;
iterable_val->stats.ops.insert("array_access");
}
if (iterable_val->is_undefined()) {
JJ_DEBUG("%s", "For loop iterable is undefined, skipping loop");
iterable_val = mk_val<value_array>();

7
common/ngram-map.cpp
View File

@ -231,10 +231,9 @@ void common_ngram_map_draft(common_ngram_map & map,
GGML_ABORT("%s: cur_len exceeds UINT32_MAX: %zu", __func__, cur_len);
}
// 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;
if (map.idx_last_check > cur_len) {
// Should not happen because of common_ngram_map_begin().
GGML_ABORT("%s: map.idx_last_check > cur_len: %zu > %zu", __func__, map.idx_last_check, cur_len);
}
map.idx_last_check = cur_len;

8
common/ngram-map.h
View File

@ -24,7 +24,6 @@
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
};
// Searches for a n-gram in the history and checks whether a draft sequence should be generated.
@ -66,15 +65,14 @@ struct common_ngram_map {
bool key_only; // true if only key n-grams are used, no values.
std::vector<common_ngram_map_key> 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
bool show_key_map_stats = false; // true, if statitics of the key_map should be printed.
bool show_key_map_stats = false; // true, if statistics of the key_map should be printed.
common_ngram_map(uint16_t sz_key, uint16_t sz_value, bool only_keys,
uint16_t check_rate, uint16_t min_hits)
uint16_t min_hits)
: size_key(sz_key), size_value(sz_value), key_only(only_keys),
check_rate(check_rate), min_hits(min_hits) {
min_hits(min_hits) {
key_map.resize(COMMON_NGRAM_HASH_MAP_SIZE); // 2^18 hash entries, 0 entries if key_map shouldn't be used
}

55
common/speculative.cpp
View File

@ -113,13 +113,14 @@ static bool common_speculative_are_compatible(
struct common_speculative_state {
const enum common_speculative_type type;
// TODO: rename to n_call_draft, n_gen_drafts, n_acc_drafts, n_gen_tokens, n_acc_tokens
// TODO: add n_call_begin, n_call_accept
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.
size_t n_call_begin = 0; // number of times this implementation was called for refresh.
size_t n_call_draft = 0; // number of times this implementation was called for generation.
size_t n_call_accept = 0; // number of times this implementation was called for accumulation.
size_t n_gen_drafts = 0; // number of times a draft or part was generated by this implementation.
size_t n_acc_drafts = 0; // number of times a draft or part was accepted by the target model.
size_t n_gen_tokens = 0; // number of tokens generated by this implementation.
size_t n_acc_tokens = 0; // number of tokens accepted by the target model.
// TODO: track performance of most recent calls
const bool gen_perf = true; // whether to generate performance stats.
@ -465,8 +466,6 @@ struct common_speculative_state_eagle3 : public common_speculative_state {
struct common_speculative_state_ngram_simple : public common_speculative_state {
common_ngram_simple_config config;
uint16_t check_id = 0; // used to control the frequency of generating drafts
common_speculative_state_ngram_simple(
enum common_speculative_type type,
common_ngram_simple_config config)
@ -481,11 +480,6 @@ struct common_speculative_state_ngram_simple : public common_speculative_state {
const llama_tokens & prompt_tgt,
llama_token id_last,
llama_tokens & result) override {
++check_id;
if (check_id < config.check_rate) {
return;
}
check_id = 0;
result = common_ngram_simple_draft(config, prompt_tgt, id_last);
GGML_UNUSED(params);
@ -752,10 +746,9 @@ static common_ngram_map get_common_ngram_map(const common_speculative_config & c
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);
return common_ngram_map(size_key, size_value, key_only, min_hits);
}
static common_speculative_state_ngram_cache create_state_ngram_cache(
@ -931,12 +924,10 @@ common_speculative * common_speculative_init(
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
/* .size_mgram = */ mgram_size_value
};
auto state = std::make_unique<common_speculative_state_ngram_simple>(
/* .type = */ config.type,
@ -997,6 +988,7 @@ void common_speculative_begin(common_speculative * spec, const llama_tokens & pr
for (auto & impl : spec->impls) {
common_time_meas tm(impl->t_begin_us, !impl->gen_perf);
impl->begin(prompt);
impl->n_call_begin++;
}
}
@ -1013,17 +1005,17 @@ llama_tokens common_speculative_draft(
{
common_time_meas tm(impl->t_draft_us, !impl->gen_perf);
impl->draft(params, prompt_tgt, id_last, result);
impl->drafts_call_count++;
impl->n_call_draft++;
}
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());
impl.get()->n_call_draft, result.size());
spec->curr_impl = impl.get(); // set current implementation for stats
impl->drafts_generated_count++;
impl->drafts_generated_tokens += result.size();
impl->n_gen_drafts++;
impl->n_gen_tokens += result.size();
break; // We have a draft, so break out of the loop and return it.
}
@ -1044,11 +1036,12 @@ void common_speculative_accept(common_speculative * spec, uint16_t n_accepted) {
{
common_time_meas tm(impl->t_accept_us, !impl->gen_perf);
if (n_accepted > 0) {
impl->drafts_accepted_count++;
impl->drafts_accepted_tokens += n_accepted;
impl->n_acc_drafts++;
impl->n_acc_tokens += n_accepted;
}
impl->accept(n_accepted);
impl->n_call_accept++;
}
}
@ -1069,13 +1062,13 @@ void common_speculative_print_stats(const common_speculative * spec) {
str_perf = "";
}
LOG_INF("statistics %s: #calls = %zu, #gen drafts = %zu, #acc drafts = %zu, #gen tokens = %zu, #acc tokens = %zu%s\n",
LOG_INF("statistics %s: #calls(b,g,a) = %zu %zu %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,
impl->n_call_begin, impl->n_call_draft, impl->n_call_accept,
impl->n_gen_drafts,
impl->n_acc_drafts,
impl->n_gen_tokens,
impl->n_acc_tokens,
str_perf.c_str());
}
}

162
convert_hf_to_gguf.py
View File

@ -1261,6 +1261,9 @@ class TextModel(ModelBase):
if chkhsh == "6c81ce329e0802883b22eabab0d3fa48357337ef1ecb45443828bf1f6254833f":
# ref: https://huggingface.co/LGAI-EXAONE/K-EXAONE-236B-A23B
res = "exaone-moe"
if chkhsh == "d30d75d9059f1aa2c19359de71047b3ae408c70875e8a3ccf8c5fba56c9d8af4":
# ref: https://huggingface.co/Qwen/Qwen3.5-9B-Instruct
res = "qwen35"
if res is None:
logger.warning("\n")
@ -4109,37 +4112,29 @@ class Qwen2MoeModel(TextModel):
# Expected GGML ne: {n_embd, n_ff_exp, n_expert} for gate/up, {n_ff_exp, n_embd, n_expert} for down
if name.endswith("mlp.experts.down_proj") or name.endswith("mlp.experts.down_proj.weight"):
mapped = f"{name}.weight" if not name.endswith(".weight") else name
# Input: (n_expert=128, n_ff_exp=768, n_embd=2048)
# Want GGML ne: {n_ff_exp, n_embd, n_expert} = {768, 2048, 128}
# Need PyTorch: (128, 2048, 768) [reversed of GGML]
# So: permute(0, 2, 1): (128, 768, 2048) -> (128, 2048, 768)
permuted = data_torch.permute(0, 2, 1).contiguous()
yield from super().modify_tensors(permuted, mapped, bid)
# HF: [n_expert, n_embd, n_ff] -> GGML: {n_ff, n_embd, n_expert}
yield from super().modify_tensors(data_torch, mapped, bid)
return
if name.endswith("mlp.experts.gate_up_proj") or name.endswith("mlp.experts.gate_up_proj.weight"):
if data_torch.ndim < 3 or data_torch.shape[-1] % 2 != 0:
if data_torch.ndim < 3 or data_torch.shape[-2] % 2 != 0:
raise ValueError(f"Unexpected gate_up_proj shape for {name}: {tuple(data_torch.shape)}")
split_dim = data_torch.shape[-1] // 2
gate = data_torch[..., :split_dim].contiguous()
up = data_torch[..., split_dim:].contiguous()
# Input gate/up: (n_expert=128, n_embd=2048, n_ff_exp=768)
# Want GGML ne: {n_embd, n_ff_exp, n_expert} = {2048, 768, 128}
# Need PyTorch: (128, 768, 2048) [reversed of GGML]
# So: permute(0, 2, 1): (128, 2048, 768) -> (128, 768, 2048)
base_name = name.removesuffix(".weight")
base = base_name.rsplit('.', 1)[0]
mapped_gate = f"{base}.gate_proj.weight"
mapped_up = f"{base}.up_proj.weight"
perm_gate = gate.permute(0, 2, 1).contiguous()
perm_up = up.permute(0, 2, 1).contiguous()
yield from super().modify_tensors(perm_gate, mapped_gate, bid)
yield from super().modify_tensors(perm_up, mapped_up, bid)
# HF: [n_expert, 2*n_ff, n_embd] -> split on dim=-2
n_ff = data_torch.shape[-2] // 2
gate = data_torch[..., :n_ff, :].contiguous()
up = data_torch[..., n_ff:, :].contiguous()
# gate/up: [n_expert, n_ff, n_embd] -> GGML: {n_embd, n_ff, n_expert}
base_name = name.removesuffix(".weight").removesuffix(".gate_up_proj")
mapped_gate = f"{base_name}.gate_proj.weight"
mapped_up = f"{base_name}.up_proj.weight"
yield from super().modify_tensors(gate, mapped_gate, bid)
yield from super().modify_tensors(up, mapped_up, bid)
return
if name.startswith("mlp") or name.startswith("vision_model") or name.startswith("model.vision_tower") or name.startswith("model.multi_modal_projector") or name.startswith("model.visual"):
# skip visual tensors
return
if name.find("experts") != -1:
n_experts = self.hparams["num_experts"]
assert bid is not None
@ -4295,6 +4290,7 @@ class Qwen3NextModel(Qwen2MoeModel):
self.gguf_writer.add_ssm_group_count(self.hparams["linear_num_key_heads"])
self.gguf_writer.add_ssm_time_step_rank(self.hparams["linear_num_value_heads"])
self.gguf_writer.add_ssm_inner_size(self.hparams["linear_value_head_dim"] * self.hparams["linear_num_value_heads"])
self.gguf_writer.add_full_attention_interval(self.hparams.get("full_attention_interval", 4))
if (rope_dim := self.hparams.get("head_dim")) is None:
rope_dim = self.hparams["hidden_size"] // self.hparams["num_attention_heads"]
self.gguf_writer.add_rope_dimension_count(int(rope_dim * self.hparams.get("partial_rotary_factor", 0.25)))
@ -4359,7 +4355,7 @@ class RND1Model(Qwen2MoeModel):
self.gguf_writer.add_mask_token_id(mask_token_id)
@ModelBase.register("Qwen3VLForConditionalGeneration", "Qwen3VLMoeForConditionalGeneration")
@ModelBase.register("Qwen3VLForConditionalGeneration", "Qwen3VLMoeForConditionalGeneration", "Qwen3_5ForConditionalGeneration", "Qwen3_5MoeForConditionalGeneration")
class Qwen3VLVisionModel(MmprojModel):
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
@ -4405,6 +4401,10 @@ class Qwen3VLVisionModel(MmprojModel):
if name.startswith("model.language_model.") or name.startswith("lm_head."):
return
# Skip MTP tensors
if name.startswith("mtp."):
return
if name.startswith("model.visual."):
name = name.replace("model.visual.", "visual.", 1)
@ -4535,9 +4535,125 @@ class Qwen3VLMoeTextModel(Qwen3MoeModel):
if name.startswith("model.visual."):
return
# Qwen3VL has transposed packed tensors, so we treat it differently from general Qwen2MoE packed tensors
if name.endswith("mlp.experts.down_proj") or name.endswith("mlp.experts.down_proj.weight"):
name = name.replace("language_model.", "")
mapped = f"{name}.weight" if not name.endswith(".weight") else name
permuted = data_torch.permute(0, 2, 1).contiguous()
yield from ModelBase.modify_tensors(self, permuted, mapped, bid)
return
if name.endswith("mlp.experts.gate_up_proj") or name.endswith("mlp.experts.gate_up_proj.weight"):
name = name.replace("language_model.", "")
if data_torch.ndim < 3 or data_torch.shape[-1] % 2 != 0:
raise ValueError(f"Unexpected gate_up_proj shape for {name}: {tuple(data_torch.shape)}")
split_dim = data_torch.shape[-1] // 2
gate = data_torch[..., :split_dim].contiguous()
up = data_torch[..., split_dim:].contiguous()
# Input gate/up: (n_expert=128, n_embd=2048, n_ff_exp=768)
# Want GGML ne: {n_embd, n_ff_exp, n_expert} = {2048, 768, 128}
# Need PyTorch: (128, 768, 2048) [reversed of GGML]
# So: permute(0, 2, 1): (128, 2048, 768) -> (128, 768, 2048)
base_name = name.removesuffix(".weight")
base = base_name.rsplit('.', 1)[0]
mapped_gate = f"{base}.gate_proj.weight"
mapped_up = f"{base}.up_proj.weight"
perm_gate = gate.permute(0, 2, 1).contiguous()
perm_up = up.permute(0, 2, 1).contiguous()
yield from ModelBase.modify_tensors(self, perm_gate, mapped_gate, bid)
yield from ModelBase.modify_tensors(self, perm_up, mapped_up, bid)
return
yield from super().modify_tensors(data_torch, name, bid)
class _LinearAttentionVReorderBase(Qwen3NextModel):
model_arch = gguf.MODEL_ARCH.QWEN3NEXT # overridden by subclasses
"""reorders V heads from grouped to tiled order for ggml broadcast
see https://github.com/ggml-org/llama.cpp/pull/19468#discussion_r2786394306
Linear attention may has num_k_heads < num_v_heads. The HF weights store
V heads grouped by K head: [G0_v0..v{r-1}, G1_v0..v{r-1}, ...].
ggml binary ops use tiled broadcast: [K0, K1, ..., K0, K1, ...].
We reorder V heads to tiled order so ggml_repeat can replace the expensive
interleaved repeat: [G0_v0, G1_v0, ..., G0_v1, G1_v1, ...].
"""
@staticmethod
def _reorder_v_heads(tensor: Tensor, dim: int, num_k_heads: int, num_v_per_k: int, head_dim: int) -> Tensor:
"""Reorder V heads from grouped (by K head) to tiled order along the given dimension."""
shape = list(tensor.shape)
if dim < 0:
dim += len(shape)
new_shape = shape[:dim] + [num_k_heads, num_v_per_k, head_dim] + shape[dim + 1:]
tensor = tensor.reshape(*new_shape)
perm = list(range(len(new_shape)))
perm[dim], perm[dim + 1] = perm[dim + 1], perm[dim]
return tensor.permute(*perm).contiguous().reshape(*shape)
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
num_k_heads = self.hparams.get("linear_num_key_heads", 0)
num_v_heads = self.hparams.get("linear_num_value_heads", 0)
if num_k_heads > 0 and num_v_heads > 0 and num_k_heads != num_v_heads and "linear_attn." in name:
head_k_dim = self.hparams["linear_key_head_dim"]
head_v_dim = self.hparams["linear_value_head_dim"]
num_v_per_k = num_v_heads // num_k_heads
if ".in_proj_qkv." in name:
# QKV weight: reorder only the V rows
q_dim = head_k_dim * num_k_heads
k_dim = head_k_dim * num_k_heads
q = data_torch[:q_dim]
k = data_torch[q_dim:q_dim + k_dim]
v = data_torch[q_dim + k_dim:]
v = self._reorder_v_heads(v, 0, num_k_heads, num_v_per_k, head_v_dim)
data_torch = torch.cat([q, k, v], dim=0)
elif ".in_proj_z." in name:
# Z gate weight: reorder rows (num_v_heads * head_v_dim)
data_torch = self._reorder_v_heads(data_torch, 0, num_k_heads, num_v_per_k, head_v_dim)
elif ".in_proj_b." in name or ".in_proj_a." in name:
# Beta/Alpha weight: reorder rows (num_v_heads, head_dim=1)
data_torch = self._reorder_v_heads(data_torch, 0, num_k_heads, num_v_per_k, 1)
elif ".A_log" in name or ".dt_bias" in name or ".dt_proj" in name:
# A_log / dt_bias: 1D parameters with num_v_heads elements
if data_torch.ndim == 1:
data_torch = self._reorder_v_heads(
data_torch.unsqueeze(-1), 0, num_k_heads, num_v_per_k, 1
).squeeze(-1)
else:
data_torch = self._reorder_v_heads(data_torch, -1, num_k_heads, num_v_per_k, 1)
elif ".conv1d" in name:
# Conv1d kernel: reorder only the V channel portion
data = data_torch.squeeze()
qk_channels = head_k_dim * num_k_heads * 2
qk_part = data[:qk_channels]
v_part = data[qk_channels:]
v_part = self._reorder_v_heads(v_part, 0, num_k_heads, num_v_per_k, head_v_dim)
data_torch = torch.cat([qk_part, v_part], dim=0)
elif ".out_proj." in name:
# Out projection weight: reorder columns (input dimension)
data_torch = self._reorder_v_heads(data_torch, 1, num_k_heads, num_v_per_k, head_v_dim)
yield from super().modify_tensors(data_torch, name, bid)
@ModelBase.register("Qwen3_5ForConditionalGeneration")
class Qwen3_5TextModel(_LinearAttentionVReorderBase):
model_arch = gguf.MODEL_ARCH.QWEN35
@ModelBase.register("Qwen3_5MoeForConditionalGeneration")
class Qwen3_5MoeTextModel(_LinearAttentionVReorderBase):
model_arch = gguf.MODEL_ARCH.QWEN35MOE
@ModelBase.register("GPT2LMHeadModel")
class GPT2Model(TextModel):
model_arch = gguf.MODEL_ARCH.GPT2

1
convert_hf_to_gguf_update.py
View File

@ -148,6 +148,7 @@ models = [
{"name": "youtu", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tencent/Youtu-LLM-2B", },
{"name": "solar-open", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/upstage/Solar-Open-100B", },
{"name": "exaone-moe", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LGAI-EXAONE/K-EXAONE-236B-A23B", },
{"name": "qwen35", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/Qwen/Qwen3.5-9B-Instruct", }
]
# some models are known to be broken upstream, so we will skip them as exceptions

180
docs/backend/VirtGPU.md Normal file
View File

@ -0,0 +1,180 @@
# GGML-VirtGPU Backend
The GGML-VirtGPU backend enables GGML applications to run machine
learning computations on host hardware while the application itself
runs inside a virtual machine. It uses host-guest shared memory to
efficiently share data buffers between the two sides.
This backend relies on the virtio-gpu, and VirglRenderer API Remoting
(APIR) component. The backend is split into two libraries:
- a GGML implementation (the "remoting frontend"), running in the
guest and interacting with the virtgpu device
- a VirglRenderer APIR compatible library (the "remoting backend"),
running in the host and interacting with Virglrenderer and an actual
GGML device backend.
## OS support
| OS | Status | Backend | CI testing | Notes
| -------- | ----------------- | ----------- | ----------- | -----
| MacOS 14 | Supported | ggml-metal | X | Working when compiled on MacOS 14
| MacOS 15 | Supported | ggml-metal | X | Working when compiled on MacOS 14 or MacOS 15
| MacOS 26 | Not tested | | |
| Linux | Under development | ggml-vulkan | not working | Working locally, CI running into deadlocks
## Architecture Overview
The GGML-VirtGPU backend consists of three main components:
```mermaid
graph TD
%% Nodes
subgraph GuestVM ["Guest VM - Frontend"]
App([GGML Application<br/>llama.cpp, etc.])
direction TB
Interface[GGML Backend Interface]
Comm["GGML-VirtGPU<br/>(hypercalls + shared mem)"]
App --> Interface
Interface --> Comm
end
API[virtio-gpu / virglrenderer API]
subgraph HostSystem [Host System - Backend]
direction TB
Dispatcher[GGML-VirtGPU-Backend]
BackendLib[GGML Backend library<br/>Metal / Vulkan / CPU / ...]
Dispatcher --> BackendLib
end
%% Connections
Comm --> API
API --> HostSystem
```
### Key Components
1. **Guest-side Frontend** (`ggml-virtgpu/`): Implements the GGML backend interface and forwards operations to the host
2. **Host-side Backend** (`ggml-virtgpu/backend/`): Receives forwarded operations and executes them on actual hardware backends
3. **Communication Layer**: Uses virtio-gpu hypercalls and shared memory for efficient data transfer
## Features
- **Dynamic backend loading** on the host side (CPU, CUDA, Metal, etc.)
- **Zero-copy data transfer** via host-guest shared memory pages
## Communication Protocol
### Hypercalls and Shared Memory
The backend uses two primary communication mechanisms:
1. **Hypercalls (`DRM_IOCTL_VIRTGPU_EXECBUFFER`)**: Trigger remote execution from guest to host
2. **Shared Memory Pages**: Zero-copy data transfer for tensors and parameters
#### Shared Memory Layout
Each connection uses two shared memory buffers:
- **Data Buffer** (24 MiB): For command/response data and tensor transfers
- **Reply Buffer** (16 KiB): For command replies and status information
- **Data Buffers**: Dynamically allocated host-guest shared buffers
served as GGML buffers.
### APIR Protocol
The Virglrender API Remoting protocol defines three command types:
- `HANDSHAKE`: Protocol version negotiation and capability discovery
- `LOADLIBRARY`: Dynamic loading of backend libraries on the host
- `FORWARD`: API function call forwarding
### Binary Serialization
Commands and data are serialized using a custom binary protocol with:
- Fixed-size encoding for basic types
- Variable-length arrays with size prefixes
- Buffer bounds checking
- Error recovery mechanisms
## Supported Operations
### Device Operations
- Device enumeration and capability queries
- Memory information (total/free)
- Backend type detection
### Buffer Operations
- Buffer allocation and deallocation
- Tensor data transfer (host ↔ guest)
- Memory copying and clearing
### Computation Operations
- Graph execution forwarding
## Build Requirements
### Guest-side Dependencies
- `libdrm` for DRM/virtio-gpu communication
- C++20 compatible compiler
- CMake 3.14+
### Host-side Dependencies
- virglrenderer with APIR support (pending upstream review)
- Target backend libraries (libggml-metal, libggml-vulkan, etc.)
## Configuration
### Environment Variables
- `GGML_VIRTGPU_BACKEND_LIBRARY`: Path to the host-side backend library
- `GGML_VIRTGPU_DEBUG`: Enable debug logging
### Build Options
- `GGML_VIRTGPU`: Enable the VirtGPU backend (`ON` or `OFF`, default: `OFF`)
- `GGML_VIRTGPU_BACKEND`: Build the host-side backend component (`ON`, `OFF` or `ONLY`, default: `OFF`)
### System Requirements
- VM with virtio-gpu support
- VirglRenderer with APIR patches
- Compatible backend libraries on host
## Limitations
- **VM-specific**: Only works in virtual machines with virtio-gpu support
- **Host dependency**: Requires properly configured host-side backend
- **Latency**: Small overhead from VM escaping for each operation
* This work is pending upstream changes in the VirglRenderer
project.
* The backend can be tested with Virglrenderer compiled from source
using this PR:
https://gitlab.freedesktop.org/virgl/virglrenderer/-/merge_requests/1590
* This work is pending changes in the VMM/hypervisor running the
virtual machine, which need to know how to route the newly
introduced APIR capset.
* The environment variable `VIRGL_ROUTE_VENUS_TO_APIR=1` allows
using the Venus capset, until the relevant hypervisors have been
patched. However, setting this flag breaks the Vulkan/Venus normal
behavior.
* The environment variable `GGML_REMOTING_USE_APIR_CAPSET` tells the
`ggml-virtgpu` backend to use the APIR capset. This will become
the default when the relevant hypervisors have been patched.
* This work focused on improving the performance of llama.cpp running
on MacOS containers, and is mainly tested on this platform. The
linux support (via `krun`) is in progress.
## See Also
- [Development and Testing](VirtGPU/development.md)
- [Backend configuration](VirtGPU/configuration.md)

174
docs/backend/VirtGPU/configuration.md Normal file
View File

@ -0,0 +1,174 @@
# GGML-VirtGPU Backend Configuration
This document describes the environment variables used by the ggml-virtgpu backend system, covering both the frontend (guest-side) and backend (host-side) components.
## Environment Variables Overview
The ggml-virtgpu backend uses environment variables for configuration across three main components:
- **Frontend (Guest)**: GGML applications running in VMs
- **Hypervisor**: Virglrenderer/APIR system
- **Backend (Host)**: Host-side GGML backend integration
## Frontend (Guest-side) Configuration
### GGML_REMOTING_USE_APIR_CAPSET
- **Location**: `ggml/src/ggml-virtgpu/virtgpu.cpp`
- **Type**: Boolean flag (presence-based)
- **Purpose**: Controls which virtio-gpu capability set to use for communication
- **Values**:
- Set (any value): Use the APIR capset (long-term setup)
- Unset: Use the Venus capset (easier for testing with an unmodified hypervisor)
- **Default**: Unset (Venus capset)
- **Usage**:
```bash
export GGML_REMOTING_USE_APIR_CAPSET=1 # Use APIR capset
# or leave unset for Venus capset
```
## Hypervisor (Virglrenderer/APIR) Configuration
These environment variables are used during the transition phase for
running with an unmodified hypervisor (not supporting the
VirglRenderer APIR component). They will be removed in the future, and
the hypervisor will instead configure VirglRenderer with the APIR
_Configuration Key_.
### VIRGL_APIR_BACKEND_LIBRARY
- **Location**: `virglrenderer/src/apir/apir-context.c`
- **Configuration Key**: `apir.load_library.path`
- **Type**: File path string
- **Purpose**: Path to the APIR backend library that virglrenderer should dynamically load
- **Required**: Yes
- **Example**:
```bash
export VIRGL_APIR_BACKEND_LIBRARY="/path/to/libggml-remotingbackend.so"
```
### VIRGL_ROUTE_VENUS_TO_APIR
- **Location**: `virglrenderer/src/apir/apir-renderer.h`
- **Type**: Boolean flag (presence-based)
- **Purpose**: Temporary workaround to route Venus capset calls to APIR during hypervisor transition period
- **Status**: will be removed once hypervisors support APIR natively
- **Warning**: Breaks normal Vulkan/Venus functionality
- **Usage**:
```bash
export VIRGL_ROUTE_VENUS_TO_APIR=1 # For testing with an unmodified hypervisor
```
### VIRGL_APIR_LOG_TO_FILE
- **Location**: `virglrenderer/src/apir/apir-renderer.c`
- **Environment Variable**: `VIRGL_APIR_LOG_TO_FILE`
- **Type**: File path string
- **Purpose**: Enable debug logging from the VirglRenderer APIR component to specified file
- **Required**: No (optional debugging)
- **Default**: Logging to `stderr`
- **Usage**:
```bash
export VIRGL_APIR_LOG_TO_FILE="/tmp/apir-debug.log"
```
## Backend (Host-side) Configuration
These environment variables are used during the transition phase for
running with an unmodified hypervisor (not supporting the
VirglRenderer APIR component). They will be removed in the future, and
the hypervisor will instead configure VirglRenderer with the APIR
_Configuration Key_.
### APIR_LLAMA_CPP_GGML_LIBRARY_PATH
- **Location**: `ggml/src/ggml-virtgpu/backend/backend.cpp`
- **Environment Variable**: `APIR_LLAMA_CPP_GGML_LIBRARY_PATH`
- **Configuration Key**: `ggml.library.path`
- **Type**: File path string
- **Purpose**: Path to the actual GGML backend library (Metal, CUDA, Vulkan, etc.)
- **Required**: **Yes** - backend initialization fails without this
- **Examples**:
```bash
# macOS with Metal backend
export APIR_LLAMA_CPP_GGML_LIBRARY_PATH="/opt/llama.cpp/lib/libggml-metal.dylib"
# Linux with CUDA backend
export APIR_LLAMA_CPP_GGML_LIBRARY_PATH="/opt/llama.cpp/lib/libggml-cuda.so"
# macOS or Linux with Vulkan backend
export APIR_LLAMA_CPP_GGML_LIBRARY_PATH="/opt/llama.cpp/lib/libggml-vulkan.so"
```
### APIR_LLAMA_CPP_GGML_LIBRARY_REG
- **Location**: `ggml/src/ggml-virtgpu/backend/backend.cpp`
- **Environment Variable**: `APIR_LLAMA_CPP_GGML_LIBRARY_REG`
- **Configuration Key**: `ggml.library.reg`
- **Type**: Function symbol name string
- **Purpose**: Name of the backend registration function to call after loading the library
- **Required**: No (defaults to `ggml_backend_init`)
- **Default**: `ggml_backend_init`
- **Examples**:
```bash
# Metal backend
export APIR_LLAMA_CPP_GGML_LIBRARY_REG="ggml_backend_metal_reg"
# CUDA backend
export APIR_LLAMA_CPP_GGML_LIBRARY_REG="ggml_backend_cuda_reg"
# Vulkan backend
export APIR_LLAMA_CPP_GGML_LIBRARY_REG="ggml_backend_vulkan_reg"
# Generic fallback (default)
# export APIR_LLAMA_CPP_GGML_LIBRARY_REG="ggml_backend_init"
```
### APIR_LLAMA_CPP_LOG_TO_FILE
- **Location**: `ggml/src/ggml-virtgpu/backend/backend.cpp:62`
- **Environment Variable**: `APIR_LLAMA_CPP_LOG_TO_FILE`
- **Type**: File path string
- **Purpose**: Enable debug logging from the GGML backend to specified file
- **Required**: No (optional debugging)
- **Usage**:
```bash
export APIR_LLAMA_CPP_LOG_TO_FILE="/tmp/ggml-backend-debug.log"
```
## Configuration Flow
The configuration system works as follows:
1. **Hypervisor Setup**: Virglrenderer loads the APIR backend library specified by `VIRGL_APIR_BACKEND_LIBRARY`
2. **Context Creation**: When an APIR context is created, it populates a configuration table with environment variables:
- `apir.load_library.path` ← `VIRGL_APIR_BACKEND_LIBRARY`
- `ggml.library.path` ← `APIR_LLAMA_CPP_GGML_LIBRARY_PATH`
- `ggml.library.reg` ← `APIR_LLAMA_CPP_GGML_LIBRARY_REG`
- this step will eventually be performed by the hypervisor itself, with command-line arguments instead of environment variables.
3. **Backend Initialization**: The backend queries the configuration via callbacks:
- `virgl_cbs->get_config(ctx_id, "ggml.library.path")` returns the library path
- `virgl_cbs->get_config(ctx_id, "ggml.library.reg")` returns the registration function
4. **Library Loading**: The backend dynamically loads and initializes the specified GGML library
## Error Messages
Common error scenarios and their messages:
- **Missing library path**: `"cannot open the GGML library: env var 'APIR_LLAMA_CPP_GGML_LIBRARY_PATH' not defined"`
- **Missing registration function**: `"cannot register the GGML library: env var 'APIR_LLAMA_CPP_GGML_LIBRARY_REG' not defined"`
## Example Complete Configuration
Here's an example configuration for a macOS host with Metal backend:
```bash
# Hypervisor environment
export VIRGL_APIR_BACKEND_LIBRARY="/opt/llama.cpp/lib/libggml-virtgpu-backend.dylib"
# Backend configuration
export APIR_LLAMA_CPP_GGML_LIBRARY_PATH="/opt/llama.cpp/lib/libggml-metal.dylib"
export APIR_LLAMA_CPP_GGML_LIBRARY_REG="ggml_backend_metal_reg"
# Optional logging
export VIRGL_APIR_LOG_TO_FILE="/tmp/apir.log"
export APIR_LLAMA_CPP_LOG_TO_FILE="/tmp/ggml.log"
# Guest configuration
export GGML_REMOTING_USE_APIR_CAPSET=1
```

220
docs/backend/VirtGPU/development.md Normal file
View File

@ -0,0 +1,220 @@
# Development and Testing
## Development
### Code Generation
The backend uses code generation from YAML configuration:
```bash
# Regenerate protocol code
cd ggml-virtgpu/
python regenerate_remoting.py
```
### Adding New Operations
1. Add function definition to `ggmlremoting_functions.yaml`
2. Regenerate code with `regenerate_remoting.py`
3. Implement guest-side forwarding in `virtgpu-forward-*.cpp`
4. Implement host-side handling in `backend-dispatched-*.cpp`
## Testing
This document provides instructions for building and testing the GGML-VirtGPU backend on macOS with containers.
### Prerequisites
The testing setup requires:
- macOS host system
- Container runtime with `libkrun` provider (podman machine)
- Access to development patchset for VirglRenderer
### Required Patchsets
The backend requires patches that are currently under review:
- **Virglrenderer APIR upstream PR**: https://gitlab.freedesktop.org/virgl/virglrenderer/-/merge_requests/1590 (for reference)
- **MacOS Virglrenderer (for krunkit)**: https://gitlab.freedesktop.org/kpouget/virglrenderer/-/tree/main-macos
- **Linux Virglrenderer (for krun)**: https://gitlab.freedesktop.org/kpouget/virglrenderer/-/tree/main-linux
### Build Instructions
#### 1. Build ggml-virtgpu-backend (Host-side, macOS)
```bash
# Build the backend that runs natively on macOS
mkdir llama.cpp
cd llama.cpp
git clone https://github.com/ggml-org/llama.cpp.git src
cd src
LLAMA_MAC_BUILD=$PWD/build/ggml-virtgpu-backend
cmake -S . -B $LLAMA_MAC_BUILD \
-DGGML_NATIVE=OFF \
-DLLAMA_CURL=ON \
-DGGML_REMOTINGBACKEND=ONLY \
-DGGML_METAL=ON
TARGETS="ggml-metal"
cmake --build $LLAMA_MAC_BUILD --parallel 8 --target $TARGETS
# Build additional tools for native benchmarking
EXTRA_TARGETS="llama-run llama-bench"
cmake --build $LLAMA_MAC_BUILD --parallel 8 --target $EXTRA_TARGETS
```
#### 2. Build virglrenderer (Host-side, macOS)
```bash
# Build virglrenderer with APIR support
mkdir virglrenderer
git clone https://gitlab.freedesktop.org/kpouget/virglrenderer -b main-macos src
cd src
VIRGL_BUILD_DIR=$PWD/build
# -Dvenus=true and VIRGL_ROUTE_VENUS_TO_APIR=1 route the APIR requests via the Venus backend, for easier testing without a patched hypervisor
meson setup $VIRGL_BUILD_DIR \
-Dvenus=true \
-Dapir=true
ninja -C $VIRGL_BUILD_DIR
```
#### 3. Build ggml-virtgpu (Guest-side, Linux)
Option A: Build from a script:
```bash
# Inside a Linux container
mkdir llama.cpp
git clone https://github.com/ggml-org/llama.cpp.git src
cd src
LLAMA_LINUX_BUILD=$PWD//build-virtgpu
cmake -S . -B $LLAMA_LINUX_BUILD \
-DGGML_VIRTGPU=ON
ninja -C $LLAMA_LINUX_BUILD
```
Option B: Build container image with frontend:
```bash
cat << EOF > remoting.containerfile
FROM quay.io/fedora/fedora:43
USER 0
WORKDIR /app/remoting
ARG LLAMA_CPP_REPO="https://github.com/ggml-org/llama.cpp.git"
ARG LLAMA_CPP_VERSION="master"
ARG LLAMA_CPP_CMAKE_FLAGS="-DGGML_VIRTGPU=ON"
ARG LLAMA_CPP_CMAKE_BUILD_FLAGS="--parallel 4"
RUN dnf install -y git cmake gcc gcc-c++ libcurl-devel libdrm-devel
RUN git clone "\${LLAMA_CPP_REPO}" src \\
&& git -C src fetch origin \${LLAMA_CPP_VERSION} \\
&& git -C src reset --hard FETCH_HEAD
RUN mkdir -p build \\
&& cd src \\
&& set -o pipefail \\
&& cmake -S . -B ../build \${LLAMA_CPP_CMAKE_FLAGS} \\
&& cmake --build ../build/ \${LLAMA_CPP_CMAKE_BUILD_FLAGS}
ENTRYPOINT ["/app/remoting/src/build/bin/llama-server"]
EOF
mkdir -p empty_dir
podman build -f remoting.containerfile ./empty_dir -t localhost/llama-cpp.virtgpu
```
### Environment Setup
#### Set krunkit Environment Variables
```bash
# Define the base directories (adapt these paths to your system)
VIRGL_BUILD_DIR=$HOME/remoting/virglrenderer/build
LLAMA_MAC_BUILD=$HOME/remoting/llama.cpp/build-backend
# For krunkit to load the custom virglrenderer library
export DYLD_LIBRARY_PATH=$VIRGL_BUILD_DIR/src
# For Virglrenderer to load the ggml-remotingbackend library
export VIRGL_APIR_BACKEND_LIBRARY="$LLAMA_MAC_BUILD/bin/libggml-virtgpu-backend.dylib"
# For llama.cpp remotingbackend to load the ggml-metal backend
export APIR_LLAMA_CPP_GGML_LIBRARY_PATH="$LLAMA_MAC_BUILD/bin/libggml-metal.dylib"
export APIR_LLAMA_CPP_GGML_LIBRARY_REG=ggml_backend_metal_reg
```
#### Launch Container Environment
```bash
# Set container provider to libkrun
export CONTAINERS_MACHINE_PROVIDER=libkrun
podman machine start
```
#### Verify Environment
Confirm that krunkit is using the correct virglrenderer library:
```bash
lsof -c krunkit | grep virglrenderer
# Expected output:
# krunkit 50574 user txt REG 1,14 2273912 10849442 ($VIRGL_BUILD_DIR/src)/libvirglrenderer.1.dylib
```
### Running Tests
#### Launch Test Container
```bash
# Optional model caching
mkdir -p models
PODMAN_CACHE_ARGS="-v models:/models --user root:root --cgroupns host --security-opt label=disable -w /models"
podman run $PODMAN_CACHE_ARGS -it --rm --device /dev/dri localhost/llama-cpp.virtgpu
```
#### Test llama.cpp in Container
```bash
# Run performance benchmark
/app/remoting/build/bin/llama-bench -m ./llama3.2
```
Expected output (performance may vary):
```
| model | size | params | backend | ngl | test | t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | ------------: | -------------------: |
| llama 3B Q4_K - Medium | 1.87 GiB | 3.21 B | ggml-virtgpu | 99 | pp512 | 991.30 ± 0.66 |
| llama 3B Q4_K - Medium | 1.87 GiB | 3.21 B | ggml-virtgpu | 99 | tg128 | 85.71 ± 0.11 |
```
### Troubleshooting
#### SSH Environment Variable Issues
⚠️ **Warning**: Setting `DYLD_LIBRARY_PATH` from SSH doesn't work on macOS. Here is a workaround:
**Workaround 1: Replace system library**
```bash
VIRGL_BUILD_DIR=$HOME/remoting/virglrenderer/build # ⚠️ adapt to your system
BREW_VIRGL_DIR=/opt/homebrew/Cellar/virglrenderer/0.10.4d/lib
VIRGL_LIB=libvirglrenderer.1.dylib
cd $BREW_VIRGL_DIR
mv $VIRGL_LIB ${VIRGL_LIB}.orig
ln -s $VIRGL_BUILD_DIR/src/$VIRGL_LIB
```

13
docs/speculative.md
View File

@ -119,8 +119,6 @@ If a draft model is combined with a draftless decoding the draftless decoding ha
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)
```
@ -153,10 +151,6 @@ 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).
@ -175,7 +169,12 @@ draft acceptance rate = 0.70312 ( 90 accepted / 128 generated)
statistics ngram_mod: #calls = 810, #gen drafts = 15, #acc drafts = 15, #gen tokens = 960, #acc tokens = 730, dur(b,g,a) = 0.149, 0.347, 0.005 ms
```
- `#calls`: number of calls of this implementations
```
statistics ngram_map_k: #calls(b,g,a) = 6 1690 26, #gen drafts = 26, #acc drafts = 26, #gen tokens = 1248, #acc tokens = 968, dur(b,g,a) = 2.234, 1.427, 0.016 ms
```
- `#calls(b,g,a)`: number of calls of begin (new prompt), generation and accumulation 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)

5
ggml/src/ggml-backend-reg.cpp
View File

@ -471,9 +471,10 @@ static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent,
int best_score = 0;
fs::path best_path;
std::error_code ec;
for (const auto & search_path : search_paths) {
if (std::error_code ec; !fs::exists(search_path, ec)) {
if (!fs::exists(search_path, ec)) {
if (ec) {
GGML_LOG_DEBUG("%s: posix_stat(%s) failure, error-message: %s\n", __func__, path_str(search_path).c_str(), ec.message().c_str());
} else {
@ -483,7 +484,7 @@ static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent,
}
fs::directory_iterator dir_it(search_path, fs::directory_options::skip_permission_denied);
for (const auto & entry : dir_it) {
if (entry.is_regular_file()) {
if (entry.is_regular_file(ec)) {
auto filename = entry.path().filename();
auto ext = entry.path().extension();
if (filename.native().find(file_prefix) == 0 && ext == file_extension) {

291
ggml/src/ggml-cann/aclnn_ops.cpp
View File

@ -3286,130 +3286,223 @@ static void ggml_cann_mul_mat_id_fp(ggml_backend_cann_context & ctx, ggml_tensor
}
/**
* @brief Performs expert-specific matrix multiplication (MoE) with
* quantized precision using the CANN backend.
* @brief Performs quantized matrix multiplication for Mixture of Experts (MoE)
* models using the CANN backend.
*
* This function executes a matrix multiplication operation tailored for
* Mixture of Experts (MoE) models, where the input tensor is multiplied
* with expert-specific quantized weight matrices. It leverages the CANN
* backend to perform efficient low-precision computations and stores the
* quantized result in the destination tensor `dst`.
* This function implements MUL_MAT_ID operation for quantized weight matrices
* (Q4_0 and Q8_0 formats). It selects expert-specific weight matrices based on
* the provided expert indices, and computes matrix multiplication using CANN's
* WeightQuantBatchMatmulV2 operator.
*
* Quantization techniques reduce memory footprint and improve performance
* by using lower-bit representations (e.g., int8) instead of floating-point.
* This function is designed to work with such formats and may incorporate
* optimizations like identity-based fast paths or routing masks for sparse
* expert selection.
* The function performs the following steps:
* 1. Converts input/output tensors to F16 format if necessary
* 2. Uses IndexSelect to extract expert-specific weights and scales based on indices
* 3. Performs quantized matrix multiplication for each expert using WeightQuantBatchMatmulV2
* 4. Converts output back to the target type if needed
*
* @param ctx The context for executing CANN backend operations.
* @param dst The destination tensor where the quantized MoE multiplication result
* will be stored.
* Tensor shapes:
* - dst: [M, K, N, 1] - output tensor
* - src0: [D, M, A, 1] - quantized weight matrices (Q4_0 or Q8_0)
* - src1: [D, B, N, 1] - input activations (B = K for per-expert input, or B = 1 for broadcast)
* - ids: [K, N] - expert indices for routing
*
* @note This function assumes quantized data types and is designed for
* MoE architectures with potential sparse expert routing.
* @param ctx The CANN backend context for operation execution.
* @param dst The destination tensor where the multiplication result will be stored.
*
* @note Only Q4_0 and Q8_0 quantization formats are supported.
* @note The function handles automatic type conversion to/from F16 as needed by the hardware.
*/
static void ggml_cann_mul_mat_id_quant(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
// TODO: Use aclnnGroupedMatMul
//dst [M, K, N, 1]
ggml_tensor * src0 = dst->src[0]; //src0 [D, M, A, 1]
ggml_tensor * src1 = dst->src[1]; //src1 [D, B, N, 1], B = K or B = 1
ggml_tensor * ids = dst->src[2]; //ids [K, N]
// dst: [M, K, N, 1]
// src0: [D, M, A, 1] - quantized weights
// src1: [D, B, N, 1] - input activations, B = K or B = 1
// ids: [K, N] - expert indices
ggml_tensor * src0 = dst->src[0];
ggml_tensor * src1 = dst->src[1];
ggml_tensor * ids = dst->src[2];
GGML_TENSOR_BINARY_OP_LOCALS
GGML_ASSERT(src0->ne[3] == 1);
GGML_ASSERT(src1->ne[3] == 1);
GGML_ASSERT(dst->ne[3] == 1);
GGML_ASSERT(src1->ne[2] == ids->ne[1]);
// copy index from npu to cpu
int64_t n_as = ne02; // A
int64_t n_ids = ids->ne[0]; // K
const int64_t n_batches = ids->ne[1];
const int64_t n_select_experts = ids->ne[0];
const enum ggml_type type = src0->type;
std::vector<char> ids_host(ggml_nbytes(ids));
ACL_CHECK(aclrtMemcpyAsync(ids_host.data(), ggml_nbytes(ids), ids->data, ggml_nbytes(ids),
ACL_MEMCPY_DEVICE_TO_HOST, ctx.stream()));
ACL_CHECK(aclrtSynchronizeStream(ctx.stream()));
const int32_t group_size = QK8_0; // Both Q4_0 and Q8_0 use group size of 32
GGML_ASSERT(group_size == QK4_0);
char * src0_original = (char *) src0->data;
char * src1_original = (char *) src1->data;
char * dst_original = (char *) dst->data;
// Calculate element size for quantized weights
const float weight_elem_size =
(type == GGML_TYPE_Q4_0) ? 0.5f :
(type == GGML_TYPE_Q8_0) ? 1.0f :
(GGML_ABORT("MUL_MAT_ID only supports Q4_0 and Q8_0"), 0.0f);
ggml_tensor src0_row = *src0;
ggml_tensor src1_row = *src1;
ggml_tensor dst_row = *dst;
// Calculate scale offset in memory
const size_t weight_size = src0->ne[0] * src0->ne[1] * src0->ne[2] * weight_elem_size;
const size_t scale_elem_size = sizeof(uint16_t);
char * scale_data = (char *) src0->data + weight_size;
const enum ggml_type type = dst->src[0]->type;
float weight_elem_size;
if (type == GGML_TYPE_Q4_0) {
weight_elem_size = float(sizeof(uint8_t)) / 2;
} else if (type == GGML_TYPE_Q8_0) {
weight_elem_size = float(sizeof(uint8_t));
} else {
GGML_ABORT("MUL_MAT_ID only support quant type Q4_0 and Q8_0 ");
}
// Allocate buffers for selected expert weights and scales
const size_t selected_weight_size = src0->ne[0] * src0->ne[1] * n_select_experts * weight_elem_size;
ggml_cann_pool_alloc selected_weight_alloc(ctx.pool(), selected_weight_size);
void * selected_weight_buffer = selected_weight_alloc.get();
// src0_row [D, M, 1, 1] weight without permute
src0_row.ne[2] = 1;
src0_row.ne[3] = 1;
src0_row.nb[0] = weight_elem_size;
src0_row.nb[1] = weight_elem_size * ne00;
src0_row.nb[2] = weight_elem_size * ne00;
src0_row.nb[3] = weight_elem_size * ne00;
size_t weight_stride = ne00 * ne01 * weight_elem_size;
size_t weight_size = weight_stride * ne02 * ne03;
const size_t selected_scale_size = (src0->ne[0] / group_size) * src0->ne[1] * n_select_experts * scale_elem_size;
ggml_cann_pool_alloc selected_scale_alloc(ctx.pool(), selected_scale_size);
void * selected_scale_buffer = selected_scale_alloc.get();
// scale [D, M, 1, 1] -> scale && permute
size_t scale_elem_size = sizeof(uint16_t);
size_t scale_stride = src0->ne[1] * src0->ne[0] / QK8_0 * scale_elem_size;
// Helper lambda to allocate and cast tensor to F16 if needed
constexpr size_t f16_elem_size = sizeof(uint16_t);
auto prepare_f16_buffer = [&](ggml_tensor * tensor, ggml_cann_pool_alloc & allocator,
bool need_cast = false) -> void * {
if (tensor->type == GGML_TYPE_F16) {
return tensor->data;
}
// src1_row [D, 1, 1, 1] -> input
src1_row.ne[1] = 1;
src1_row.ne[2] = 1;
src1_row.ne[3] = 1;
src1_row.nb[2] = nb11;
src1_row.nb[3] = nb11;
size_t total_size = f16_elem_size;
for (int i = 0; i < GGML_MAX_DIMS; i++) {
total_size *= tensor->ne[i];
}
void * buffer = allocator.alloc(total_size);
// dst_row [M, 1, 1, 1] -> out
dst_row.ne[1] = 1;
dst_row.ne[2] = 1;
dst_row.ne[3] = 1;
dst_row.nb[2] = nb1;
dst_row.nb[3] = nb1;
if (need_cast == false) {
return buffer;
}
//create weight for one row
ggml_cann_pool_alloc weight_allocator(ctx.pool());
void * weight_buffer = weight_allocator.alloc(nb02);
for (int64_t iid1 = 0; iid1 < ids->ne[1]; iid1++) {
for (int64_t id = 0; id < n_ids; id++) {
// expert index
int32_t i02 = *(int32_t *) (ids_host.data() + iid1 * ids->nb[1] + id * ids->nb[0]);
GGML_ASSERT(i02 >= 0 && i02 < n_as);
int64_t ne[GGML_MAX_DIMS];
size_t nb[GGML_MAX_DIMS] = { f16_elem_size };
for (int i = 0; i < GGML_MAX_DIMS; i++) {
ne[i] = tensor->ne[i];
if (i > 0) {
nb[i] = nb[i - 1] * ne[i - 1];
}
}
// If B = 1 (broadcast), always use 0; otherwise, use id.
int64_t i11 = (ne11 == 1 ? 0 : id);
int64_t i12 = iid1;
acl_tensor_ptr src_tensor = ggml_cann_create_tensor(tensor);
acl_tensor_ptr f16_tensor = ggml_cann_create_tensor(buffer, ACL_FLOAT16, f16_elem_size, ne, nb, GGML_MAX_DIMS);
aclnn_cast(ctx, src_tensor.get(), f16_tensor.get(), ACL_FLOAT16);
int64_t i1 = id;
int64_t i2 = i12;
return buffer;
};
void * src0_tmp_ptr = src0_original + i02 * weight_stride;
void * scale_tmp_ptr = src0_original + weight_size + i02 * scale_stride;
void * src1_tmp_ptr = src1_original + i11 * nb11 + i12 * nb12;
void * dst_tmp_ptr = dst_original + i1 * nb1 + i2 * nb2;
// Prepare input and output buffers
ggml_cann_pool_alloc input_alloc(ctx.pool());
void * input_buffer = prepare_f16_buffer(src1, input_alloc, true);
// mem cpy
ACL_CHECK(aclrtMemcpyAsync(weight_buffer, weight_stride, src0_tmp_ptr, weight_stride,
ACL_MEMCPY_DEVICE_TO_DEVICE, ctx.stream()));
void * scale_buffer = (char *) weight_buffer + weight_stride;
ACL_CHECK(aclrtMemcpyAsync(scale_buffer, scale_stride, scale_tmp_ptr, scale_stride,
ACL_MEMCPY_DEVICE_TO_DEVICE, ctx.stream()));
ggml_cann_pool_alloc output_alloc(ctx.pool());
void * output_buffer = prepare_f16_buffer(dst, output_alloc, false);
src0_row.data = weight_buffer;
src1_row.data = src1_tmp_ptr;
dst_row.data = dst_tmp_ptr;
dst_row.src[0] = &src0_row;
dst_row.src[1] = &src1_row;
// Process each batch
for (int64_t batch_idx = 0; batch_idx < n_batches; batch_idx++) {
// Create index tensor for current batch
const size_t index_offset = batch_idx * ids->nb[1];
acl_tensor_ptr batch_indices = ggml_cann_create_tensor(ids, ids->ne, ids->nb, 1, ACL_FORMAT_ND, index_offset);
ggml_cann_mul_mat(ctx, &dst_row);
// Select quantized weights using expert indices
// Q4_0 stores 2 values per byte, Q8_0 stores 1 value per byte
const int64_t weight_d = (type == GGML_TYPE_Q4_0) ? src0->ne[0] / 2 : src0->ne[0];
const int64_t weight_m = src0->ne[1];
const int64_t weight_n_experts = src0->ne[2];
int64_t weight_ne[3] = { weight_d, weight_m, weight_n_experts };
size_t weight_nb[3] = { sizeof(int8_t), weight_d * sizeof(int8_t), weight_d * weight_m * sizeof(int8_t) };
acl_tensor_ptr all_weights =
ggml_cann_create_tensor(src0->data, ACL_INT8, sizeof(int8_t), weight_ne, weight_nb, 3);
int64_t selected_weight_ne[3] = { weight_d, weight_m, n_select_experts };
size_t selected_weight_nb[3] = { sizeof(int8_t), weight_d * sizeof(int8_t),
weight_d * weight_m * sizeof(int8_t) };
acl_tensor_ptr selected_weights = ggml_cann_create_tensor(selected_weight_buffer, ACL_INT8, sizeof(int8_t),
selected_weight_ne, selected_weight_nb, 3);
GGML_CANN_CALL_ACLNN_OP(ctx, IndexSelect, all_weights.get(), 0, batch_indices.get(), selected_weights.get());
// Select scales using the same expert indices
const int64_t scale_d = src0->ne[0] / group_size;
int64_t scale_ne[3] = { scale_d, weight_m, weight_n_experts };
size_t scale_nb[3] = { scale_elem_size, scale_d * scale_elem_size, scale_d * weight_m * scale_elem_size };
acl_tensor_ptr all_scales =
ggml_cann_create_tensor(scale_data, ACL_FLOAT16, scale_elem_size, scale_ne, scale_nb, 3);
int64_t selected_scale_ne[3] = { scale_d, weight_m, n_select_experts };
size_t selected_scale_nb[3] = { scale_elem_size, scale_d * scale_elem_size,
scale_d * weight_m * scale_elem_size };
acl_tensor_ptr selected_scales = ggml_cann_create_tensor(selected_scale_buffer, ACL_FLOAT16, scale_elem_size,
selected_scale_ne, selected_scale_nb, 3);
GGML_CANN_CALL_ACLNN_OP(ctx, IndexSelect, all_scales.get(), 0, batch_indices.get(), selected_scales.get());
// Process each expert for current batch
// IndexSelect output layout: [D, M, K] in contiguous format
// WeightQuantBatchMatmulV2 expects: [M, D] with row-major stride
for (int64_t expert_idx = 0; expert_idx < n_select_experts; expert_idx++) {
// Determine input offset: broadcast if src1->ne[1]==1, otherwise use per-expert input
const size_t input_offset =
(batch_idx * src1->ne[1] + (src1->ne[1] == 1 ? 0 : expert_idx)) * src1->ne[0] * f16_elem_size;
const size_t output_offset = (batch_idx * dst->ne[1] + expert_idx) * dst->ne[0] * f16_elem_size;
// Create weight view for current expert: [D, M, K] -> [M, D]
int64_t weight_view_ne[2] = { weight_m, src0->ne[0] };
float weight_view_nb[2] = { src0->ne[0] * weight_elem_size, weight_elem_size };
const size_t weight_view_offset = expert_idx * selected_weight_nb[2];
acl_tensor_ptr weight_view =
ggml_cann_create_tensor(selected_weight_buffer, ggml_cann_type_mapping(type), weight_elem_size,
weight_view_ne, weight_view_nb, 2, ACL_FORMAT_ND, weight_view_offset);
// Create scale view for current expert: [D, M, K] -> [M, D]
int64_t scale_view_ne[2] = { weight_m, scale_d };
size_t scale_view_nb[2] = { selected_scale_nb[1], selected_scale_nb[0] };
const size_t scale_view_offset = expert_idx * selected_scale_nb[2];
acl_tensor_ptr scale_view =
ggml_cann_create_tensor(selected_scale_buffer, ACL_FLOAT16, scale_elem_size, scale_view_ne,
scale_view_nb, 2, ACL_FORMAT_ND, scale_view_offset);
// Create input activation tensor [D, 1]
int64_t input_ne[2] = { src1->ne[0], 1 };
size_t input_nb[2] = { f16_elem_size, src1->ne[0] * f16_elem_size };
acl_tensor_ptr input_tensor = ggml_cann_create_tensor(input_buffer, ACL_FLOAT16, f16_elem_size, input_ne,
input_nb, 2, ACL_FORMAT_ND, input_offset);
// Create output tensor [M, 1]
int64_t output_ne[2] = { dst->ne[0], 1 };
size_t output_nb[2] = { f16_elem_size, dst->ne[0] * f16_elem_size };
acl_tensor_ptr output_tensor = ggml_cann_create_tensor(output_buffer, ACL_FLOAT16, f16_elem_size, output_ne,
output_nb, 2, ACL_FORMAT_ND, output_offset);
// Perform quantized matrix multiplication
GGML_CANN_CALL_ACLNN_OP(ctx, WeightQuantBatchMatmulV2, input_tensor.get(), weight_view.get(),
scale_view.get(), nullptr, nullptr, nullptr, nullptr, group_size,
output_tensor.get());
}
}
return;
// Cast output back to original type if we used a temporary F16 buffer
if (dst->type != GGML_TYPE_F16) {
int64_t ne[GGML_MAX_DIMS];
size_t nb[GGML_MAX_DIMS] = { f16_elem_size };
for (int i = 0; i < GGML_MAX_DIMS; i++) {
ne[i] = dst->ne[i];
if (i > 0) {
nb[i] = nb[i - 1] * ne[i - 1];
}
}
acl_tensor_ptr f16_output =
ggml_cann_create_tensor(output_buffer, ACL_FLOAT16, f16_elem_size, ne, nb, GGML_MAX_DIMS);
acl_tensor_ptr dst_tensor = ggml_cann_create_tensor(dst);
aclnn_cast(ctx, f16_output.get(), dst_tensor.get(), ggml_cann_type_mapping(dst->type));
}
}
void ggml_cann_mul_mat_id(ggml_backend_cann_context & ctx, ggml_tensor * dst) {

89
ggml/src/ggml-cann/ggml-cann.cpp
View File

@ -794,19 +794,44 @@ struct ggml_backend_cann_buffer_context {
~ggml_backend_cann_buffer_context() { ACL_CHECK(aclrtFree(dev_ptr)); }
};
// cann buffer type
/**
* @brief Check if a buffer is a CANN buffer.
*
* This function checks if a given buffer is a CANN buffer by comparing its
* `get_name` function pointer to `ggml_backend_cann_buffer_get_name`.
*
* @param buffer The buffer to check.
* @return true if the buffer is a CANN buffer, false otherwise.
* @brief Structure representing context information for a specific backend
* buffer type.
*/
static bool ggml_backend_buft_is_cann(ggml_backend_buffer_type_t buft);
struct ggml_backend_cann_buffer_type_context {
int32_t device; /**< Device identifier associated with the buffer context. */
std::string name; /**< Name associated with the buffer context. */
};
static bool ggml_backend_buffer_is_cann(ggml_backend_buffer_t buffer) {
return ggml_backend_buft_is_cann(buffer->buft);
/**
* @brief Retrieves the name associated with a CANN buffer type.
*
* This function returns the descriptive name associated with the specified
* CANN buffer type context.
*
* @param buft Pointer to the buffer type context.
* @return Const pointer to the C-style string containing the name.
*/
static const char * ggml_backend_cann_buffer_type_name(ggml_backend_buffer_type_t buft) {
ggml_backend_cann_buffer_type_context * buft_ctx = (ggml_backend_cann_buffer_type_context *) buft->context;
return buft_ctx->name.c_str();
}
/**
* @brief Checks if the backend buffer type is associated with the CANN backend.
*
* This function checks whether the provided backend buffer type is associated
* with the CANN backend based on the comparison of its name retrieval function
* pointer.
*
* @param buft Pointer to the backend buffer type to check.
* @return bool Returns true if the buffer type is associated with the CANN
* backend, otherwise false.
*/
static bool ggml_backend_buft_is_cann(ggml_backend_buffer_type_t buft) {
return buft->iface.get_name == ggml_backend_cann_buffer_type_name;
}
/**
@ -1271,7 +1296,7 @@ static void ggml_backend_cann_buffer_get_tensor(ggml_backend_buffer_t buffer,
static bool ggml_backend_cann_buffer_cpy_tensor(ggml_backend_buffer_t buffer,
const ggml_tensor * src,
ggml_tensor * dst) {
if (ggml_backend_buffer_is_cann(src->buffer)) {
if (ggml_backend_buft_is_cann(src->buffer->buft)) {
ggml_backend_cann_buffer_context * src_ctx = (ggml_backend_cann_buffer_context *) src->buffer->context;
ggml_backend_cann_buffer_context * dst_ctx = (ggml_backend_cann_buffer_context *) buffer->context;
@ -1335,31 +1360,6 @@ static const ggml_backend_buffer_i ggml_backend_cann_buffer_interface = {
/* .reset = */ NULL,
};
// cann buffer type
/**
* @brief Structure representing context information for a specific backend
* buffer type.
*/
struct ggml_backend_cann_buffer_type_context {
int32_t device; /**< Device identifier associated with the buffer context. */
std::string name; /**< Name associated with the buffer context. */
};
/**
* @brief Retrieves the name associated with a CANN buffer type.
*
* This function returns the descriptive name associated with the specified
* CANN buffer type context.
*
* @param buft Pointer to the buffer type context.
* @return Const pointer to the C-style string containing the name.
*/
static const char * ggml_backend_cann_buffer_type_name(ggml_backend_buffer_type_t buft) {
ggml_backend_cann_buffer_type_context * buft_ctx = (ggml_backend_cann_buffer_type_context *) buft->context;
return buft_ctx->name.c_str();
}
/**
* @brief Allocates a new CANN buffer of the specified type and size.
*
@ -1997,7 +1997,7 @@ static bool ggml_backend_cann_cpy_tensor_async(ggml_backend_t backend_src,
GGML_ASSERT(!is_matmul_weight((const ggml_tensor *) src));
if (!ggml_backend_buffer_is_cann(src->buffer) || !ggml_backend_buffer_is_cann(dst->buffer)) {
if (!ggml_backend_buft_is_cann(src->buffer->buft) || !ggml_backend_buft_is_cann(dst->buffer->buft)) {
return false;
}
@ -2523,21 +2523,6 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev, const ggml_ten
GGML_UNUSED(dev);
}
/**
* @brief Checks if the backend buffer type is associated with the CANN backend.
*
* This function checks whether the provided backend buffer type is associated
* with the CANN backend based on the comparison of its name retrieval function
* pointer.
*
* @param buft Pointer to the backend buffer type to check.
* @return bool Returns true if the buffer type is associated with the CANN
* backend, otherwise false.
*/
static bool ggml_backend_buft_is_cann(ggml_backend_buffer_type_t buft) {
return buft->iface.get_name == ggml_backend_cann_buffer_type_name;
}
/**
* @brief Records an event on the CANN backend stream.
*

16
ggml/src/ggml-cpu/arch-fallback.h
View File

@ -43,6 +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_8x4_q8_K_generic ggml_gemv_q6_K_8x4_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
@ -55,7 +56,8 @@
#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_q6_K_8x4_q8_K_generic ggml_gemm_q6_K_8x4_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
@ -76,6 +78,7 @@
#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_8x4_q8_K_generic ggml_gemv_q6_K_8x4_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
@ -84,6 +87,7 @@
#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_8x4_q8_K_generic ggml_gemm_q6_K_8x4_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
@ -107,6 +111,7 @@
#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_8x4_q8_K_generic ggml_gemv_q6_K_8x4_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
@ -119,6 +124,7 @@
#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_8x4_q8_K_generic ggml_gemm_q6_K_8x4_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
@ -143,6 +149,7 @@
#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_8x4_q8_K_generic ggml_gemv_q6_K_8x4_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
@ -155,6 +162,7 @@
#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_8x4_q8_K_generic ggml_gemm_q6_K_8x4_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
@ -186,6 +194,7 @@
#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_8x4_q8_K_generic ggml_gemv_q6_K_8x4_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
@ -197,6 +206,7 @@
#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_8x4_q8_K_generic ggml_gemm_q6_K_8x4_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
@ -227,6 +237,7 @@
#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_8x4_q8_K_generic ggml_gemv_q6_K_8x4_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
@ -239,6 +250,7 @@
#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_8x4_q8_K_generic ggml_gemm_q6_K_8x4_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
@ -271,6 +283,7 @@
#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_8x4_q8_K_generic ggml_gemv_q6_K_8x4_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
@ -283,6 +296,7 @@
#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_8x4_q8_K_generic ggml_gemm_q6_K_8x4_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

400
ggml/src/ggml-cpu/arch/arm/repack.cpp
View File

@ -1072,6 +1072,195 @@ 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_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) {
constexpr int qk = QK_K;
const int nb = n / qk;
constexpr int ncols_interleaved = 8;
constexpr int blocklen = 4;
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_groups = ncols_interleaved / 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);
// 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);
for (int i = 0; i < col_groups; i++) {
acc_f32[i] = 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);
int32x4_t acc[col_groups];
for (int i = 0; i < col_groups; i++) {
acc[i] = vdupq_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 four interleaved columns of q6)
int8x16_t q8_l[4];
int8x16_t q8_h[4];
for (int i = 0; i < 4; i++) {
q8_l[i] = (int8x16_t) vld1q_dup_s32((const int32_t *) (q8_base_l + i * 4));
q8_h[i] = (int8x16_t) vld1q_dup_s32((const int32_t *) (q8_base_h + i * 4));
}
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)
uint8x16x4_t q6_ql_0 = vld1q_u8_x4(ql_base + ql_off_base);
uint8x16x4_t q6_ql_1 = vld1q_u8_x4(ql_base + ql_off_base + 64);
uint8x16x4_t q6_qh_0 = vld1q_u8_x4(qh_base + qh_off_base);
uint8x16x4_t q6_qh_1 = 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);
}
const uint8x16_t q6_ql[8] = { q6_ql_0.val[0], q6_ql_0.val[1], q6_ql_0.val[2], q6_ql_0.val[3],
q6_ql_1.val[0], q6_ql_1.val[1], q6_ql_1.val[2], q6_ql_1.val[3] };
const uint8x16_t q6_qh[8] = { q6_qh_0.val[0], q6_qh_0.val[1], q6_qh_0.val[2], q6_qh_0.val[3],
q6_qh_1.val[0], q6_qh_1.val[1], q6_qh_1.val[2], q6_qh_1.val[3] };
// Process column groups (0-3, 4-7)
for (int g = 0; g < col_groups; g++) {
int32x4_t sb_acc_l = vdupq_n_s32(0);
int32x4_t sb_acc_h = vdupq_n_s32(0);
for (int chunk = 0; chunk < 4; chunk++) {
const int idx = chunk * 2 + g;
const uint8x16_t q6_qs_l = q6_ql[idx];
const uint8x16_t q6_qs_h = q6_qh[idx];
// Extract high 2 bits for upper nibble reconstruction
const uint8x16_t q6_qs_hh = vandq_u8(q6_qs_h, mask_hi);
// q6 = (low4 | high2<<4), without -32 bias (handled via bsums)
const int8x16_t q6_l =
vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(q6_qs_l, m4b), vandq_u8(q6_qs_h, mask_lo), 4));
const int8x16_t q6_h = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6_qs_l, 4), q6_qs_hh));
sb_acc_l = vdotq_s32(sb_acc_l, q6_l, q8_l[chunk]);
sb_acc_h = vdotq_s32(sb_acc_h, q6_h, q8_h[chunk]);
}
const int scale_idx_l = half * 8 + sb;
const int scale_idx_h = half * 8 + sb + 4;
const int32x4_t scale_vec_l = vmovl_s16(vld1_s16(q6_scales + scale_idx_l * 8 + g * 4));
const int32x4_t scale_vec_h = vmovl_s16(vld1_s16(q6_scales + scale_idx_h * 8 + g * 4));
acc[g] = vmlaq_s32(acc[g], sb_acc_l, scale_vec_l);
acc[g] = vmlaq_s32(acc[g], sb_acc_h, scale_vec_h);
}
}
} // for half
// Bias correction
acc[0] = vsubq_s32(acc[0], bias_lo);
acc[1] = vsubq_s32(acc[1], bias_hi);
// Apply superblock scale (no mins for q6_K)
// acc[g] has [c0, c1, c2, c3]
float32x4_t w_0123 = vmulq_f32(vcvtq_f32_s32(acc[0]), sb_scale_0);
float32x4_t w_4567 = vmulq_f32(vcvtq_f32_s32(acc[1]), sb_scale_1);
acc_f32[0] = vaddq_f32(acc_f32[0], w_0123);
acc_f32[1] = vaddq_f32(acc_f32[1], w_4567);
} // 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_8x4_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,
@ -1177,15 +1366,14 @@ void ggml_gemv_q6_K_8x8_q8_K(int n,
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);
uint8x16x4_t q6_ql_0 = vld1q_u8_x4(ql_base + ql_off_base);
uint8x16x4_t q6_ql_1 = vld1q_u8_x4(ql_base + ql_off_base + 64);
uint8x16x4_t q6_qh_0 = vld1q_u8_x4(qh_base + qh_off_base);
uint8x16x4_t q6_qh_1 = vld1q_u8_x4(qh_base + qh_off_base + 64);
// Adjust qh for subblocks 2 and 3 (shift right by 2)
if (sb > 1) {
@ -3474,6 +3662,208 @@ 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_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) {
constexpr int qk = QK_K;
const int nb = n / qk;
constexpr int ncols_interleaved = 8;
constexpr int blocklen = 4;
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_DOTPROD)
constexpr int q8_k_blocklen = 4;
constexpr int col_groups = ncols_interleaved / 4;
constexpr int acc_size = q8_k_blocklen * col_groups; // 4 rows, 2 column groups
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);
float32x4_t acc_f32[acc_size];
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 < acc_size; i++) {
acc_f32[i] = vdupq_n_f32(0);
}
for (int b = 0; b < nb; b++) {
float32x4_t q6_d_0123 = vcvt_f32_f16(vld1_f16((const __fp16 *) q6_ptr[b].d));
float32x4_t q6_d_4567 = vcvt_f32_f16(vld1_f16((const __fp16 *) q6_ptr[b].d + 4));
float32x4_t q8_d_0123 = vld1q_f32(q8_ptr[b].d);
float32x4_t sbd_scale_0123[q8_k_blocklen];
float32x4_t sbd_scale_4567[q8_k_blocklen];
sbd_scale_0123[0] = vmulq_laneq_f32(q6_d_0123, q8_d_0123, 0);
sbd_scale_4567[0] = vmulq_laneq_f32(q6_d_4567, q8_d_0123, 0);
sbd_scale_0123[1] = vmulq_laneq_f32(q6_d_0123, q8_d_0123, 1);
sbd_scale_4567[1] = vmulq_laneq_f32(q6_d_4567, q8_d_0123, 1);
sbd_scale_0123[2] = vmulq_laneq_f32(q6_d_0123, q8_d_0123, 2);
sbd_scale_4567[2] = vmulq_laneq_f32(q6_d_4567, q8_d_0123, 2);
sbd_scale_0123[3] = vmulq_laneq_f32(q6_d_0123, q8_d_0123, 3);
sbd_scale_4567[3] = vmulq_laneq_f32(q6_d_4567, q8_d_0123, 3);
int32x4_t acc_s32[acc_size];
for (int i = 0; i < acc_size; i++) {
acc_s32[i] = vdupq_n_s32(0);
}
int16_t q6_scales[8 * 16];
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);
}
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;
for (int sb = 0; sb < QK_K / 64; sb++) {
int32x4_t acc_lo[acc_size];
int32x4_t acc_hi[acc_size];
for (int i = 0; i < acc_size; i++) {
acc_lo[i] = vdupq_n_s32(0);
acc_hi[i] = vdupq_n_s32(0);
}
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;
// 4 rows * 16 elements per scale
// 4 reads of 16 bytes each
constexpr int reads_per_sb = 4;
int8x16_t q8_l[reads_per_sb];
int8x16_t q8_h[reads_per_sb];
for (int k = 0; k < reads_per_sb; k++) {
q8_l[k] = vld1q_s8(q8_base_l + 16 * k);
q8_h[k] = vld1q_s8(q8_base_h + 16 * k);
}
const int ql_off_base = sb * QK_K / 2;
const int qh_off_base = ql_off_base & 255;
uint8x16_t q6_ql_0123[reads_per_sb];
uint8x16_t q6_ql_4567[reads_per_sb];
uint8x16_t q6_qh_0123[reads_per_sb];
uint8x16_t q6_qh_4567[reads_per_sb];
for (int k = 0; k < reads_per_sb; k++) {
q6_ql_0123[k] = vld1q_u8(ql_base + ql_off_base + k * 32);
q6_ql_4567[k] = vld1q_u8(ql_base + ql_off_base + k * 32 + 16);
q6_qh_0123[k] = vld1q_u8(qh_base + qh_off_base + k * 32);
q6_qh_4567[k] = vld1q_u8(qh_base + qh_off_base + k * 32 + 16);
}
if (sb > 1) {
for (int k = 0; k < reads_per_sb; k++) {
q6_qh_0123[k] = vshrq_n_u8(q6_qh_0123[k], 2);
q6_qh_4567[k] = vshrq_n_u8(q6_qh_4567[k], 2);
}
}
for (int k = 0; k < reads_per_sb; k++) {
// q = (ql | qh) - 32
const uint8x16_t hbit_lo_0123 = vandq_u8(q6_qh_0123[k], mask_lo);
const uint8x16_t hbit_hi_0123 = vandq_u8(q6_qh_0123[k], mask_hi);
const uint8x16_t hbit_lo_4567 = vandq_u8(q6_qh_4567[k], mask_lo);
const uint8x16_t hbit_hi_4567 = vandq_u8(q6_qh_4567[k], mask_hi);
const int8x16_t q6_0123_lo = vsubq_s8(
vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(q6_ql_0123[k], m4b), hbit_lo_0123, 4)), m32s);
const int8x16_t q6_0123_hi = vsubq_s8(
vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6_ql_0123[k], 4), hbit_hi_0123)), m32s);
acc_lo[0] = vdotq_laneq_s32(acc_lo[0], q6_0123_lo, q8_l[k], 0); // 0..3 r0 c0123
acc_lo[1] = vdotq_laneq_s32(acc_lo[1], q6_0123_lo, q8_l[k], 1); // 0..3 r1 c0123
acc_lo[2] = vdotq_laneq_s32(acc_lo[2], q6_0123_lo, q8_l[k], 2); // 0..3 r2 c0123
acc_lo[3] = vdotq_laneq_s32(acc_lo[3], q6_0123_lo, q8_l[k], 3); // 0..3 r3 c0123
acc_hi[0] = vdotq_laneq_s32(acc_hi[0], q6_0123_hi, q8_h[k], 0); // 64..67 r0 c0123
acc_hi[1] = vdotq_laneq_s32(acc_hi[1], q6_0123_hi, q8_h[k], 1); // 64..67 r1 c0123
acc_hi[2] = vdotq_laneq_s32(acc_hi[2], q6_0123_hi, q8_h[k], 2); // 64..67 r2 c0123
acc_hi[3] = vdotq_laneq_s32(acc_hi[3], q6_0123_hi, q8_h[k], 3); // 64..67 r3 c0123
const int8x16_t q6_4567_lo = vsubq_s8(
vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(q6_ql_4567[k], m4b), hbit_lo_4567, 4)), m32s);
const int8x16_t q6_4567_hi = vsubq_s8(
vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6_ql_4567[k], 4), hbit_hi_4567)), m32s);
acc_lo[4] = vdotq_laneq_s32(acc_lo[4], q6_4567_lo, q8_l[k], 0); // 0..3 r0 c4567
acc_lo[5] = vdotq_laneq_s32(acc_lo[5], q6_4567_lo, q8_l[k], 1); // 0..3 r1 c4567
acc_lo[6] = vdotq_laneq_s32(acc_lo[6], q6_4567_lo, q8_l[k], 2); // 0..3 r2 c4567
acc_lo[7] = vdotq_laneq_s32(acc_lo[7], q6_4567_lo, q8_l[k], 3); // 0..3 r3 c4567
acc_hi[4] = vdotq_laneq_s32(acc_hi[4], q6_4567_hi, q8_h[k], 0); // 64..67 r0 c4567
acc_hi[5] = vdotq_laneq_s32(acc_hi[5], q6_4567_hi, q8_h[k], 1); // 64..67 r1 c4567
acc_hi[6] = vdotq_laneq_s32(acc_hi[6], q6_4567_hi, q8_h[k], 2); // 64..67 r2 c4567
acc_hi[7] = vdotq_laneq_s32(acc_hi[7], q6_4567_hi, q8_h[k], 3); // 64..67 r3 c4567
}
// Scale and bias
const int scale_idx_l = half * 8 + sb;
const int scale_idx_h = half * 8 + sb + 4;
for (int g = 0; g < col_groups; g++) {
const int16x4_t scales_l16 = vld1_s16(q6_scales + scale_idx_l * 8 + g * 4);
const int16x4_t scales_h16 = vld1_s16(q6_scales + scale_idx_h * 8 + g * 4);
const int32x4_t scale_vec_l = vmovl_s16(scales_l16);
const int32x4_t scale_vec_h = vmovl_s16(scales_h16);
const int acc_offset = g * q8_k_blocklen;
for (int row = 0; row < q8_k_blocklen; row++) {
const int idx = row * 2 + g;
acc_s32[idx] = vmlaq_s32(acc_s32[idx], acc_lo[acc_offset + row], scale_vec_l);
acc_s32[idx] = vmlaq_s32(acc_s32[idx], acc_hi[acc_offset + row], scale_vec_h);
}
}
}
}
// Finally we apply the superblock scales
for (int row = 0; row < q8_k_blocklen; row++) {
const int idx0 = 2 * row;
const int idx1 = 2 * row + 1;
const int32x4_t acc_0123 = acc_s32[idx0];
const int32x4_t acc_4567 = acc_s32[idx1];
acc_f32[idx0] = vmlaq_f32(acc_f32[idx0], vcvtq_f32_s32(acc_0123), sbd_scale_0123[row]);
acc_f32[idx1] = vmlaq_f32(acc_f32[idx1], vcvtq_f32_s32(acc_4567), sbd_scale_4567[row]);
}
} // for b
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_DOTPROD)
ggml_gemm_q6_K_8x4_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,

3
ggml/src/ggml-cpu/ops.cpp
View File

@ -7629,8 +7629,7 @@ static void ggml_compute_forward_pad_f32(
const ggml_tensor * src0 = dst->src[0];
GGML_ASSERT(src0->nb[0] == sizeof(float));
GGML_ASSERT( dst->nb[0] == sizeof(float));
assert(dst->nb[0] == sizeof(float));
const int ith = params->ith;
const int nth = params->nth;

425
ggml/src/ggml-cpu/repack.cpp
View File

@ -256,6 +256,200 @@ template <> void ggml_quantize_mat_t<8, GGML_TYPE_Q8_K>(const float * GGML_RESTR
ggml_quantize_mat_q8_K_4x8(x, vy, n_per_row);
}
template <int M, int N>
static void ggml_gemv_q6_K_NxM_q8_K_generic_impl(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 blocklen = M;
constexpr int ncols_interleaved = N;
const int qk = QK_K;
const int nb = n / qk;
const int blocks_per_half = 64 / blocklen;
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 < (qk / (2 * blocklen)); k++) {
const int base_l = (k / blocks_per_half) * 128 + (k % blocks_per_half) * blocklen;
const int base_h = base_l + 64;
const int scale_idx_l = base_l / 16;
const int scale_idx_h = base_h / 16;
const int qh_shift_l = ((base_l % 128) / 32) * 2;
const int qh_shift_h = ((base_h % 128) / 32) * 2;
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++) {
const int8_t scale_l = b_ptr[l].scales[scale_idx_l * ncols_interleaved + j];
const int8_t scale_h = b_ptr[l].scales[scale_idx_h * ncols_interleaved + j];
int sumi_l = 0;
int sumi_h = 0;
for (int i = 0; i < blocklen; i++) {
const int ql_pos = k * ncols_interleaved * blocklen + j * blocklen + 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 / blocklen;
const int qh_pos_l = qh_idx_l % blocklen;
const int qh_offset_l = qh_chunk_l * (blocklen * ncols_interleaved) + j * blocklen + 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 / blocklen;
const int qh_pos_h = qh_idx_h % blocklen;
const int qh_offset_h = qh_chunk_h * (blocklen * ncols_interleaved) + j * blocklen + 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];
}
}
}
template <int M, int N>
static void ggml_gemm_q6_K_NxM_q8_K_generic_impl(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 blocklen = M;
constexpr int ncols_interleaved = N;
const int qk = QK_K;
const int nb = n / qk;
const int blocks_per_half = 64 / blocklen;
const int q8_half_stride = 512;
const int q8_low_high_step = 256;
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 < (qk / (2 * blocklen)); k++) {
const int base_l = (k / blocks_per_half) * 128 + (k % blocks_per_half) * blocklen;
const int base_h = base_l + 64;
const int scale_idx_l = base_l / 16;
const int scale_idx_h = base_h / 16;
const int qh_shift_l = ((base_l % 128) / 32) * 2;
const int qh_shift_h = ((base_h % 128) / 32) * 2;
const int qh_half_l = (base_l / 128) * 32;
const int qh_half_h = (base_h / 128) * 32;
const int q8_base = (k / blocks_per_half) * q8_half_stride + (k % blocks_per_half) * (blocklen * 4);
for (int m = 0; m < 4; m++) {
for (int j = 0; j < ncols_interleaved; j++) {
const int8_t scale_l = b_ptr[l].scales[scale_idx_l * ncols_interleaved + j];
const int8_t scale_h = b_ptr[l].scales[scale_idx_h * ncols_interleaved + j];
int sumi_l = 0;
int sumi_h = 0;
for (int i = 0; i < blocklen; i++) {
const int ql_pos = k * ncols_interleaved * blocklen + j * blocklen + 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 / blocklen;
const int qh_pos_l = qh_idx_l % blocklen;
const int qh_offset_l =
qh_chunk_l * (blocklen * ncols_interleaved) + j * blocklen + 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 / blocklen;
const int qh_pos_h = qh_idx_h % blocklen;
const int qh_offset_h =
qh_chunk_h * (blocklen * ncols_interleaved) + j * blocklen + 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 * blocklen + i];
const int8_t q8_h = a_ptr[l].qs[q8_base + m * blocklen + i + q8_low_high_step];
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];
}
}
}
}
}
extern "C" {
void ggml_gemv_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) {
@ -704,94 +898,12 @@ void ggml_gemv_q5_K_8x8_q8_K_generic(int n,
}
void ggml_gemv_q6_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) {
ggml_gemv_q6_K_NxM_q8_K_generic_impl<4, 8>(n, s, bs, vx, vy, nr, 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) {
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];
}
}
ggml_gemv_q6_K_NxM_q8_K_generic_impl<8, 8>(n, s, bs, vx, vy, nr, 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) {
@ -1485,109 +1597,12 @@ 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;
void ggml_gemm_q6_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) {
ggml_gemm_q6_K_NxM_q8_K_generic_impl<4, 8>(n, s, bs, vx, vy, nr, nc);
}
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_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) {
ggml_gemm_q6_K_NxM_q8_K_generic_impl<8, 8>(n, s, bs, vx, vy, nr, 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) {
@ -2097,18 +2112,18 @@ static block_q6_Kx8 make_block_q6_Kx8(block_q6_K * in, unsigned int blck_size_in
}
const int end_ls = QK_K * 4 / blck_size_interleave;
// Interleave Q6_K quants by taking 8 bytes at a time
// Interleave Q6_K quants by taking blck_size_interleave 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));
memcpy(&elem_ls, &in[src_id].ql[src_offset], blck_size_interleave);
memcpy(&out.ql[dst_offset], &elem_ls, blck_size_interleave);
}
// Interleave high bits using same 8-byte pattern as low bits
// Interleave high bits using same chunk size as low bits
const int end_hs = end_ls / 2;
for (int i = 0; i < end_hs; ++i) {
int src_id = i % n_blocks;
@ -2116,8 +2131,8 @@ static block_q6_Kx8 make_block_q6_Kx8(block_q6_K * in, unsigned int blck_size_in
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));
memcpy(&elem_hs, &in[src_id].qh[src_offset], blck_size_interleave);
memcpy(&out.qh[dst_offset], &elem_hs, blck_size_interleave);
}
// The below logic is designed so as to unpack and rearrange scales in Q6_K
@ -2262,7 +2277,7 @@ static int repack_q5_K_to_q5_K_8_bl(struct ggml_tensor * t,
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);
GGML_ASSERT(interleave_block == 4 || interleave_block == 8);
constexpr int nrows_interleaved = 8;
block_q6_Kx8 * dst = (block_q6_Kx8 *)t->data;
@ -2511,6 +2526,10 @@ template <> int repack<block_q5_K, 8, 8>(struct ggml_tensor * t, const void * da
return repack_q5_K_to_q5_K_8_bl(t, 8, data, data_size);
}
template <> int repack<block_q6_K, 4, 8>(struct ggml_tensor * t, const void * data, size_t data_size) {
return repack_q6_K_to_q6_K_8_bl(t, 4, data, data_size);
}
template <> int repack<block_q6_K, 8, 8>(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);
}
@ -2575,6 +2594,10 @@ template <> void gemv<block_q5_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t
ggml_gemv_q5_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
}
template <> void gemv<block_q6_K, 4, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
ggml_gemv_q6_K_8x4_q8_K(n, s, bs, vx, vy, nr, nc);
}
template <> void gemv<block_q6_K, 8, 8, GGML_TYPE_Q8_K>(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);
}
@ -2634,6 +2657,10 @@ template <> void gemm<block_q5_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t
ggml_gemm_q5_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
}
template <> void gemm<block_q6_K, 4, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
ggml_gemm_q6_K_8x4_q8_K(n, s, bs, vx, vy, nr, nc);
}
template <> void gemm<block_q6_K, 8, 8, GGML_TYPE_Q8_K>(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);
}
@ -3043,6 +3070,7 @@ static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(cons
static const ggml::cpu::repack::tensor_traits<block_q5_K, 8, 8, GGML_TYPE_Q8_K> q5_K_8x8_q8_K;
// instance for Q6_K
static const ggml::cpu::repack::tensor_traits<block_q6_K, 4, 8, GGML_TYPE_Q8_K> q6_K_8x4_q8_K;
static const ggml::cpu::repack::tensor_traits<block_q6_K, 8, 8, GGML_TYPE_Q8_K> q6_K_8x8_q8_K;
// instance for Q2
@ -3107,6 +3135,11 @@ static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(cons
return &q6_K_8x8_q8_K;
}
}
if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
if (cur->ne[1] % 8 == 0) {
return &q6_K_8x4_q8_K;
}
}
} else if (cur->type == GGML_TYPE_IQ4_NL) {
if (ggml_cpu_has_avx2()) {
if (cur->ne[1] % 8 == 0) {

4
ggml/src/ggml-cpu/repack.h
View File

@ -112,6 +112,7 @@ void ggml_gemv_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
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_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_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);
@ -122,6 +123,7 @@ 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_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_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);
@ -142,6 +144,7 @@ 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_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_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);
@ -152,6 +155,7 @@ 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_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_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);

3
ggml/src/ggml-cuda/ggml-cuda.cu
View File

@ -4834,8 +4834,9 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
case GGML_OP_SUM_ROWS:
case GGML_OP_MEAN:
case GGML_OP_GROUP_NORM:
case GGML_OP_PAD:
return ggml_is_contiguous(op->src[0]);
case GGML_OP_PAD:
return true;
case GGML_OP_UPSCALE:
case GGML_OP_PAD_REFLECT_1D:
case GGML_OP_ARANGE:

23
ggml/src/ggml-cuda/pad.cu
View File

@ -7,7 +7,7 @@ __device__ __forceinline__ int64_t wrap_around(int64_t coord, int64_t size) {
return (coord + size) % size;
}
static __global__ void pad_f32(const float * src, float * dst,
static __global__ void pad_f32(const float * src, size_t s00, size_t s01, size_t s02, size_t s03, float * dst,
const int lp0, const int rp0, const int lp1, const int rp1,
const int lp2, const int rp2, const int lp3, const int rp3,
const int ne0, const int ne1, const int ne2, const int ne3,
@ -34,11 +34,8 @@ static __global__ void pad_f32(const float * src, float * dst,
const int64_t i01 = i1 - lp1;
const int64_t i02 = i2 - lp2;
const int64_t i03 = i3 - lp3;
const int64_t ne02 = ne2 - lp2 - rp2;
const int64_t ne01 = ne1 - lp1 - rp1;
const int64_t ne00 = ne0 - lp0 - rp0;
const int64_t src_idx = i03 * (ne00 * ne01 * ne02) + i02 * (ne00 * ne01) + i01 * ne00 + i00;
const int64_t src_idx = i03 * s03 + i02 * s02 + i01 * s01 + i00 * s00;
dst[dst_idx] = src[src_idx];
} else {
@ -57,21 +54,21 @@ static __global__ void pad_f32(const float * src, float * dst,
const int64_t i02 = wrap_around(i2 - lp2, ne02);
const int64_t i03 = wrap_around(i3 - lp3, ne03);
const int64_t src_idx = i03 * (ne00 * ne01 * ne02) + i02 * (ne00 * ne01) + i01 * ne00 + i00;
const int64_t src_idx = i03 * s03 + i02 * s02 + i01 * s01 + i00 * s00;
dst[dst_idx] = src[src_idx];
}
}
static void pad_f32_cuda(const float * src, float * dst,
static void pad_f32_cuda(const float * src, size_t s00, size_t s01, size_t s02, size_t s03, float * dst,
const int lp0, const int rp0, const int lp1, const int rp1,
const int lp2, const int rp2, const int lp3, const int rp3,
const int ne0, const int ne1, const int ne2, const int ne3,
const bool circular, cudaStream_t stream) {
int num_blocks = (ne0 + CUDA_PAD_BLOCK_SIZE - 1) / CUDA_PAD_BLOCK_SIZE;
dim3 gridDim(num_blocks, ne1, ne2 * ne3);
pad_f32<<<gridDim, CUDA_PAD_BLOCK_SIZE, 0, stream>>>(src, dst,
pad_f32<<<gridDim, CUDA_PAD_BLOCK_SIZE, 0, stream>>>(src, s00, s01, s02, s03, dst,
lp0, rp0, lp1, rp1, lp2, rp2, lp3, rp3,
ne0, ne1, ne2, ne3, circular);
}
@ -82,9 +79,10 @@ void ggml_cuda_op_pad(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
float * dst_d = (float *) dst->data;
cudaStream_t stream = ctx.stream();
GGML_TENSOR_UNARY_OP_LOCALS;
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT(dst->type == GGML_TYPE_F32);
GGML_ASSERT(ggml_is_contiguous(src0));
const int32_t lp0 = ((const int32_t *) (dst->op_params))[0];
const int32_t rp0 = ((const int32_t *) (dst->op_params))[1];
@ -96,7 +94,12 @@ void ggml_cuda_op_pad(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const int32_t rp3 = ((const int32_t *) (dst->op_params))[7];
const int32_t circular = ((const int32_t *) (dst->op_params))[8];
pad_f32_cuda(src0_d, dst_d,
const size_t s00 = nb00 / ggml_type_size(src0->type);
const size_t s01 = nb01 / ggml_type_size(src0->type);
const size_t s02 = nb02 / ggml_type_size(src0->type);
const size_t s03 = nb03 / ggml_type_size(src0->type);
pad_f32_cuda(src0_d, s00, s01, s02, s03, dst_d,
lp0, rp0, lp1, rp1, lp2, rp2, lp3, rp3,
dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3],
(bool) circular, stream);

81
ggml/src/ggml-webgpu/ggml-webgpu-shader-lib.hpp
View File

@ -4,6 +4,7 @@
#include "ggml.h"
#include "pre_wgsl.hpp"
#include <memory>
#include <string>
#include <vector>
@ -18,9 +19,9 @@
#define GGML_WEBGPU_ARGSORT_MERGE_MAX_WG_SIZE 512u
struct ggml_webgpu_processed_shader {
std::string wgsl;
std::string variant;
void * decisions;
std::string wgsl;
std::string variant;
std::shared_ptr<void> decisions;
};
// Same hash combine function as in boost
@ -192,13 +193,13 @@ inline ggml_webgpu_processed_shader ggml_webgpu_preprocess_flash_attn_shader(
defines.push_back(std::string("WG_SIZE=") + std::to_string(wg_size));
ggml_webgpu_processed_shader result;
result.wgsl = preprocessor.preprocess(shader_src, defines);
result.variant = variant;
ggml_webgpu_flash_attn_shader_decisions * decisions = new ggml_webgpu_flash_attn_shader_decisions();
decisions->q_tile = q_tile;
decisions->kv_tile = kv_tile;
decisions->wg_size = wg_size;
result.decisions = decisions;
result.wgsl = preprocessor.preprocess(shader_src, defines);
result.variant = variant;
auto decisions = std::make_shared<ggml_webgpu_flash_attn_shader_decisions>();
decisions->q_tile = q_tile;
decisions->kv_tile = kv_tile;
decisions->wg_size = wg_size;
result.decisions = decisions;
return result;
}
@ -270,11 +271,11 @@ inline ggml_webgpu_processed_shader ggml_webgpu_preprocess_pad_shader(
defines.push_back(std::string("WG_SIZE=") + std::to_string(context.max_wg_size));
ggml_webgpu_processed_shader result;
result.wgsl = preprocessor.preprocess(shader_src, defines);
result.variant = variant;
ggml_webgpu_generic_shader_decisions * decisions = new ggml_webgpu_generic_shader_decisions();
decisions->wg_size = context.max_wg_size;
result.decisions = decisions;
result.wgsl = preprocessor.preprocess(shader_src, defines);
result.variant = variant;
auto decisions = std::make_shared<ggml_webgpu_generic_shader_decisions>();
decisions->wg_size = context.max_wg_size;
result.decisions = decisions;
return result;
}
@ -305,11 +306,11 @@ inline ggml_webgpu_processed_shader ggml_webgpu_preprocess_argsort_shader(
}
defines.push_back(std::string("WG_SIZE=") + std::to_string(wg_size));
ggml_webgpu_processed_shader result;
result.wgsl = preprocessor.preprocess(shader_src, defines);
result.variant = variant;
ggml_webgpu_argsort_shader_decisions * decisions = new ggml_webgpu_argsort_shader_decisions();
decisions->wg_size = wg_size;
result.decisions = decisions;
result.wgsl = preprocessor.preprocess(shader_src, defines);
result.variant = variant;
auto decisions = std::make_shared<ggml_webgpu_argsort_shader_decisions>();
decisions->wg_size = wg_size;
result.decisions = decisions;
return result;
}
@ -324,11 +325,11 @@ inline ggml_webgpu_processed_shader ggml_webgpu_preprocess_argsort_merge_shader(
uint32_t wg_size = std::min(GGML_WEBGPU_ARGSORT_MERGE_MAX_WG_SIZE, context.max_wg_size);
defines.push_back(std::string("WG_SIZE=") + std::to_string(wg_size));
ggml_webgpu_processed_shader result;
result.wgsl = preprocessor.preprocess(shader_src, defines);
result.variant = variant;
ggml_webgpu_argsort_shader_decisions * decisions = new ggml_webgpu_argsort_shader_decisions();
decisions->wg_size = wg_size;
result.decisions = decisions;
result.wgsl = preprocessor.preprocess(shader_src, defines);
result.variant = variant;
auto decisions = std::make_shared<ggml_webgpu_argsort_shader_decisions>();
decisions->wg_size = wg_size;
result.decisions = decisions;
return result;
}
@ -391,11 +392,11 @@ inline ggml_webgpu_processed_shader ggml_webgpu_preprocess_set_rows_shader(
defines.push_back(std::string("WG_SIZE=") + std::to_string(context.max_wg_size));
ggml_webgpu_processed_shader result;
result.wgsl = preprocessor.preprocess(shader_src, defines);
result.variant = variant;
ggml_webgpu_generic_shader_decisions * decisions = new ggml_webgpu_generic_shader_decisions();
decisions->wg_size = context.max_wg_size;
result.decisions = decisions;
result.wgsl = preprocessor.preprocess(shader_src, defines);
result.variant = variant;
auto decisions = std::make_shared<ggml_webgpu_generic_shader_decisions>();
decisions->wg_size = context.max_wg_size;
result.decisions = decisions;
return result;
}
@ -457,11 +458,11 @@ inline ggml_webgpu_processed_shader ggml_webgpu_preprocess_unary_shader(
defines.push_back(std::string("WG_SIZE=") + std::to_string(context.max_wg_size));
ggml_webgpu_processed_shader result;
result.wgsl = preprocessor.preprocess(shader_src, defines);
result.variant = variant;
ggml_webgpu_generic_shader_decisions * decisions = new ggml_webgpu_generic_shader_decisions();
decisions->wg_size = context.max_wg_size;
result.decisions = decisions;
result.wgsl = preprocessor.preprocess(shader_src, defines);
result.variant = variant;
auto decisions = std::make_shared<ggml_webgpu_generic_shader_decisions>();
decisions->wg_size = context.max_wg_size;
result.decisions = decisions;
return result;
}
@ -527,11 +528,11 @@ inline ggml_webgpu_processed_shader ggml_webgpu_preprocess_binary_shader(
defines.push_back(std::string("WG_SIZE=") + std::to_string(context.max_wg_size));
ggml_webgpu_processed_shader result;
result.wgsl = preprocessor.preprocess(shader_src, defines);
result.variant = variant;
ggml_webgpu_generic_shader_decisions * decisions = new ggml_webgpu_generic_shader_decisions();
decisions->wg_size = context.max_wg_size;
result.decisions = decisions;
result.wgsl = preprocessor.preprocess(shader_src, defines);
result.variant = variant;
auto decisions = std::make_shared<ggml_webgpu_generic_shader_decisions>();
decisions->wg_size = context.max_wg_size;
result.decisions = decisions;
return result;
}
#endif // GGML_WEBGPU_SHADER_LIB_HPP

89
ggml/src/ggml-webgpu/ggml-webgpu.cpp
View File

@ -186,11 +186,17 @@ struct webgpu_buf_pool {
void cleanup() {
std::lock_guard<std::mutex> lock(mutex);
for (auto & bufs : free) {
bufs.host_buf.Destroy();
bufs.dev_buf.Destroy();
if (bufs.host_buf) {
bufs.host_buf.Destroy();
}
if (bufs.dev_buf) {
bufs.dev_buf.Destroy();
}
}
free.clear();
}
~webgpu_buf_pool() { this->cleanup(); }
};
#ifdef GGML_WEBGPU_GPU_PROFILE
@ -252,13 +258,15 @@ struct webgpu_gpu_profile_buf_pool {
}
free.clear();
}
~webgpu_gpu_profile_buf_pool() { this->cleanup(); }
};
#endif
struct webgpu_pipeline {
wgpu::ComputePipeline pipeline;
std::string name;
void * context = nullptr;
std::shared_ptr<void> context = nullptr;
};
struct webgpu_command {
@ -319,6 +327,23 @@ struct webgpu_global_context_struct {
wgpu::Buffer debug_host_buf;
wgpu::Buffer debug_dev_buf;
#endif
~webgpu_global_context_struct() {
if (this->get_tensor_staging_buf) {
this->get_tensor_staging_buf.Destroy();
this->get_tensor_staging_buf = nullptr;
}
#ifdef GGML_WEBGPU_DEBUG
if (this->debug_host_buf) {
this->debug_host_buf.Destroy();
this->debug_host_buf = nullptr;
}
if (this->debug_dev_buf) {
this->debug_dev_buf.Destroy();
this->debug_dev_buf = nullptr;
}
#endif
}
};
typedef std::shared_ptr<webgpu_global_context_struct> webgpu_global_context;
@ -744,7 +769,6 @@ static const char * ggml_backend_webgpu_name(ggml_backend_t backend) {
return ctx->name.c_str();
}
// TODO: implement proper cleanup
static void ggml_backend_webgpu_free(ggml_backend_t backend) {
ggml_backend_webgpu_context * ctx = (ggml_backend_webgpu_context *) backend->context;
WEBGPU_LOG_DEBUG("ggml_backend_webgpu_free(" << ctx->name << ")");
@ -788,9 +812,8 @@ static void ggml_backend_webgpu_free(ggml_backend_t backend) {
std::cout << "ggml_webgpu: gpu/cpu ratio: " << (total_cpu > 0.0 ? total_gpu / total_cpu : 0.0) << "\n";
#endif
#if !defined(GGML_WEBGPU_CPU_PROFILE) && !defined(GGML_WEBGPU_GPU_PROFILE)
GGML_UNUSED(ctx);
#endif
delete ctx;
delete backend;
}
static size_t ggml_webgpu_tensor_offset(const ggml_tensor * tensor) {
@ -896,8 +919,7 @@ static webgpu_command ggml_webgpu_pad(webgpu_context & ctx, ggml_tensor * src, g
ctx->pad_pipelines.emplace(pipeline_key, pipeline);
}
ggml_webgpu_generic_shader_decisions decisions =
*static_cast<ggml_webgpu_generic_shader_decisions *>(pipeline.context);
auto * decisions = static_cast<ggml_webgpu_generic_shader_decisions *>(pipeline.context.get());
const uint32_t ne = (uint32_t) ggml_nelements(dst);
@ -941,7 +963,7 @@ static webgpu_command ggml_webgpu_pad(webgpu_context & ctx, ggml_tensor * src, g
.size = ggml_webgpu_tensor_binding_size(ctx, dst) }
};
uint32_t wg_x = CEIL_DIV(ne, decisions.wg_size);
uint32_t wg_x = CEIL_DIV(ne, decisions->wg_size);
return ggml_backend_webgpu_build(ctx->global_ctx, ctx->param_buf_pool, pipeline, params, entries, wg_x);
}
@ -975,8 +997,7 @@ static std::optional<webgpu_command> ggml_webgpu_set_rows(webgpu_context & ctx,
ctx->set_rows_pipelines.emplace(key, pipeline);
}
ggml_webgpu_generic_shader_decisions decisions =
*static_cast<ggml_webgpu_generic_shader_decisions *>(pipeline.context);
auto * decisions = static_cast<ggml_webgpu_generic_shader_decisions *>(pipeline.context.get());
std::optional<webgpu_pool_bufs> error_bufs = std::nullopt;
if (key.i64_idx) {
@ -1028,7 +1049,7 @@ static std::optional<webgpu_command> ggml_webgpu_set_rows(webgpu_context & ctx,
} else {
threads = src->ne[0] * src->ne[1] * src->ne[2] * src->ne[3];
}
uint32_t wg_x = CEIL_DIV(threads, decisions.wg_size);
uint32_t wg_x = CEIL_DIV(threads, decisions->wg_size);
return ggml_backend_webgpu_build(ctx->global_ctx, ctx->param_buf_pool, pipeline, params, entries, wg_x, 1,
error_bufs);
}
@ -1297,10 +1318,9 @@ static webgpu_command ggml_webgpu_flash_attn(webgpu_context & ctx,
ctx->flash_attn_pipelines.emplace(key, pipeline);
}
ggml_webgpu_flash_attn_shader_decisions decisions =
*static_cast<ggml_webgpu_flash_attn_shader_decisions *>(pipeline.context);
auto * decisions = static_cast<ggml_webgpu_flash_attn_shader_decisions *>(pipeline.context.get());
uint32_t wg_per_head = CEIL_DIV(Q->ne[1], decisions.q_tile);
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->global_ctx, ctx->param_buf_pool, pipeline, params, entries, wg_x);
}
@ -1331,8 +1351,7 @@ static webgpu_command ggml_webgpu_unary_op(webgpu_context & ctx, ggml_tensor * s
ctx->unary_pipelines.emplace(pipeline_key, pipeline);
}
ggml_webgpu_generic_shader_decisions decisions =
*static_cast<ggml_webgpu_generic_shader_decisions *>(pipeline.context);
auto * decisions = static_cast<ggml_webgpu_generic_shader_decisions *>(pipeline.context.get());
uint32_t ne = (uint32_t) ggml_nelements(dst);
@ -1392,7 +1411,7 @@ static webgpu_command ggml_webgpu_unary_op(webgpu_context & ctx, ggml_tensor * s
.size = ggml_webgpu_tensor_binding_size(ctx, dst) });
}
uint32_t wg_x = CEIL_DIV(ne, decisions.wg_size);
uint32_t wg_x = CEIL_DIV(ne, decisions->wg_size);
return ggml_backend_webgpu_build(ctx->global_ctx, ctx->param_buf_pool, pipeline, params, entries, wg_x);
}
@ -1425,8 +1444,7 @@ static webgpu_command ggml_webgpu_binary_op(webgpu_context & ctx,
ctx->binary_pipelines.emplace(pipeline_key, pipeline);
}
ggml_webgpu_generic_shader_decisions decisions =
*static_cast<ggml_webgpu_generic_shader_decisions *>(pipeline.context);
auto * decisions = static_cast<ggml_webgpu_argsort_shader_decisions *>(pipeline.context.get());
uint32_t ne = (uint32_t) ggml_nelements(dst);
@ -1471,7 +1489,7 @@ static webgpu_command ggml_webgpu_binary_op(webgpu_context & ctx,
.size = ggml_webgpu_tensor_binding_size(ctx, dst) });
}
uint32_t wg_x = CEIL_DIV(ne, decisions.wg_size);
uint32_t wg_x = CEIL_DIV(ne, decisions->wg_size);
return ggml_backend_webgpu_build(ctx->global_ctx, ctx->param_buf_pool, pipeline, params, entries, wg_x);
}
@ -1821,8 +1839,7 @@ static webgpu_command ggml_webgpu_argsort(webgpu_context & ctx, ggml_tensor * sr
argsort_pipeline.context = processed.decisions;
ctx->argsort_pipelines.emplace(order, argsort_pipeline);
}
ggml_webgpu_argsort_shader_decisions argsort_decisions =
*static_cast<ggml_webgpu_argsort_shader_decisions *>(argsort_pipeline.context);
auto * argsort_decisions = static_cast<ggml_webgpu_argsort_shader_decisions *>(argsort_pipeline.context.get());
webgpu_pipeline argsort_merge_pipeline;
it = ctx->argsort_merge_pipelines.find(order);
@ -1839,13 +1856,13 @@ static webgpu_command ggml_webgpu_argsort(webgpu_context & ctx, ggml_tensor * sr
const uint32_t src_ne0 = (uint32_t) src->ne[0];
const uint32_t nrows = (uint32_t) ggml_nrows(src);
const uint32_t npr = CEIL_DIV(src_ne0, argsort_decisions.wg_size);
const uint32_t npr = CEIL_DIV(src_ne0, argsort_decisions->wg_size);
const uint32_t block_size =
is_top_k ? std::min(argsort_decisions.wg_size, (uint32_t) dst->ne[0]) : argsort_decisions.wg_size;
is_top_k ? std::min(argsort_decisions->wg_size, (uint32_t) dst->ne[0]) : argsort_decisions->wg_size;
uint32_t out_ne0 = src_ne0;
if (is_top_k) {
if (npr > 1) {
const uint32_t last_tile = src_ne0 - (npr - 1) * argsort_decisions.wg_size;
const uint32_t last_tile = src_ne0 - (npr - 1) * argsort_decisions->wg_size;
out_ne0 = (npr - 1) * block_size + std::min(last_tile, block_size);
} else {
out_ne0 = block_size;
@ -2198,7 +2215,10 @@ static ggml_backend_i ggml_backend_webgpu_i = {
static void ggml_backend_webgpu_buffer_free_buffer(ggml_backend_buffer_t buffer) {
ggml_backend_webgpu_buffer_context * ctx = static_cast<ggml_backend_webgpu_buffer_context *>(buffer->context);
ctx->buffer.Destroy();
if (ctx != nullptr && ctx->buffer != nullptr) {
ctx->buffer.Destroy();
delete ctx;
}
}
// Returns the "fake" base pointer.
@ -2926,12 +2946,12 @@ static bool create_webgpu_device(ggml_backend_webgpu_reg_context * ctx) {
dev_desc.SetDeviceLostCallback(
wgpu::CallbackMode::AllowSpontaneous,
[](const wgpu::Device & device, wgpu::DeviceLostReason reason, wgpu::StringView message) {
if (reason == wgpu::DeviceLostReason::Destroyed) {
return;
}
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<int>(reason),
//std::string(message).c_str());
GGML_LOG_ERROR("ggml_webgpu: Device lost! Reason: %d, Message: %s\n", static_cast<int>(reason),
std::string(message).c_str());
});
dev_desc.SetUncapturedErrorCallback(
[](const wgpu::Device & device, wgpu::ErrorType reason, wgpu::StringView message) {
@ -3365,10 +3385,7 @@ static size_t ggml_backend_webgpu_reg_get_device_count(ggml_backend_reg_t reg) {
return ctx->device_count;
}
// 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()");

66
gguf-py/gguf/constants.py
View File

@ -142,6 +142,7 @@ class Keys:
EMBEDDING_SCALE = "{arch}.embedding_scale"
TOKEN_SHIFT_COUNT = "{arch}.token_shift_count"
INTERLEAVE_MOE_LAYER_STEP = "{arch}.interleave_moe_layer_step"
FULL_ATTENTION_INTERVAL = "{arch}.full_attention_interval"
ACTIVATION_SPARSITY_SCALE = "{arch}.activation_sparsity_scale"
ALTUP_ACTIVE_IDX = "{arch}.altup.active_idx"
ALTUP_NUM_INPUTS = "{arch}.altup.num_inputs"
@ -384,6 +385,8 @@ class MODEL_ARCH(IntEnum):
QWEN3NEXT = auto()
QWEN3VL = auto()
QWEN3VLMOE = auto()
QWEN35 = auto()
QWEN35MOE = auto()
PHI2 = auto()
PHI3 = auto()
PHIMOE = auto()
@ -557,13 +560,14 @@ class MODEL_TENSOR(IntEnum):
SSM_D = auto()
SSM_NORM = auto()
SSM_OUT = auto()
SSM_ALPHA = auto() # qwen3.5
SSM_BETA_ALPHA = auto() # qwen3next
SSM_CONV1D_Q = auto() # Kimi Linear
SSM_CONV1D_K = auto() # Kimi Linear
SSM_CONV1D_V = auto() # Kimi Linear
SSM_F_A = auto() # Kimi Linear
SSM_F_B = auto() # Kimi Linear
SSM_BETA = auto() # Kimi Linear
SSM_BETA = auto() # Kimi Linear qwen3.5
SSM_G_A = auto() # Kimi Linear
SSM_G_B = auto() # Kimi Linear
TIME_MIX_W0 = auto()
@ -814,6 +818,8 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
MODEL_ARCH.QWEN3NEXT: "qwen3next",
MODEL_ARCH.QWEN3VL: "qwen3vl",
MODEL_ARCH.QWEN3VLMOE: "qwen3vlmoe",
MODEL_ARCH.QWEN35: "qwen35",
MODEL_ARCH.QWEN35MOE: "qwen35moe",
MODEL_ARCH.PHI2: "phi2",
MODEL_ARCH.PHI3: "phi3",
MODEL_ARCH.PHIMOE: "phimoe",
@ -985,13 +991,14 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
MODEL_TENSOR.SSM_D: "blk.{bid}.ssm_d",
MODEL_TENSOR.SSM_NORM: "blk.{bid}.ssm_norm",
MODEL_TENSOR.SSM_OUT: "blk.{bid}.ssm_out",
MODEL_TENSOR.SSM_ALPHA: "blk.{bid}.ssm_alpha", # qwen3.5
MODEL_TENSOR.SSM_BETA_ALPHA: "blk.{bid}.ssm_ba",
MODEL_TENSOR.SSM_CONV1D_Q: "blk.{bid}.ssm_conv1d_q", # Kimi Linear
MODEL_TENSOR.SSM_CONV1D_K: "blk.{bid}.ssm_conv1d_k", # Kimi Linear
MODEL_TENSOR.SSM_CONV1D_V: "blk.{bid}.ssm_conv1d_v", # Kimi Linear
MODEL_TENSOR.SSM_F_A: "blk.{bid}.ssm_f_a", # Kimi Linear
MODEL_TENSOR.SSM_F_B: "blk.{bid}.ssm_f_b", # Kimi Linear
MODEL_TENSOR.SSM_BETA: "blk.{bid}.ssm_beta", # Kimi Linear
MODEL_TENSOR.SSM_BETA: "blk.{bid}.ssm_beta", # Kimi Linear qwen3.5
MODEL_TENSOR.SSM_G_A: "blk.{bid}.ssm_g_a", # Kimi Linear
MODEL_TENSOR.SSM_G_B: "blk.{bid}.ssm_g_b", # Kimi Linear
MODEL_TENSOR.TIME_MIX_W0: "blk.{bid}.time_mix_w0",
@ -1818,6 +1825,61 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.FFN_DOWN_EXP,
MODEL_TENSOR.FFN_UP_EXP,
],
MODEL_ARCH.QWEN35: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
MODEL_TENSOR.OUTPUT,
MODEL_TENSOR.ATTN_NORM,
MODEL_TENSOR.ATTN_Q,
MODEL_TENSOR.ATTN_Q_NORM,
MODEL_TENSOR.ATTN_K,
MODEL_TENSOR.ATTN_K_NORM,
MODEL_TENSOR.ATTN_V,
MODEL_TENSOR.ATTN_OUT,
MODEL_TENSOR.ATTN_POST_NORM,
MODEL_TENSOR.ATTN_GATE,
MODEL_TENSOR.ATTN_QKV,
MODEL_TENSOR.FFN_GATE,
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_UP,
MODEL_TENSOR.SSM_A,
MODEL_TENSOR.SSM_CONV1D,
MODEL_TENSOR.SSM_DT,
MODEL_TENSOR.SSM_NORM,
MODEL_TENSOR.SSM_BETA,
MODEL_TENSOR.SSM_ALPHA,
MODEL_TENSOR.SSM_OUT
],
MODEL_ARCH.QWEN35MOE: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
MODEL_TENSOR.OUTPUT,
MODEL_TENSOR.ATTN_NORM,
MODEL_TENSOR.ATTN_Q,
MODEL_TENSOR.ATTN_Q_NORM,
MODEL_TENSOR.ATTN_K,
MODEL_TENSOR.ATTN_K_NORM,
MODEL_TENSOR.ATTN_V,
MODEL_TENSOR.ATTN_OUT,
MODEL_TENSOR.ATTN_POST_NORM,
MODEL_TENSOR.ATTN_GATE,
MODEL_TENSOR.ATTN_QKV,
MODEL_TENSOR.FFN_GATE_INP,
MODEL_TENSOR.FFN_GATE_INP_SHEXP,
MODEL_TENSOR.FFN_UP_SHEXP,
MODEL_TENSOR.FFN_DOWN_SHEXP,
MODEL_TENSOR.FFN_GATE_SHEXP,
MODEL_TENSOR.FFN_DOWN_EXP,
MODEL_TENSOR.FFN_UP_EXP,
MODEL_TENSOR.FFN_GATE_EXP,
MODEL_TENSOR.SSM_A,
MODEL_TENSOR.SSM_CONV1D,
MODEL_TENSOR.SSM_DT,
MODEL_TENSOR.SSM_NORM,
MODEL_TENSOR.SSM_BETA,
MODEL_TENSOR.SSM_ALPHA,
MODEL_TENSOR.SSM_OUT
],
MODEL_ARCH.PLAMO: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,

3
gguf-py/gguf/gguf_writer.py
View File

@ -708,6 +708,9 @@ class GGUFWriter:
def add_leading_dense_block_count(self, length: int) -> None:
self.add_uint32(Keys.LLM.LEADING_DENSE_BLOCK_COUNT.format(arch=self.arch), length)
def add_full_attention_interval(self, interval: int) -> None:
self.add_uint32(Keys.LLM.FULL_ATTENTION_INTERVAL.format(arch=self.arch), interval)
def add_feed_forward_length(self, length: int | Sequence[int]) -> None:
if isinstance(length, int):
self.add_uint32(Keys.LLM.FEED_FORWARD_LENGTH.format(arch=self.arch), length)

9
gguf-py/gguf/tensor_mapping.py
View File

@ -228,6 +228,7 @@ class TensorNameMap:
"transformer_encoder.{bid}.qkv", # neobert
"layers.{bid}.attn.Wqkv", # modern-bert
"model.layers.{bid}.self_attn.language_expert_query_key_value", # cogvlm
"model.layers.{bid}.linear_attn.in_proj_qkv", # qwen3.5
),
# Attention query
@ -359,6 +360,7 @@ class TensorNameMap:
MODEL_TENSOR.ATTN_GATE: (
"model.layers.{bid}.self_attn.gate_proj", # afmoe
"model.layers.{bid}.linear_attn.in_proj_z", # qwen3.5
"model.layers.{bid}.self_attn.g_proj", # step3.5 head-wise attention gate
),
@ -823,6 +825,10 @@ class TensorNameMap:
"model.layers.layers.{bid}.mixer.out_proj", # plamo2
),
MODEL_TENSOR.SSM_ALPHA: (
"model.layers.{bid}.linear_attn.in_proj_a", # qwen3.5
),
MODEL_TENSOR.SSM_BETA_ALPHA: (
"model.layers.{bid}.linear_attn.in_proj_ba", # qwen3next
),
@ -844,7 +850,8 @@ class TensorNameMap:
"model.layers.{bid}.self_attn.f_b_proj",
),
MODEL_TENSOR.SSM_BETA: (
"model.layers.{bid}.self_attn.b_proj",
"model.layers.{bid}.linear_attn.in_proj_b", # qwen3.5
"model.layers.{bid}.self_attn.b_proj", # Kimi Linear
),
MODEL_TENSOR.SSM_G_A: (
"model.layers.{bid}.self_attn.g_a_proj",

2
src/CMakeLists.txt
View File

@ -122,6 +122,8 @@ add_library(llama
models/qwen3vl-moe.cpp
models/qwen3moe.cpp
models/qwen3next.cpp
models/qwen35.cpp
models/qwen35moe.cpp
models/refact.cpp
models/rnd1.cpp
models/rwkv6-base.cpp

65
src/llama-arch.cpp
View File

@ -37,6 +37,8 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
{ LLM_ARCH_QWEN3NEXT, "qwen3next" },
{ LLM_ARCH_QWEN3VL, "qwen3vl" },
{ LLM_ARCH_QWEN3VLMOE, "qwen3vlmoe" },
{ LLM_ARCH_QWEN35, "qwen35" },
{ LLM_ARCH_QWEN35MOE, "qwen35moe" },
{ LLM_ARCH_PHI2, "phi2" },
{ LLM_ARCH_PHI3, "phi3" },
{ LLM_ARCH_PHIMOE, "phimoe" },
@ -195,6 +197,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
{ LLM_KV_EMBEDDING_SCALE, "%s.embedding_scale" },
{ LLM_KV_TOKEN_SHIFT_COUNT, "%s.token_shift_count" },
{ LLM_KV_INTERLEAVE_MOE_LAYER_STEP, "%s.interleave_moe_layer_step" },
{ LLM_KV_FULL_ATTENTION_INTERVAL, "%s.full_attention_interval" },
{ LLM_KV_ATTENTION_HEAD_COUNT, "%s.attention.head_count" },
{ LLM_KV_ATTENTION_HEAD_COUNT_KV, "%s.attention.head_count_kv" },
@ -366,6 +369,7 @@ static const std::map<llm_tensor, const char *> LLM_TENSOR_NAMES = {
{ LLM_TENSOR_SSM_CONV1D, "blk.%d.ssm_conv1d" },
{ LLM_TENSOR_SSM_DT, "blk.%d.ssm_dt" },
{ LLM_TENSOR_SSM_BETA_ALPHA, "blk.%d.ssm_ba" },
{ LLM_TENSOR_SSM_ALPHA, "blk.%d.ssm_alpha" },
{ LLM_TENSOR_SSM_IN, "blk.%d.ssm_in" },
{ LLM_TENSOR_SSM_NORM, "blk.%d.ssm_norm" },
{ LLM_TENSOR_SSM_OUT, "blk.%d.ssm_out" },
@ -968,7 +972,6 @@ static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
LLM_TENSOR_ATTN_OUT,
LLM_TENSOR_ATTN_QKV,
LLM_TENSOR_ATTN_GATE,
LLM_TENSOR_FFN_NORM,
LLM_TENSOR_FFN_GATE_INP,
LLM_TENSOR_FFN_GATE_EXPS,
LLM_TENSOR_FFN_DOWN_EXPS,
@ -985,6 +988,63 @@ static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
LLM_TENSOR_SSM_NORM,
LLM_TENSOR_SSM_OUT,
};
case LLM_ARCH_QWEN35:
return {
LLM_TENSOR_TOKEN_EMBD,
LLM_TENSOR_OUTPUT_NORM,
LLM_TENSOR_OUTPUT,
LLM_TENSOR_ATTN_NORM,
LLM_TENSOR_ATTN_POST_NORM,
LLM_TENSOR_ATTN_Q,
LLM_TENSOR_ATTN_Q_NORM,
LLM_TENSOR_ATTN_K,
LLM_TENSOR_ATTN_K_NORM,
LLM_TENSOR_ATTN_V,
LLM_TENSOR_ATTN_OUT,
LLM_TENSOR_ATTN_QKV,
LLM_TENSOR_ATTN_GATE,
LLM_TENSOR_FFN_GATE,
LLM_TENSOR_FFN_DOWN,
LLM_TENSOR_FFN_UP,
LLM_TENSOR_SSM_A_NOSCAN,
LLM_TENSOR_SSM_CONV1D,
LLM_TENSOR_SSM_DT,
LLM_TENSOR_SSM_BETA,
LLM_TENSOR_SSM_ALPHA,
LLM_TENSOR_SSM_NORM,
LLM_TENSOR_SSM_OUT,
};
case LLM_ARCH_QWEN35MOE:
return {
LLM_TENSOR_TOKEN_EMBD,
LLM_TENSOR_OUTPUT_NORM,
LLM_TENSOR_OUTPUT,
LLM_TENSOR_ATTN_NORM,
LLM_TENSOR_ATTN_POST_NORM,
LLM_TENSOR_ATTN_Q,
LLM_TENSOR_ATTN_Q_NORM,
LLM_TENSOR_ATTN_K,
LLM_TENSOR_ATTN_K_NORM,
LLM_TENSOR_ATTN_V,
LLM_TENSOR_ATTN_OUT,
LLM_TENSOR_ATTN_QKV,
LLM_TENSOR_ATTN_GATE,
LLM_TENSOR_FFN_GATE_INP,
LLM_TENSOR_FFN_GATE_EXPS,
LLM_TENSOR_FFN_DOWN_EXPS,
LLM_TENSOR_FFN_UP_EXPS,
LLM_TENSOR_FFN_GATE_INP_SHEXP,
LLM_TENSOR_FFN_GATE_SHEXP,
LLM_TENSOR_FFN_DOWN_SHEXP,
LLM_TENSOR_FFN_UP_SHEXP,
LLM_TENSOR_SSM_A_NOSCAN,
LLM_TENSOR_SSM_CONV1D,
LLM_TENSOR_SSM_DT,
LLM_TENSOR_SSM_BETA,
LLM_TENSOR_SSM_ALPHA,
LLM_TENSOR_SSM_NORM,
LLM_TENSOR_SSM_OUT,
};
case LLM_ARCH_QWEN3VL:
case LLM_ARCH_CHAMELEON:
case LLM_ARCH_HUNYUAN_DENSE:
@ -2456,6 +2516,7 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
{LLM_TENSOR_SSM_X, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_SSM_DT, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_SSM_OUT, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_SSM_ALPHA, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_SSM_BETA_ALPHA, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_TIME_MIX_W1, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
{LLM_TENSOR_TIME_MIX_W2, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
@ -2675,6 +2736,8 @@ bool llm_arch_is_hybrid(const llm_arch & arch) {
case LLM_ARCH_NEMOTRON_H_MOE:
case LLM_ARCH_QWEN3NEXT:
case LLM_ARCH_KIMI_LINEAR:
case LLM_ARCH_QWEN35:
case LLM_ARCH_QWEN35MOE:
return true;
default:
return false;

6
src/llama-arch.h
View File

@ -41,6 +41,8 @@ enum llm_arch {
LLM_ARCH_QWEN3NEXT,
LLM_ARCH_QWEN3VL,
LLM_ARCH_QWEN3VLMOE,
LLM_ARCH_QWEN35,
LLM_ARCH_QWEN35MOE,
LLM_ARCH_PHI2,
LLM_ARCH_PHI3,
LLM_ARCH_PHIMOE,
@ -199,6 +201,7 @@ enum llm_kv {
LLM_KV_EMBEDDING_SCALE,
LLM_KV_TOKEN_SHIFT_COUNT,
LLM_KV_INTERLEAVE_MOE_LAYER_STEP,
LLM_KV_FULL_ATTENTION_INTERVAL,
LLM_KV_ATTENTION_HEAD_COUNT,
LLM_KV_ATTENTION_HEAD_COUNT_KV,
@ -404,13 +407,14 @@ enum llm_tensor {
LLM_TENSOR_SSM_NORM,
LLM_TENSOR_SSM_OUT,
LLM_TENSOR_SSM_BETA_ALPHA, // qwen3next
LLM_TENSOR_SSM_ALPHA, // qwen3.5
// Kimi Linear KDA (using SSM_ prefix for consistency)
LLM_TENSOR_SSM_CONV1D_Q, // kimi: Q conv1d weight
LLM_TENSOR_SSM_CONV1D_K, // kimi: K conv1d weight
LLM_TENSOR_SSM_CONV1D_V, // kimi: V conv1d weight
LLM_TENSOR_SSM_F_A, // kimi: forget gate projection A
LLM_TENSOR_SSM_F_B, // kimi: forget gate projection B
LLM_TENSOR_SSM_BETA, // kimi: beta mixing coefficient
LLM_TENSOR_SSM_BETA, // kimi: beta mixing coefficient and qwen3.5
LLM_TENSOR_SSM_G_A, // kimi: output gate projection A
LLM_TENSOR_SSM_G_B, // kimi: output gate projection B
LLM_TENSOR_TIME_MIX_W0,

2
src/llama-context.cpp
View File

@ -2013,7 +2013,7 @@ void llama_context::output_reorder() {
//
uint32_t llama_context::graph_max_nodes(uint32_t n_tokens) const {
if (model.arch == LLM_ARCH_QWEN3NEXT || model.arch == LLM_ARCH_KIMI_LINEAR) {
if (model.arch == LLM_ARCH_QWEN3NEXT || model.arch == LLM_ARCH_KIMI_LINEAR || model.arch == LLM_ARCH_QWEN35 || model.arch == LLM_ARCH_QWEN35MOE) {
return std::max<uint32_t>(n_tokens * 40, 32u * model.n_tensors());
}
uint32_t res = std::max<uint32_t>(1024u, 8u*model.n_tensors());

202
src/llama-model.cpp
View File

@ -125,6 +125,7 @@ const char * llm_type_name(llm_type type) {
case LLM_TYPE_21B_A3B: return "21B.A3B";
case LLM_TYPE_30B_A3B: return "30B.A3B";
case LLM_TYPE_31B_A3_5B: return "31B.A3.5B";
case LLM_TYPE_35B_A3B: return "35B.A3B";
case LLM_TYPE_48B_A3B: return "48B.A3B";
case LLM_TYPE_80B_A3B: return "80B.A3B";
case LLM_TYPE_100B_A6B: return "100B.A6B";
@ -2403,8 +2404,12 @@ void llama_model::load_hparams(llama_model_loader & ml) {
ml.get_key(LLM_KV_SSM_GROUP_COUNT, hparams.ssm_n_group);
// Mark recurrent layers (linear attention layers)
for (uint32_t i = 0; i < hparams.n_layer; ++i) {
hparams.recurrent_layer_arr[i] = ((i + 1) % 4 != 0); // TODO: extract the magic 4 from "full_attention_interval"
{
uint32_t full_attn_interval = 4;
ml.get_key(LLM_KV_FULL_ATTENTION_INTERVAL, full_attn_interval, false);
for (uint32_t i = 0; i < hparams.n_layer; ++i) {
hparams.recurrent_layer_arr[i] = ((i + 1) % full_attn_interval != 0);
}
}
switch (hparams.n_layer) {
@ -2412,6 +2417,62 @@ void llama_model::load_hparams(llama_model_loader & ml) {
default: type = LLM_TYPE_UNKNOWN;
}
} break;
case LLM_ARCH_QWEN35:
{
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
ml.get_key_or_arr(LLM_KV_ROPE_DIMENSION_SECTIONS, hparams.rope_sections, 4, true);
// Load linear attention (gated delta net) parameters
ml.get_key(LLM_KV_SSM_CONV_KERNEL, hparams.ssm_d_conv);
ml.get_key(LLM_KV_SSM_INNER_SIZE, hparams.ssm_d_inner);
ml.get_key(LLM_KV_SSM_STATE_SIZE, hparams.ssm_d_state);
ml.get_key(LLM_KV_SSM_TIME_STEP_RANK, hparams.ssm_dt_rank);
ml.get_key(LLM_KV_SSM_GROUP_COUNT, hparams.ssm_n_group);
// Mark recurrent layers (linear attention layers)
{
uint32_t full_attn_interval = 4;
ml.get_key(LLM_KV_FULL_ATTENTION_INTERVAL, full_attn_interval, false);
for (uint32_t i = 0; i < hparams.n_layer; ++i) {
hparams.recurrent_layer_arr[i] = ((i + 1) % full_attn_interval != 0);
}
}
switch (hparams.n_layer) {
case 24: type = LLM_TYPE_2B; break;
default: type = LLM_TYPE_UNKNOWN;
}
} break;
case LLM_ARCH_QWEN35MOE:
{
ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp, false);
ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, false);
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
ml.get_key_or_arr(LLM_KV_ROPE_DIMENSION_SECTIONS, hparams.rope_sections, 4, true);
// Load linear attention (gated delta net) parameters
ml.get_key(LLM_KV_SSM_CONV_KERNEL, hparams.ssm_d_conv);
ml.get_key(LLM_KV_SSM_INNER_SIZE, hparams.ssm_d_inner);
ml.get_key(LLM_KV_SSM_STATE_SIZE, hparams.ssm_d_state);
ml.get_key(LLM_KV_SSM_TIME_STEP_RANK, hparams.ssm_dt_rank);
ml.get_key(LLM_KV_SSM_GROUP_COUNT, hparams.ssm_n_group);
// Mark recurrent layers (linear attention layers)
{
uint32_t full_attn_interval = 4;
ml.get_key(LLM_KV_FULL_ATTENTION_INTERVAL, full_attn_interval, false);
for (uint32_t i = 0; i < hparams.n_layer; ++i) {
hparams.recurrent_layer_arr[i] = ((i + 1) % full_attn_interval != 0);
}
}
switch (hparams.n_layer) {
case 28: type = LLM_TYPE_35B_A3B; break;
case 48: type = LLM_TYPE_80B_A3B; break;
default: type = LLM_TYPE_UNKNOWN;
}
} break;
case LLM_ARCH_MISTRAL3:
{
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
@ -7101,6 +7162,131 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), { hparams.n_ff_shexp, n_embd }, 0);
}
} break;
case LLM_ARCH_QWEN35MOE:
{
tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, 0);
// output
output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), { n_embd }, 0);
output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), { n_embd, n_vocab }, TENSOR_NOT_REQUIRED);
// if output is NULL, init from the input tok embed
if (output == NULL) {
output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, TENSOR_DUPLICATED);
}
const int64_t n_ff_exp = hparams.n_ff_exp ? hparams.n_ff_exp : n_ff / n_expert_used;
// Calculate dimensions from hyperparameters
const int64_t head_k_dim = hparams.ssm_d_state;
const int64_t head_v_dim = hparams.ssm_d_state;
const int64_t n_k_heads = hparams.ssm_n_group;
const int64_t n_v_heads = hparams.ssm_dt_rank;
const int64_t key_dim = head_k_dim * n_k_heads;
const int64_t value_dim = head_v_dim * n_v_heads;
const int64_t conv_dim = key_dim * 2 + value_dim;
for (int i = 0; i < n_layer; ++i) {
auto & layer = layers[i];
layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), { n_embd }, 0);
layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), { n_embd }, 0);
if (!hparams.is_recurrent(i)) {
// Attention layers
layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), { n_embd, n_embd_head_k * n_head * 2 }, 0);
layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), { n_embd, n_embd_k_gqa }, 0);
layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), { n_embd, n_embd_v_gqa }, 0);
layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd_head_k * n_head, n_embd }, 0);
// Q/K normalization for attention layers
layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), { n_embd_head_k }, 0);
layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), { n_embd_head_k }, 0);
} else {
// Linear attention (gated delta net) specific tensors
// Create tensors with calculated dimensions
layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), { n_embd, key_dim * 2 + value_dim }, TENSOR_NOT_REQUIRED);
layer.wqkv_gate = create_tensor(tn(LLM_TENSOR_ATTN_GATE, "weight", i), { n_embd, value_dim }, TENSOR_NOT_REQUIRED);
layer.ssm_conv1d = create_tensor(tn(LLM_TENSOR_SSM_CONV1D, "weight", i), { hparams.ssm_d_conv, conv_dim }, 0);
layer.ssm_dt = create_tensor(tn(LLM_TENSOR_SSM_DT, "bias", i), { hparams.ssm_dt_rank }, 0);
layer.ssm_a = create_tensor(tn(LLM_TENSOR_SSM_A_NOSCAN, i), { hparams.ssm_dt_rank }, 0);
layer.ssm_beta = create_tensor(tn(LLM_TENSOR_SSM_BETA, "weight", i), { n_embd, n_v_heads }, 0);
layer.ssm_alpha = create_tensor(tn(LLM_TENSOR_SSM_ALPHA, "weight", i), { n_embd, n_v_heads }, 0);
layer.ssm_norm = create_tensor(tn(LLM_TENSOR_SSM_NORM, "weight", i), { head_v_dim }, 0);
layer.ssm_out = create_tensor(tn(LLM_TENSOR_SSM_OUT, "weight", i), { value_dim, n_embd }, 0);
}
layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), { n_embd, n_expert }, 0);
layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), { n_embd, n_ff_exp, n_expert }, 0);
layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), { n_ff_exp, n_embd, n_expert }, 0);
layer.ffn_up_exps = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), { n_embd, n_ff_exp, n_expert }, 0);
// Shared experts
const int64_t n_ff_shexp = hparams.n_ff_shexp ? hparams.n_ff_shexp : n_ff;
layer.ffn_gate_inp_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP_SHEXP, "weight", i), { n_embd }, 0);
layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), { n_embd, n_ff_shexp }, 0);
layer.ffn_up_shexp = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP, "weight", i), { n_embd, n_ff_shexp }, 0);
layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), { n_ff_shexp, n_embd }, 0);
}
} break;
case LLM_ARCH_QWEN35:
{
tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, 0);
// output
output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), { n_embd }, 0);
output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), { n_embd, n_vocab }, TENSOR_NOT_REQUIRED);
// if output is NULL, init from the input tok embed
if (output == NULL) {
output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, TENSOR_DUPLICATED);
}
// Calculate dimensions from hyperparameters
const int64_t head_k_dim = hparams.ssm_d_state;
const int64_t head_v_dim = hparams.ssm_d_state;
const int64_t n_k_heads = hparams.ssm_n_group;
const int64_t n_v_heads = hparams.ssm_dt_rank;
const int64_t key_dim = head_k_dim * n_k_heads;
const int64_t value_dim = head_v_dim * n_v_heads;
const int64_t conv_dim = key_dim * 2 + value_dim;
for (int i = 0; i < n_layer; ++i) {
auto & layer = layers[i];
layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), { n_embd }, 0);
layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), { n_embd }, 0);
if (!hparams.is_recurrent(i)) {
// Attention layers
layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), { n_embd, n_embd_head_k * n_head * 2 }, 0);
layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), { n_embd, n_embd_k_gqa }, 0);
layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), { n_embd, n_embd_v_gqa }, 0);
layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd_head_k * n_head, n_embd }, 0);
// Q/K normalization for attention layers
layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), { n_embd_head_k }, 0);
layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), { n_embd_head_k }, 0);
} else {
// Linear attention (gated delta net) specific tensors
// Create tensors with calculated dimensions
layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), { n_embd, key_dim * 2 + value_dim }, TENSOR_NOT_REQUIRED);
layer.wqkv_gate = create_tensor(tn(LLM_TENSOR_ATTN_GATE, "weight", i), { n_embd, value_dim }, TENSOR_NOT_REQUIRED);
layer.ssm_conv1d = create_tensor(tn(LLM_TENSOR_SSM_CONV1D, "weight", i), { hparams.ssm_d_conv, conv_dim }, 0);
layer.ssm_dt = create_tensor(tn(LLM_TENSOR_SSM_DT, "bias", i), { hparams.ssm_dt_rank }, 0);
layer.ssm_a = create_tensor(tn(LLM_TENSOR_SSM_A_NOSCAN, i), { hparams.ssm_dt_rank }, 0);
layer.ssm_beta = create_tensor(tn(LLM_TENSOR_SSM_BETA, "weight", i), { n_embd, n_v_heads }, 0);
layer.ssm_alpha = create_tensor(tn(LLM_TENSOR_SSM_ALPHA, "weight", i), { n_embd, n_v_heads }, 0);
layer.ssm_norm = create_tensor(tn(LLM_TENSOR_SSM_NORM, "weight", i), { head_v_dim }, 0);
layer.ssm_out = create_tensor(tn(LLM_TENSOR_SSM_OUT, "weight", i), { value_dim, n_embd }, 0);
}
layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0);
layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, 0);
layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, 0);
}
} break;
case LLM_ARCH_MIMO2:
{
tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
@ -7545,6 +7731,8 @@ void llama_model::print_info() const {
arch == LLM_ARCH_PLAMO2 ||
arch == LLM_ARCH_GRANITE_HYBRID ||
arch == LLM_ARCH_QWEN3NEXT ||
arch == LLM_ARCH_QWEN35 ||
arch == LLM_ARCH_QWEN35MOE ||
arch == LLM_ARCH_NEMOTRON_H ||
arch == LLM_ARCH_NEMOTRON_H_MOE) {
LLAMA_LOG_INFO("%s: ssm_d_conv = %u\n", __func__, hparams.ssm_d_conv);
@ -8343,6 +8531,14 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
{
llm = std::make_unique<llm_build_qwen3next>(*this, params);
} break;
case LLM_ARCH_QWEN35:
{
llm = std::make_unique<llm_build_qwen35>(*this, params);
} break;
case LLM_ARCH_QWEN35MOE:
{
llm = std::make_unique<llm_build_qwen35moe>(*this, params);
} break;
case LLM_ARCH_MISTRAL3:
{
llm = std::make_unique<llm_build_mistral3>(*this, params);
@ -8611,6 +8807,8 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
return LLAMA_ROPE_TYPE_MROPE;
case LLM_ARCH_QWEN3VL:
case LLM_ARCH_QWEN3VLMOE:
case LLM_ARCH_QWEN35:
case LLM_ARCH_QWEN35MOE:
return LLAMA_ROPE_TYPE_IMROPE;
case LLM_ARCH_GLM4:

4
src/llama-model.h
View File

@ -118,6 +118,7 @@ enum llm_type {
LLM_TYPE_21B_A3B, // Ernie MoE small
LLM_TYPE_30B_A3B,
LLM_TYPE_31B_A3_5B,
LLM_TYPE_35B_A3B, // Qwen3.5
LLM_TYPE_48B_A3B, // Kimi Linear
LLM_TYPE_80B_A3B, // Qwen3 Next
LLM_TYPE_100B_A6B,
@ -322,6 +323,9 @@ struct llama_layer {
// qwen3next
struct ggml_tensor * ssm_beta_alpha = nullptr;
// qwen3.5
struct ggml_tensor * ssm_alpha = nullptr;
// rwkv
struct ggml_tensor * time_mix_w1 = nullptr;
struct ggml_tensor * time_mix_w2 = nullptr;

11
src/llama-vocab.cpp
View File

@ -368,6 +368,13 @@ struct llm_tokenizer_bpe : llm_tokenizer {
"(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
};
break;
case LLAMA_VOCAB_PRE_TYPE_QWEN35:
regex_exprs = {
// original regex from tokenizer.json
// "(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?[\\p{L}\\p{M}]+|\\p{N}| ?[^\\s\\p{L}\\p{M}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+"
"(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?[\\p{L}\\p{M}]+|\\p{N}| ?[^\\s\\p{L}\\p{M}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
};
break;
case LLAMA_VOCAB_PRE_TYPE_PORO:
case LLAMA_VOCAB_PRE_TYPE_BLOOM:
case LLAMA_VOCAB_PRE_TYPE_GPT3_FINNISH:
@ -1926,6 +1933,10 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
tokenizer_pre == "kormo") {
pre_type = LLAMA_VOCAB_PRE_TYPE_QWEN2;
clean_spaces = false;
} else if (
tokenizer_pre == "qwen35") {
pre_type = LLAMA_VOCAB_PRE_TYPE_QWEN35;
clean_spaces = false;
} else if (
tokenizer_pre == "stablelm2") {
pre_type = LLAMA_VOCAB_PRE_TYPE_STABLELM2;

1
src/llama-vocab.h
View File

@ -54,6 +54,7 @@ enum llama_vocab_pre_type {
LLAMA_VOCAB_PRE_TYPE_SOLAR_OPEN = 43,
LLAMA_VOCAB_PRE_TYPE_YOUTU = 44,
LLAMA_VOCAB_PRE_TYPE_EXAONE_MOE = 45,
LLAMA_VOCAB_PRE_TYPE_QWEN35 = 46,
};
struct LLM_KV;

119
src/models/models.h
View File

@ -476,6 +476,7 @@ struct llm_build_qwen3vl : public llm_graph_context {
struct llm_build_qwen3vlmoe : public llm_graph_context {
llm_build_qwen3vlmoe(const llama_model & model, const llm_graph_params & params);
};
struct llm_build_qwen3next : public llm_graph_context_mamba {
llm_build_qwen3next(const llama_model & model, const llm_graph_params & params);
private:
@ -534,6 +535,124 @@ private:
const llama_model & model;
};
struct llm_build_qwen35 : public llm_graph_context_mamba {
llm_build_qwen35(const llama_model & model, const llm_graph_params & params);
private:
ggml_tensor * build_layer_attn(
llm_graph_input_attn_kv * inp_attn,
ggml_tensor * cur,
ggml_tensor * inp_pos,
int * sections,
int il);
ggml_tensor * build_layer_attn_linear(
llm_graph_input_rs * inp,
ggml_tensor * cur,
ggml_tensor * causal_mask,
ggml_tensor * identity,
ggml_tensor * diag_mask,
int il);
ggml_tensor * build_layer_ffn(
ggml_tensor * cur,
int il);
// returns pair of output and new state
std::pair<ggml_tensor *, ggml_tensor *> build_delta_net_chunking(
ggml_tensor * q,
ggml_tensor * k,
ggml_tensor * v,
ggml_tensor * g,
ggml_tensor * beta,
ggml_tensor * state,
ggml_tensor * causal_mask,
ggml_tensor * identity,
ggml_tensor * diag_mask,
int il);
// returns pair of output and new state
std::pair<ggml_tensor *, ggml_tensor *> build_delta_net_autoregressive(
ggml_tensor * q,
ggml_tensor * k,
ggml_tensor * v,
ggml_tensor * g,
ggml_tensor * beta,
ggml_tensor * state,
int il);
ggml_tensor * build_norm_gated(
ggml_tensor * input,
ggml_tensor * weights,
ggml_tensor * gate,
int layer);
// returns pair of qkv, z
std::pair<ggml_tensor *, ggml_tensor *> build_qkvz(
ggml_tensor * input,
int il);
const llama_model & model;
};
struct llm_build_qwen35moe : public llm_graph_context_mamba {
llm_build_qwen35moe(const llama_model & model, const llm_graph_params & params);
private:
ggml_tensor * build_layer_attn(
llm_graph_input_attn_kv * inp_attn,
ggml_tensor * cur,
ggml_tensor * inp_pos,
int * sections,
int il);
ggml_tensor * build_layer_attn_linear(
llm_graph_input_rs * inp,
ggml_tensor * cur,
ggml_tensor * causal_mask,
ggml_tensor * identity,
ggml_tensor * diag_mask,
int il);
ggml_tensor * build_layer_ffn(
ggml_tensor * cur,
int il);
// returns pair of output and new state
std::pair<ggml_tensor *, ggml_tensor *> build_delta_net_chunking(
ggml_tensor * q,
ggml_tensor * k,
ggml_tensor * v,
ggml_tensor * g,
ggml_tensor * beta,
ggml_tensor * state,
ggml_tensor * causal_mask,
ggml_tensor * identity,
ggml_tensor * diag_mask,
int il);
// returns pair of output and new state
std::pair<ggml_tensor *, ggml_tensor *> build_delta_net_autoregressive(
ggml_tensor * q,
ggml_tensor * k,
ggml_tensor * v,
ggml_tensor * g,
ggml_tensor * beta,
ggml_tensor * state,
int il);
ggml_tensor * build_norm_gated(
ggml_tensor * input,
ggml_tensor * weights,
ggml_tensor * gate,
int layer);
// returns pair of qkv, z
std::pair<ggml_tensor *, ggml_tensor *> build_qkvz(
ggml_tensor * input,
int il);
const llama_model & model;
};
struct llm_build_qwen : public llm_graph_context {
llm_build_qwen(const llama_model & model, const llm_graph_params & params);
};

740
src/models/qwen35.cpp Normal file
View File

@ -0,0 +1,740 @@
#include "ggml.h"
#include "models.h"
#define CHUNK_SIZE 64
llm_build_qwen35::llm_build_qwen35(const llama_model & model, const llm_graph_params & params) :
llm_graph_context_mamba(params), model(model) {
const int64_t n_embd_head = hparams.n_embd_head_v;
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
int sections[4];
std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections);
ggml_tensor * cur;
ggml_tensor * inpL;
inpL = build_inp_embd(model.tok_embd);
cb(inpL, "model.input_embed", -1);
auto * inp = build_inp_mem_hybrid();
ggml_tensor * inp_pos = build_inp_pos();
ggml_tensor * inp_out_ids = build_inp_out_ids();
ggml_tensor * causal_mask =
ggml_tri(ctx0, ggml_fill(ctx0, ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, CHUNK_SIZE, CHUNK_SIZE), 1.0f),
GGML_TRI_TYPE_LOWER);
ggml_tensor * identity = ggml_diag(ctx0, ggml_fill(ctx0, ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, CHUNK_SIZE), 1.0f));
ggml_tensor * diag_mask = ggml_add(ctx0, causal_mask, identity);
ggml_build_forward_expand(gf, causal_mask);
ggml_build_forward_expand(gf, identity);
ggml_build_forward_expand(gf, diag_mask);
for (int il = 0; il < n_layer; ++il) {
ggml_tensor * inpSA = inpL;
cur = build_norm(inpL, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il);
cb(cur, "attn_norm", il);
// Determine layer type and build appropriate attention mechanism
if (hparams.is_recurrent(il)) {
// Linear attention layer (gated delta net)
cur = build_layer_attn_linear(inp->get_recr(), cur, causal_mask, identity, diag_mask, il);
} else {
// Full attention layer
cur = build_layer_attn(inp->get_attn(), cur, inp_pos, sections, il);
}
if (il == n_layer - 1 && inp_out_ids) {
cur = ggml_get_rows(ctx0, cur, inp_out_ids);
inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
}
// Residual connection
cur = ggml_add(ctx0, cur, inpSA);
cb(cur, "attn_residual", il);
// Save the tensor before post-attention norm for residual connection
ggml_tensor * ffn_residual = cur;
// Post-attention norm
ggml_tensor * attn_post_norm = build_norm(cur, model.layers[il].attn_post_norm, nullptr, LLM_NORM_RMS, il);
cb(attn_post_norm, "attn_post_norm", il);
// Dense FFN layer - without residual connection
cur = build_layer_ffn(attn_post_norm, il);
cb(cur, "ffn_out", il);
// Residual connection for FFN - add to the tensor from before post_attention_layernorm
cur = ggml_add(ctx0, cur, ffn_residual);
cb(cur, "post_ffn", il);
// Input for next layer
inpL = cur;
}
cur = inpL;
// Final norm
cur = build_norm(cur, model.output_norm, nullptr, LLM_NORM_RMS, -1);
cb(cur, "result_norm", -1);
res->t_embd = cur;
// LM head
cur = build_lora_mm(model.output, cur);
cb(cur, "result_output", -1);
res->t_logits = cur;
ggml_build_forward_expand(gf, cur);
}
// utility to get one slice from the third dimension
// input dim: [x, y, c, b]
// output dim: [x, y, 1, b]
static ggml_tensor * get_slice_2d(ggml_context * ctx0, ggml_tensor * t, int64_t c) {
return ggml_view_4d(ctx0, t, t->ne[0], t->ne[1], 1, t->ne[3],
t->nb[1], t->nb[2], t->nb[3], t->nb[2] * c);
}
std::pair<ggml_tensor *, ggml_tensor *> llm_build_qwen35::build_delta_net_chunking(
ggml_tensor * q,
ggml_tensor * k,
ggml_tensor * v,
ggml_tensor * g,
ggml_tensor * beta,
ggml_tensor * state,
ggml_tensor * causal_mask,
ggml_tensor * identity,
ggml_tensor * diag_mask,
int il) {
const int64_t S_k = q->ne[0];
const int64_t H_k = q->ne[1];
const int64_t n_tokens = q->ne[2];
const int64_t n_seqs = q->ne[3];
const int64_t S_v = v->ne[0];
const int64_t H_v = v->ne[1];
GGML_ASSERT(v->ne[2] == n_tokens);
GGML_ASSERT(k->ne[2] == n_tokens);
GGML_ASSERT(g->ne[0] == H_v && g->ne[1] == n_tokens && g->ne[2] == n_seqs);
GGML_ASSERT(beta->ne[0] == H_v && beta->ne[2] == n_tokens && beta->ne[3] == n_seqs);
GGML_ASSERT(state->ne[0] == S_v && state->ne[1] == S_v * H_v && state->ne[2] == 1 && state->ne[3] == n_seqs);
GGML_ASSERT(q->ne[0] == S_k && q->ne[1] == H_k && q->ne[2] == n_tokens && q->ne[3] == n_seqs);
GGML_ASSERT(k->ne[0] == S_k && k->ne[1] == H_k && k->ne[2] == n_tokens && k->ne[3] == n_seqs);
GGML_ASSERT(H_k == H_v); // we did a repeat to make sure this is the case
const float eps_norm = hparams.f_norm_rms_eps;
q = ggml_l2_norm(ctx0, q, eps_norm);
k = ggml_l2_norm(ctx0, k, eps_norm);
const float scale = 1.0f / sqrtf(S_v);
q = ggml_scale(ctx0, q, scale);
beta = ggml_sigmoid(ctx0, beta);
cb(q, "q_in", il);
cb(k, "k_in", il);
cb(v, "v_in", il);
cb(beta, "beta_in", il);
cb(g, "g_in", il);
q = ggml_cont_4d(ctx0, ggml_permute(ctx0, q, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs);
k = ggml_cont_4d(ctx0, ggml_permute(ctx0, k, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs);
v = ggml_cont_4d(ctx0, ggml_permute(ctx0, v, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs);
g = ggml_cont_4d(ctx0, ggml_permute(ctx0, g, 2, 0, 3, 1), n_tokens, 1, H_k, n_seqs);
beta = ggml_cont(ctx0, ggml_permute(ctx0, beta, 2, 0, 1, 3));
state = ggml_reshape_4d(ctx0, state, S_v, S_v, H_v, n_seqs);
cb(q, "q_perm", il);
cb(k, "k_perm", il);
cb(v, "v_perm", il);
cb(beta, "beta_perm", il);
cb(g, "g_perm", il);
cb(state, "state_in", il);
GGML_ASSERT(q->ne[1] == n_tokens && q->ne[0] == S_k && q->ne[2] == H_k && q->ne[3] == n_seqs);
GGML_ASSERT(k->ne[1] == n_tokens && k->ne[0] == S_k && k->ne[2] == H_k && k->ne[3] == n_seqs);
GGML_ASSERT(v->ne[1] == n_tokens && v->ne[0] == S_v && v->ne[2] == H_k && v->ne[3] == n_seqs);
GGML_ASSERT(beta->ne[1] == n_tokens && beta->ne[2] == H_k && beta->ne[0] == 1 && beta->ne[3] == n_seqs);
// Do padding
const int64_t chunk_size = CHUNK_SIZE;
const int64_t pad = (chunk_size - n_tokens % chunk_size) % chunk_size;
const int64_t n_chunks = (n_tokens + pad) / chunk_size;
q = ggml_pad(ctx0, q, 0, pad, 0, 0);
k = ggml_pad(ctx0, k, 0, pad, 0, 0);
v = ggml_pad(ctx0, v, 0, pad, 0, 0);
g = ggml_pad(ctx0, g, pad, 0, 0, 0);
beta = ggml_pad(ctx0, beta, 0, pad, 0, 0);
cb(q, "q_pad", il);
cb(k, "k_pad", il);
cb(v, "v_pad", il);
cb(beta, "beta_pad", il);
cb(g, "g_pad", il);
ggml_tensor * v_beta = ggml_mul(ctx0, v, beta);
ggml_tensor * k_beta = ggml_mul(ctx0, k, beta);
cb(v_beta, "v_beta", il);
cb(k_beta, "k_beta", il);
q = ggml_reshape_4d(ctx0, q, S_k, chunk_size, n_chunks, H_k * n_seqs);
k = ggml_reshape_4d(ctx0, k, S_k, chunk_size, n_chunks, H_k * n_seqs);
k_beta = ggml_reshape_4d(ctx0, k_beta, S_k, chunk_size, n_chunks, H_k * n_seqs);
v = ggml_reshape_4d(ctx0, v, S_v, chunk_size, n_chunks, H_v * n_seqs);
v_beta = ggml_reshape_4d(ctx0, v_beta, S_v, chunk_size, n_chunks, H_v * n_seqs);
g = ggml_reshape_4d(ctx0, g, chunk_size, 1, n_chunks, H_k * n_seqs);
beta = ggml_reshape_4d(ctx0, beta, 1, chunk_size, n_chunks, H_k * n_seqs);
ggml_tensor * g_cumsum = ggml_cumsum(ctx0, g);
cb(g_cumsum, "g_cumsum", il); // shape: (chunk_size, 1, n_chunks, H_v * n_seqs)
ggml_tensor * gcs_i = g_cumsum; // ggml_reshape_4d(ctx0, g_cumsum, chunk_size, 1, n_chunks, H_v * n_seqs);
ggml_tensor * gcs_j = ggml_reshape_4d(ctx0, g_cumsum, 1, chunk_size, n_chunks, H_v * n_seqs);
ggml_tensor * gcs_j_broadcast =
ggml_repeat_4d(ctx0, gcs_j, chunk_size, chunk_size, n_chunks, H_v * n_seqs);
ggml_tensor * decay_mask = ggml_sub(ctx0, gcs_j_broadcast, gcs_i);
cb(decay_mask, "decay_mask", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs)
decay_mask = ggml_mul(ctx0, decay_mask, diag_mask);
decay_mask = ggml_exp(ctx0, decay_mask);
decay_mask = ggml_mul(ctx0, decay_mask, diag_mask);
ggml_tensor * kmulkbeta = ggml_mul_mat(ctx0, k, k_beta);
ggml_tensor * k_decay = ggml_mul(ctx0, kmulkbeta, decay_mask);
ggml_tensor * attn = ggml_neg(ctx0, ggml_mul(ctx0, k_decay, causal_mask));
cb(attn, "attn_pre_solve", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs)
ggml_tensor * attn_lower = ggml_mul(ctx0, attn, causal_mask);
ggml_tensor * lhs = ggml_sub(ctx0, ggml_repeat(ctx0, identity, attn_lower), attn_lower);
ggml_tensor * lin_solve = ggml_solve_tri(ctx0, lhs, attn, true, true, false);
attn = ggml_mul(ctx0, lin_solve, causal_mask);
attn = ggml_add(ctx0, attn, identity);
cb(attn, "attn_solved", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs)
v = ggml_mul_mat(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, v_beta)), attn);
ggml_tensor * g_cumsum_t = ggml_cont(ctx0, ggml_transpose(ctx0, g_cumsum));
ggml_tensor * gexp = ggml_exp(ctx0, g_cumsum_t);
ggml_tensor * kbeta_gexp = ggml_mul(ctx0, k_beta, gexp);
cb(kbeta_gexp, "kbeta_gexp", il); // shape: (S_k, chunk_size, n_chunks, H_v * n_seqs)
ggml_tensor * k_cumdecay =
ggml_cont(ctx0, ggml_transpose(ctx0, ggml_mul_mat(ctx0, attn, ggml_cont(ctx0, ggml_transpose(ctx0, kbeta_gexp)))));
cb(k_cumdecay, "k_cumdecay", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs)
ggml_tensor * attn_kq = ggml_mul_mat(ctx0, k, q);
attn_kq = ggml_mul(ctx0, attn_kq, decay_mask);
attn_kq = ggml_mul(ctx0, attn_kq, diag_mask);
cb(attn_kq, "attn_kq", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs)
// vectorized calculation of key_gdiff
// improved from the chunked version:
// g_last = torch.clamp(g_cum[:, :, -1], max=50.0).exp().unsqueeze(-1).unsqueeze(-1)
// g_diff = torch.clamp(g_cum[:, :, -1:] - g_cum, max=50.0).exp()
// key_gdiff = key * g_diff.unsqueeze(-1)
// kgdmulvnew = (key_gdiff).transpose(-1, -2) @ v_new
// last_recurrent_state = last_recurrent_state * g_last + kgdmulvnew
// get last element in g_cumsum along chunk_size dimension (ne0)
// example: [[x, y, z, ..., last], ...] -> [[last], ...]
ggml_tensor * g_last = ggml_view_4d(ctx0, g_cumsum, 1, 1, g_cumsum->ne[2], g_cumsum->ne[3],
g_cumsum->nb[1], g_cumsum->nb[2], g_cumsum->nb[3],
(g_cumsum->ne[0] - 1) * ggml_element_size(g_cumsum));
g_last = ggml_cont(ctx0, g_last);
cb(g_last, "g_last", il); // shape: (1, 1, n_chunks, H_v * n_seqs)
ggml_tensor * g_last_exp = ggml_exp(ctx0, g_last);
cb(g_last_exp, "g_last_exp", il); // shape: (1, 1, n_chunks, H_v * n_seqs)
ggml_tensor * g_diff = ggml_neg(ctx0, ggml_sub(ctx0, g_cumsum, g_last));
cb(g_diff, "g_diff", il); // shape: (chunk_size, 1, n_chunks, H_v * n_seqs)
ggml_tensor * g_diff_exp = ggml_exp(ctx0, g_diff);
ggml_tensor * g_diff_exp_t = ggml_reshape_4d(ctx0, g_diff_exp,
1, chunk_size, n_chunks, g_diff_exp->ne[3]);
ggml_tensor * key_gdiff = ggml_mul(ctx0, k, g_diff_exp_t);
cb(key_gdiff, "key_gdiff", il); // shape: (S_k, chunk_size, n_chunks, H_v * n_seqs)
ggml_tensor * key_gdiff_t = ggml_cont(ctx0, ggml_transpose(ctx0, key_gdiff));
cb(key_gdiff_t, "key_gdiff_t", il); // shape: (chunk_size, S_k, n_chunks, H_v * n_seqs)
// state to be updated per chunk
ggml_tensor * new_state = state; // ggml_dup(ctx0, state);
cb(new_state, "new_state", il); // shape: (S_v, S_v, H_v, n_seqs)
// shape after loop of chunks: (S_v, chunk_size, n_chunks, H_v * n_seqs)
ggml_tensor * core_attn_out = nullptr;
for (int64_t chunk = 0; chunk < n_chunks; chunk++) {
// shape: (S_k, chunk_size, 1, H_k * n_seqs)
ggml_tensor * q_chunk = get_slice_2d(ctx0, q, chunk); // (no cont), next op: ggml_mul
// shape: (S_v, chunk_size, 1, H_v * n_seqs)
ggml_tensor * v_chunk = get_slice_2d(ctx0, v, chunk); // (no cont), next op: ggml_repeat
// shape: (chunk_size, 1, n_chunks, H_v * n_seqs)
ggml_tensor * gexp_chunk = get_slice_2d(ctx0, gexp, chunk); // (no cont), next op: ggml_mul
// shape: (chunk_size, 1, H_v * n_seqs)
ggml_tensor * k_cumdecay_chunk = get_slice_2d(ctx0, k_cumdecay, chunk); // (no cont), next op: ggml_mul_mat
// attn = (q_i @ k_i.transpose(-1, -2) * decay_mask[:, :, i]).masked_fill_(mask, 0)
// replaced by precomputed attn_kq
ggml_tensor * attn_chunk = get_slice_2d(ctx0, attn_kq, chunk);
cb(attn_chunk, "attn_chunk", il);
ggml_tensor * state_t = ggml_cont_4d(ctx0, ggml_permute(ctx0, new_state, 1, 0, 2, 3), S_v, S_v, 1, H_v * n_seqs);
// v_prime = (k_cumdecay[:, :, i]) @ last_recurrent_state
ggml_tensor * v_prime = ggml_mul_mat(ctx0, state_t, k_cumdecay_chunk);
cb(v_prime, "v_prime_chunk", il); // shape: (S_v, 1, H_v * n_seqs)
// v_new = v_i - v_prime
ggml_tensor * v_new = ggml_sub(ctx0, ggml_repeat(ctx0, v_chunk, v_prime), v_prime);
ggml_tensor * v_new_t = ggml_cont(ctx0, ggml_transpose(ctx0, v_new));
cb(v_new, "v_new_chunk", il);
// attn_inter = (q_i * g[:, :, i, :, None].exp()) @ last_recurrent_state
ggml_tensor * q_g_exp = ggml_mul(ctx0, q_chunk, gexp_chunk);
ggml_tensor * attn_inter = ggml_mul_mat(ctx0, state_t, q_g_exp);
cb(attn_inter, "attn_inter_chunk", il);
// core_attn_out[:, :, i] = attn_inter + attn @ v_new
ggml_tensor * v_attn = ggml_mul_mat(ctx0, v_new_t, attn_chunk);
cb(v_attn, "v_attn_chunk", il);
ggml_tensor * core_attn_out_chunk = ggml_add(ctx0, attn_inter, v_attn);
cb(core_attn_out_chunk, "core_attn_out_chunk", il); // shape: (S_v, chunk_size, 1, H_v * n_seqs)
core_attn_out = core_attn_out == nullptr
? core_attn_out_chunk
: ggml_concat(ctx0, core_attn_out, core_attn_out_chunk, 2);
// kgdmulvnew = (key_gdiff).transpose(-1, -2) @ v_new
ggml_tensor * k_gdiff_t = get_slice_2d(ctx0, key_gdiff_t, chunk);
//ggml_tensor * kgdmulvnew = ggml_mul_mat(ctx0, k_gdiff, v_new); // this is slower on metal, why?
ggml_tensor * kgdmulvnew = ggml_mul_mat(ctx0, v_new_t, k_gdiff_t);
// last_recurrent_state = last_recurrent_state * g_last + kgdmulvnew
ggml_tensor * gexp_last_chunk = ggml_cont(ctx0, get_slice_2d(ctx0, g_last_exp, chunk));
new_state = ggml_add(ctx0,
ggml_mul(ctx0, new_state, ggml_reshape_4d(ctx0, gexp_last_chunk, gexp_last_chunk->ne[0], gexp_last_chunk->ne[1], H_v, n_seqs)),
ggml_reshape_4d(ctx0, kgdmulvnew, kgdmulvnew->ne[0], kgdmulvnew->ne[1], H_v, n_seqs));
}
// truncate padded tokens
ggml_tensor * output_tokens = ggml_view_4d(ctx0, core_attn_out,
S_v, n_tokens, H_v, n_seqs,
ggml_row_size(core_attn_out->type, S_v),
ggml_row_size(core_attn_out->type, S_v * chunk_size * n_chunks),
ggml_row_size(core_attn_out->type, S_v * chunk_size * n_chunks * H_v), 0);
output_tokens = ggml_cont(ctx0, output_tokens);
cb(output_tokens, "output_tokens", il);
// permute back to (S_v, H_v, n_tokens, n_seqs)
output_tokens = ggml_permute(ctx0, output_tokens, 0, 2, 1, 3);
output_tokens = ggml_cont(ctx0, output_tokens);
return {output_tokens, new_state};
}
std::pair<ggml_tensor *, ggml_tensor *> llm_build_qwen35::build_delta_net_autoregressive(
ggml_tensor * q,
ggml_tensor * k,
ggml_tensor * v,
ggml_tensor * g,
ggml_tensor * beta,
ggml_tensor * state,
int il) {
const int64_t S_k = q->ne[0];
const int64_t H_k = q->ne[1];
const int64_t n_tokens = q->ne[2];
const int64_t n_seqs = q->ne[3];
const int64_t S_v = v->ne[0];
const int64_t H_v = v->ne[1];
GGML_ASSERT(n_tokens == 1); // This function is optimized for single token processing
GGML_ASSERT(v->ne[2] == n_tokens);
GGML_ASSERT(k->ne[2] == n_tokens);
GGML_ASSERT(g->ne[0] == H_v && g->ne[1] == n_tokens && g->ne[2] == n_seqs);
GGML_ASSERT(beta->ne[0] == H_v && beta->ne[2] == n_tokens && beta->ne[3] == n_seqs);
GGML_ASSERT(state->ne[0] == S_v && state->ne[1] == S_v * H_v && state->ne[2] == 1 && state->ne[3] == n_seqs);
GGML_ASSERT(q->ne[0] == S_k && q->ne[1] == H_k && q->ne[2] == n_tokens && q->ne[3] == n_seqs);
GGML_ASSERT(k->ne[0] == S_k && k->ne[1] == H_k && k->ne[2] == n_tokens && k->ne[3] == n_seqs);
GGML_ASSERT(H_k == H_v); // we did a repeat to make sure this is the case
const float eps_norm = hparams.f_norm_rms_eps;
q = ggml_l2_norm(ctx0, q, eps_norm);
k = ggml_l2_norm(ctx0, k, eps_norm);
const float scale = 1.0f / sqrtf(S_v);
q = ggml_scale(ctx0, q, scale);
beta = ggml_sigmoid(ctx0, beta);
cb(q, "q_in", il);
cb(k, "k_in", il);
cb(v, "v_in", il);
cb(beta, "beta_in", il);
cb(g, "g_in", il);
state = ggml_reshape_4d(ctx0, state, S_v, S_v, H_v, n_seqs);
ggml_tensor * g_t = ggml_reshape_4d(ctx0, ggml_transpose(ctx0, g), 1, 1, H_k, n_seqs);
ggml_tensor * beta_t = ggml_reshape_4d(ctx0, ggml_transpose(ctx0, beta), 1, 1, H_k, n_seqs);
// Apply exponential to g_t
g_t = ggml_exp(ctx0, g_t);
// Apply the gated delta rule for the single timestep
// last_recurrent_state = last_recurrent_state * g_t
state = ggml_mul(ctx0, state, g_t);
// kv_mem = (last_recurrent_state * k_t.unsqueeze(-1)).sum(dim=-2)
ggml_tensor * k_t_unsqueezed = ggml_reshape_4d(ctx0, k, 1, S_v, H_v, n_seqs);
ggml_tensor * kv_mem = ggml_mul(ctx0, state, k_t_unsqueezed);
// we need to sum over dim=-2, so we transpose, sum, then transpose again
kv_mem = ggml_transpose(ctx0, ggml_sum_rows(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, kv_mem))));
// v_t = v.unsqueeze(2) (we insert the singleton dimension after n_seqs and H_v)
ggml_tensor * v_t = ggml_reshape_4d(ctx0, v, S_v, 1, H_v, n_seqs);
// delta = (v_t - kv_mem) * beta_t
ggml_tensor * v_diff = ggml_sub(ctx0, v_t, kv_mem); // both should be [S_v, 1, H_v, n_seqs]
ggml_tensor * delta = ggml_mul(ctx0, v_diff, beta_t);
// last_recurrent_state = last_recurrent_state + k_t.unsqueeze(-1) * delta
ggml_tensor * k_t_delta = ggml_mul(ctx0, ggml_repeat_4d(ctx0, k_t_unsqueezed, S_v, S_v, H_v, n_seqs), delta);
state = ggml_add(ctx0, state, k_t_delta);
// Compute the attention output
// core_attn_out = (last_recurrent_state * q_t.unsqueeze(-1)).sum(dim=-2)
ggml_tensor * q_t_unsqueezed = ggml_reshape_4d(ctx0, q, 1, S_v, H_v, n_seqs); // unsqueeze q_t
ggml_tensor * state_q = ggml_mul(ctx0, state, q_t_unsqueezed);
// again, since it's over dim = -2, transpose, sum, transpose back
ggml_tensor * core_attn_out =
ggml_transpose(ctx0, ggml_sum_rows(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, state_q))));
// core_attn_out should be [S_v, 1, H_v, n_seqs] after this
cb(core_attn_out, "output_tokens", il);
cb(state, "new_state", il);
return {core_attn_out, state};
}
std::pair<ggml_tensor *, ggml_tensor *> llm_build_qwen35::build_qkvz(
ggml_tensor * input,
int il) {
const int64_t n_seqs = ubatch.n_seqs;
const int64_t n_seq_tokens = ubatch.n_seq_tokens;
ggml_tensor * qkv_mixed = build_lora_mm(model.layers[il].wqkv, input);
qkv_mixed = ggml_reshape_3d(ctx0, qkv_mixed, qkv_mixed->ne[0], n_seq_tokens, n_seqs);
cb(qkv_mixed, "linear_attn_qkv_mixed", il);
ggml_tensor * z = build_lora_mm(model.layers[il].wqkv_gate, input);
cb(z, "z", il);
return { qkv_mixed, z };
}
ggml_tensor * llm_build_qwen35::build_norm_gated(
ggml_tensor * input,
ggml_tensor * weights,
ggml_tensor * gate,
int layer) {
ggml_tensor * normalized = build_norm(input, weights, nullptr, LLM_NORM_RMS, layer);
ggml_tensor * gated_silu = ggml_silu(ctx0, gate);
return ggml_mul(ctx0, normalized, gated_silu);
}
ggml_tensor * llm_build_qwen35::build_layer_attn(
llm_graph_input_attn_kv * inp,
ggml_tensor * cur,
ggml_tensor * inp_pos,
int * sections,
int il) {
const int64_t n_embd_head = hparams.n_embd_head_v;
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
// Order: joint QG projection, QG split, Q norm, KV projection, K norm, RoPE, attention
// Qwen3Next uses a single Q projection that outputs query + gate
ggml_tensor * Qcur_full = build_lora_mm(model.layers[il].wq, cur); // [ (n_embd_head * 2) * n_head, n_tokens ]
cb(Qcur_full, "Qcur_full", il);
ggml_tensor * Qcur = ggml_view_3d(ctx0, Qcur_full, n_embd_head, n_head, n_tokens,
ggml_element_size(Qcur_full) * n_embd_head * 2,
ggml_element_size(Qcur_full) * n_embd_head * 2 * n_head, 0);
cb(Qcur, "Qcur_reshaped", il);
// Apply Q normalization
Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, nullptr, LLM_NORM_RMS, il);
cb(Qcur, "Qcur_normed", il);
ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
cb(Kcur, "Kcur", il);
ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
cb(Vcur, "Vcur", il);
// Apply K normalization
Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, nullptr, LLM_NORM_RMS, il);
cb(Kcur, "Kcur_normed", il);
ggml_tensor * gate = ggml_view_3d(ctx0, Qcur_full, n_embd_head, n_head, n_tokens,
ggml_element_size(Qcur_full) * n_embd_head * 2,
ggml_element_size(Qcur_full) * n_embd_head * 2 * n_head,
ggml_element_size(Qcur_full) * n_embd_head);
gate = ggml_cont_2d(ctx0, gate, n_embd_head * n_head, n_tokens);
cb(gate, "gate_reshaped", il);
Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
// Apply MRoPE
Qcur = ggml_rope_multi(
ctx0, Qcur, inp_pos, nullptr,
n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
Kcur = ggml_rope_multi(
ctx0, Kcur, inp_pos, nullptr,
n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
cb(Qcur, "Qcur", il);
cb(Kcur, "Kcur", il);
cb(Vcur, "Vcur", il);
// Attention computation
const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
cur = build_attn(inp,
nullptr, nullptr,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
cb(cur, "attn_pregate", il);
ggml_tensor * gate_sigmoid = ggml_sigmoid(ctx0, gate);
cb(gate_sigmoid, "gate_sigmoid", il);
cur = ggml_mul(ctx0, cur, gate_sigmoid);
cb(cur, "attn_gated", il);
cur = build_lora_mm(model.layers[il].wo, cur);
cb(cur, "attn_output", il);
return cur;
}
ggml_tensor * llm_build_qwen35::build_layer_attn_linear(
llm_graph_input_rs * inp,
ggml_tensor * cur,
ggml_tensor * causal_mask,
ggml_tensor * identity,
ggml_tensor * diag_mask,
int il) {
const auto * mctx_cur = inp->mctx;
const int64_t d_inner = hparams.ssm_d_inner;
const int64_t n_seqs = ubatch.n_seqs;
const int64_t head_k_dim = hparams.ssm_d_state;
const int64_t num_k_heads = hparams.ssm_n_group;
const int64_t num_v_heads = hparams.ssm_dt_rank;
const int64_t head_v_dim = d_inner / num_v_heads;
const int64_t n_seq_tokens = ubatch.n_seq_tokens;
const auto kv_head = mctx_cur->get_head();
GGML_ASSERT(n_seqs != 0);
GGML_ASSERT(ubatch.equal_seqs());
GGML_ASSERT(ubatch.n_tokens == n_seq_tokens * n_seqs);
// Input projections
auto qkvz = build_qkvz(cur, il);
ggml_tensor * qkv_mixed = qkvz.first;
ggml_tensor * z = qkvz.second;
ggml_tensor * beta = build_lora_mm(model.layers[il].ssm_beta, cur);
beta = ggml_reshape_4d(ctx0, beta, num_v_heads, 1, n_seq_tokens, n_seqs);
cb(beta, "beta", il);
ggml_tensor * alpha = build_lora_mm(model.layers[il].ssm_alpha, cur);
alpha = ggml_cont_3d(ctx0, alpha, num_v_heads, n_seq_tokens, n_seqs);
cb(alpha, "alpha", il);
ggml_tensor * alpha_biased = ggml_add(ctx0, alpha, model.layers[il].ssm_dt);
ggml_tensor * alpha_softplus = ggml_softplus(ctx0, alpha_biased);
cb(alpha_softplus, "a_softplus", il);
ggml_tensor * gate = ggml_mul(ctx0, alpha_softplus, model.layers[il].ssm_a); // -A_log.exp() * softplus
cb(gate, "gate", il);
// Get convolution states from cache
ggml_tensor * conv_states_all = mctx_cur->get_r_l(il);
ggml_tensor * ssm_states_all = mctx_cur->get_s_l(il);
// bool use_precomputed_states = n_seq_tokens == 1 && mctx_cur->has_previous_state();
// Build the convolution states tensor
ggml_tensor * conv_states = build_rs(inp, conv_states_all, hparams.n_embd_r(), n_seqs);
cb(conv_states, "conv_states", il);
// Calculate convolution kernel size
ggml_tensor * conv_kernel = model.layers[il].ssm_conv1d;
const int64_t conv_kernel_size = conv_kernel->ne[0];
const int64_t conv_channels = d_inner + 2 * hparams.ssm_n_group * hparams.ssm_d_state;
conv_states = ggml_reshape_3d(ctx0, conv_states, conv_kernel_size - 1, conv_channels, n_seqs);
cb(conv_states, "conv_states_reshaped", il);
qkv_mixed = ggml_permute(ctx0, qkv_mixed, 1, 0, 2, 3);
cb(qkv_mixed, "qkv_mixed_permuted", il);
ggml_tensor * conv_input = ggml_concat(ctx0, conv_states, qkv_mixed, 0);
cb(conv_input, "conv_input", il);
// Update convolution state cache
// Extract the last (conv_kernel_size - 1) states from conv_input
ggml_tensor * last_conv_states =
ggml_view_3d(ctx0, conv_input, conv_kernel_size - 1, conv_channels, n_seqs, conv_input->nb[1],
conv_input->nb[2], (conv_input->ne[0] - conv_states->ne[0]) * ggml_element_size(conv_input));
cb(last_conv_states, "last_conv_states", il);
ggml_tensor * state_update_target =
ggml_view_1d(ctx0, conv_states_all, (conv_kernel_size - 1) * conv_channels * n_seqs,
kv_head * (conv_kernel_size - 1) * conv_channels * ggml_element_size(conv_states_all));
cb(state_update_target, "state_update_target", il);
ggml_build_forward_expand(gf, ggml_cpy(ctx0, last_conv_states, state_update_target));
cb(conv_states_all, "conv_states_updated", il);
// Apply SSM convolution
ggml_tensor * conv_output_proper = ggml_ssm_conv(ctx0, conv_input, conv_kernel);
cb(conv_output_proper, "conv_output_raw", il);
ggml_tensor * conv_output_silu = ggml_silu(ctx0, conv_output_proper);
cb(conv_output_silu, "conv_output_silu", il);
ggml_tensor * conv_qkv_mix = conv_output_silu;
// Calculate the total conv dimension
int64_t qkv_dim = head_k_dim * num_k_heads * 2 + head_v_dim * num_v_heads;
int64_t nb1_qkv = ggml_row_size(conv_qkv_mix->type, qkv_dim);
// Extract the convolved Q, K, V from conv_output
ggml_tensor * q_conv =
ggml_view_2d(ctx0, conv_qkv_mix, head_k_dim * num_k_heads, n_seq_tokens * n_seqs, nb1_qkv, 0);
cb(q_conv, "q_conv", il);
ggml_tensor * k_conv =
ggml_view_2d(ctx0, conv_qkv_mix, head_k_dim * num_k_heads, n_seq_tokens * n_seqs, nb1_qkv,
head_k_dim * num_k_heads * ggml_element_size(conv_qkv_mix));
cb(k_conv, "k_conv", il);
ggml_tensor * v_conv =
ggml_view_2d(ctx0, conv_qkv_mix, head_v_dim * num_v_heads, n_seq_tokens * n_seqs, nb1_qkv,
2 * head_k_dim * num_k_heads * ggml_element_size(conv_qkv_mix));
cb(v_conv, "v_conv", il);
// Unsqueeze them
q_conv = ggml_cont_4d(ctx0, q_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs);
k_conv = ggml_cont_4d(ctx0, k_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs);
v_conv = ggml_cont_4d(ctx0, v_conv, head_v_dim, num_v_heads, n_seq_tokens, n_seqs);
ggml_tensor * state = build_rs(inp, ssm_states_all, hparams.n_embd_s(), n_seqs);
state = ggml_reshape_4d(ctx0, state, head_v_dim, head_v_dim * num_v_heads, 1, n_seqs);
cb(state, "state_predelta", il);
// if head keys and value keys are different, repeat Q/K to match V's head count
// V heads are in tiled order (from conversion), so simple tiled repeat works
if (num_k_heads != num_v_heads) {
GGML_ASSERT(num_v_heads % num_k_heads == 0);
q_conv = ggml_repeat_4d(ctx0, q_conv, head_k_dim, num_v_heads, n_seq_tokens, n_seqs);
k_conv = ggml_repeat_4d(ctx0, k_conv, head_k_dim, num_v_heads, n_seq_tokens, n_seqs);
}
cb(q_conv, "q_conv_predelta", il);
cb(k_conv, "k_conv_predelta", il);
cb(v_conv, "v_conv_predelta", il);
// Choose between build_delta_net_chunking, build_delta_net_recurrent, and build_delta_net_autoregressive based on n_tokens
std::pair<ggml_tensor *, ggml_tensor *> attn_out; // pair of (output, new_state)
if (n_seq_tokens == 1) {
attn_out = build_delta_net_autoregressive(q_conv, k_conv, v_conv, gate, beta, state, il);
} else {
attn_out = build_delta_net_chunking(q_conv, k_conv, v_conv, gate, beta, state, causal_mask, identity, diag_mask, il);
}
ggml_tensor * output = attn_out.first;
ggml_tensor * new_state = attn_out.second;
cb(output, "attn_output", il);
cb(new_state, "new_state", il);
// Update the recurrent states
ggml_build_forward_expand(gf,
ggml_cpy(ctx0, new_state,
ggml_view_1d(ctx0, ssm_states_all, hparams.n_embd_s() * n_seqs,
kv_head * hparams.n_embd_s() * ggml_element_size(ssm_states_all))));
// Reshape both attn_out_final and z to 2D tensors for normalization
// attn_out_final: [head_dim, n_heads, n_tokens, n_seqs] -> [n_heads * n_tokens * n_seqs, head_dim]
ggml_tensor * attn_out_2d_final = ggml_reshape_2d(ctx0, output, head_v_dim, num_v_heads * n_seq_tokens * n_seqs);
// z: [head_dim, n_heads, n_tokens, n_seqs] -> [n_heads * n_tokens * n_seqs, head_dim]
ggml_tensor * z_2d = ggml_reshape_2d(ctx0, z, head_v_dim, num_v_heads * n_seq_tokens * n_seqs);
// Apply gated normalization: self.norm(core_attn_out, z)
ggml_tensor * attn_out_norm = build_norm_gated(attn_out_2d_final, model.layers[il].ssm_norm, z_2d, il);
// Final reshape: [head_dim, n_heads, n_tokens, n_seqs] -> [n_tokens, n_seqs, n_heads * head_dim]
ggml_tensor * final_output = ggml_reshape_3d(ctx0, attn_out_norm, head_v_dim * num_v_heads, n_seq_tokens, n_seqs);
cb(final_output, "final_output", il);
// Output projection
cur = build_lora_mm(model.layers[il].ssm_out, final_output);
cb(cur, "linear_attn_out", il);
// Reshape back to original dimensions
cur = ggml_cont_2d(ctx0, cur, n_embd, n_seq_tokens * n_seqs);
return cur;
}
ggml_tensor * llm_build_qwen35::build_layer_ffn(ggml_tensor * cur, const int il) {
// Qwen3.5 does not use MoE FFN
GGML_ASSERT(model.layers[il].ffn_gate_inp == nullptr);
cur = build_ffn(cur,
model.layers[il].ffn_up, NULL, NULL,
model.layers[il].ffn_gate, NULL, NULL,
model.layers[il].ffn_down, NULL, NULL,
NULL,
LLM_FFN_SILU, LLM_FFN_PAR, il);
cb(cur, "ffn_out", il);
return cur;
}

774
src/models/qwen35moe.cpp Normal file
View File

@ -0,0 +1,774 @@
#include "ggml.h"
#include "models.h"
#define CHUNK_SIZE 64
llm_build_qwen35moe::llm_build_qwen35moe(const llama_model & model, const llm_graph_params & params) :
llm_graph_context_mamba(params), model(model) {
const int64_t n_embd_head = hparams.n_embd_head_v;
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
int sections[4];
std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections);
ggml_tensor * cur;
ggml_tensor * inpL;
inpL = build_inp_embd(model.tok_embd);
cb(inpL, "model.input_embed", -1);
auto * inp = build_inp_mem_hybrid();
ggml_tensor * inp_pos = build_inp_pos();
ggml_tensor * inp_out_ids = build_inp_out_ids();
ggml_tensor * causal_mask =
ggml_tri(ctx0, ggml_fill(ctx0, ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, CHUNK_SIZE, CHUNK_SIZE), 1.0f),
GGML_TRI_TYPE_LOWER);
ggml_tensor * identity = ggml_diag(ctx0, ggml_fill(ctx0, ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, CHUNK_SIZE), 1.0f));
ggml_tensor * diag_mask = ggml_add(ctx0, causal_mask, identity);
ggml_build_forward_expand(gf, causal_mask);
ggml_build_forward_expand(gf, identity);
ggml_build_forward_expand(gf, diag_mask);
for (int il = 0; il < n_layer; ++il) {
ggml_tensor * inpSA = inpL;
cur = build_norm(inpL, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il);
cb(cur, "attn_norm", il);
// Determine layer type and build appropriate attention mechanism
if (hparams.is_recurrent(il)) {
// Linear attention layer (gated delta net)
cur = build_layer_attn_linear(inp->get_recr(), cur, causal_mask, identity, diag_mask, il);
} else {
// Full attention layer
cur = build_layer_attn(inp->get_attn(), cur, inp_pos, sections, il);
}
if (il == n_layer - 1 && inp_out_ids) {
cur = ggml_get_rows(ctx0, cur, inp_out_ids);
inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
}
// Residual connection
cur = ggml_add(ctx0, cur, inpSA);
cb(cur, "attn_residual", il);
// Save the tensor before post-attention norm for residual connection
ggml_tensor * ffn_residual = cur;
// Post-attention norm
ggml_tensor * attn_post_norm = build_norm(cur, model.layers[il].attn_post_norm, nullptr, LLM_NORM_RMS, il);
cb(attn_post_norm, "attn_post_norm", il);
// MOE FFN layer
cur = build_layer_ffn(attn_post_norm, il);
cb(cur, "ffn_out", il);
// Residual connection for FFN - add to the tensor from before post_attention_layernorm
cur = ggml_add(ctx0, cur, ffn_residual);
cb(cur, "post_moe", il);
// Input for next layer
inpL = cur;
}
cur = inpL;
// Final norm
cur = build_norm(cur, model.output_norm, nullptr, LLM_NORM_RMS, -1);
cb(cur, "result_norm", -1);
res->t_embd = cur;
// LM head
cur = build_lora_mm(model.output, cur);
cb(cur, "result_output", -1);
res->t_logits = cur;
ggml_build_forward_expand(gf, cur);
}
// utility to get one slice from the third dimension
// input dim: [x, y, c, b]
// output dim: [x, y, 1, b]
static ggml_tensor * get_slice_2d(ggml_context * ctx0, ggml_tensor * t, int64_t c) {
return ggml_view_4d(ctx0, t, t->ne[0], t->ne[1], 1, t->ne[3],
t->nb[1], t->nb[2], t->nb[3], t->nb[2] * c);
}
std::pair<ggml_tensor *, ggml_tensor *> llm_build_qwen35moe::build_delta_net_chunking(
ggml_tensor * q,
ggml_tensor * k,
ggml_tensor * v,
ggml_tensor * g,
ggml_tensor * beta,
ggml_tensor * state,
ggml_tensor * causal_mask,
ggml_tensor * identity,
ggml_tensor * diag_mask,
int il) {
const int64_t S_k = q->ne[0];
const int64_t H_k = q->ne[1];
const int64_t n_tokens = q->ne[2];
const int64_t n_seqs = q->ne[3];
const int64_t S_v = v->ne[0];
const int64_t H_v = v->ne[1];
GGML_ASSERT(v->ne[2] == n_tokens);
GGML_ASSERT(k->ne[2] == n_tokens);
GGML_ASSERT(g->ne[0] == H_v && g->ne[1] == n_tokens && g->ne[2] == n_seqs);
GGML_ASSERT(beta->ne[0] == H_v && beta->ne[2] == n_tokens && beta->ne[3] == n_seqs);
GGML_ASSERT(state->ne[0] == S_v && state->ne[1] == S_v * H_v && state->ne[2] == 1 && state->ne[3] == n_seqs);
GGML_ASSERT(q->ne[0] == S_k && q->ne[1] == H_k && q->ne[2] == n_tokens && q->ne[3] == n_seqs);
GGML_ASSERT(k->ne[0] == S_k && k->ne[1] == H_k && k->ne[2] == n_tokens && k->ne[3] == n_seqs);
GGML_ASSERT(H_k == H_v); // we did a repeat to make sure this is the case
const float eps_norm = hparams.f_norm_rms_eps;
q = ggml_l2_norm(ctx0, q, eps_norm);
k = ggml_l2_norm(ctx0, k, eps_norm);
const float scale = 1.0f / sqrtf(S_v);
q = ggml_scale(ctx0, q, scale);
beta = ggml_sigmoid(ctx0, beta);
cb(q, "q_in", il);
cb(k, "k_in", il);
cb(v, "v_in", il);
cb(beta, "beta_in", il);
cb(g, "g_in", il);
q = ggml_cont_4d(ctx0, ggml_permute(ctx0, q, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs);
k = ggml_cont_4d(ctx0, ggml_permute(ctx0, k, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs);
v = ggml_cont_4d(ctx0, ggml_permute(ctx0, v, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs);
g = ggml_cont_4d(ctx0, ggml_permute(ctx0, g, 2, 0, 3, 1), n_tokens, 1, H_k, n_seqs);
beta = ggml_cont(ctx0, ggml_permute(ctx0, beta, 2, 0, 1, 3));
state = ggml_reshape_4d(ctx0, state, S_v, S_v, H_v, n_seqs);
cb(q, "q_perm", il);
cb(k, "k_perm", il);
cb(v, "v_perm", il);
cb(beta, "beta_perm", il);
cb(g, "g_perm", il);
cb(state, "state_in", il);
GGML_ASSERT(q->ne[1] == n_tokens && q->ne[0] == S_k && q->ne[2] == H_k && q->ne[3] == n_seqs);
GGML_ASSERT(k->ne[1] == n_tokens && k->ne[0] == S_k && k->ne[2] == H_k && k->ne[3] == n_seqs);
GGML_ASSERT(v->ne[1] == n_tokens && v->ne[0] == S_v && v->ne[2] == H_k && v->ne[3] == n_seqs);
GGML_ASSERT(beta->ne[1] == n_tokens && beta->ne[2] == H_k && beta->ne[0] == 1 && beta->ne[3] == n_seqs);
// Do padding
const int64_t chunk_size = CHUNK_SIZE;
const int64_t pad = (chunk_size - n_tokens % chunk_size) % chunk_size;
const int64_t n_chunks = (n_tokens + pad) / chunk_size;
q = ggml_pad(ctx0, q, 0, pad, 0, 0);
k = ggml_pad(ctx0, k, 0, pad, 0, 0);
v = ggml_pad(ctx0, v, 0, pad, 0, 0);
g = ggml_pad(ctx0, g, pad, 0, 0, 0);
beta = ggml_pad(ctx0, beta, 0, pad, 0, 0);
cb(q, "q_pad", il);
cb(k, "k_pad", il);
cb(v, "v_pad", il);
cb(beta, "beta_pad", il);
cb(g, "g_pad", il);
ggml_tensor * v_beta = ggml_mul(ctx0, v, beta);
ggml_tensor * k_beta = ggml_mul(ctx0, k, beta);
cb(v_beta, "v_beta", il);
cb(k_beta, "k_beta", il);
q = ggml_reshape_4d(ctx0, q, S_k, chunk_size, n_chunks, H_k * n_seqs);
k = ggml_reshape_4d(ctx0, k, S_k, chunk_size, n_chunks, H_k * n_seqs);
k_beta = ggml_reshape_4d(ctx0, k_beta, S_k, chunk_size, n_chunks, H_k * n_seqs);
v = ggml_reshape_4d(ctx0, v, S_v, chunk_size, n_chunks, H_v * n_seqs);
v_beta = ggml_reshape_4d(ctx0, v_beta, S_v, chunk_size, n_chunks, H_v * n_seqs);
g = ggml_reshape_4d(ctx0, g, chunk_size, 1, n_chunks, H_k * n_seqs);
beta = ggml_reshape_4d(ctx0, beta, 1, chunk_size, n_chunks, H_k * n_seqs);
ggml_tensor * g_cumsum = ggml_cumsum(ctx0, g);
cb(g_cumsum, "g_cumsum", il); // shape: (chunk_size, 1, n_chunks, H_v * n_seqs)
ggml_tensor * gcs_i = g_cumsum; // ggml_reshape_4d(ctx0, g_cumsum, chunk_size, 1, n_chunks, H_v * n_seqs);
ggml_tensor * gcs_j = ggml_reshape_4d(ctx0, g_cumsum, 1, chunk_size, n_chunks, H_v * n_seqs);
ggml_tensor * gcs_j_broadcast =
ggml_repeat_4d(ctx0, gcs_j, chunk_size, chunk_size, n_chunks, H_v * n_seqs);
ggml_tensor * decay_mask = ggml_sub(ctx0, gcs_j_broadcast, gcs_i);
cb(decay_mask, "decay_mask", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs)
decay_mask = ggml_mul(ctx0, decay_mask, diag_mask);
decay_mask = ggml_exp(ctx0, decay_mask);
decay_mask = ggml_mul(ctx0, decay_mask, diag_mask);
ggml_tensor * kmulkbeta = ggml_mul_mat(ctx0, k, k_beta);
ggml_tensor * k_decay = ggml_mul(ctx0, kmulkbeta, decay_mask);
ggml_tensor * attn = ggml_neg(ctx0, ggml_mul(ctx0, k_decay, causal_mask));
cb(attn, "attn_pre_solve", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs)
ggml_tensor * attn_lower = ggml_mul(ctx0, attn, causal_mask);
ggml_tensor * lhs = ggml_sub(ctx0, ggml_repeat(ctx0, identity, attn_lower), attn_lower);
ggml_tensor * lin_solve = ggml_solve_tri(ctx0, lhs, attn, true, true, false);
attn = ggml_mul(ctx0, lin_solve, causal_mask);
attn = ggml_add(ctx0, attn, identity);
cb(attn, "attn_solved", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs)
v = ggml_mul_mat(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, v_beta)), attn);
ggml_tensor * g_cumsum_t = ggml_cont(ctx0, ggml_transpose(ctx0, g_cumsum));
ggml_tensor * gexp = ggml_exp(ctx0, g_cumsum_t);
ggml_tensor * kbeta_gexp = ggml_mul(ctx0, k_beta, gexp);
cb(kbeta_gexp, "kbeta_gexp", il); // shape: (S_k, chunk_size, n_chunks, H_v * n_seqs)
ggml_tensor * k_cumdecay =
ggml_cont(ctx0, ggml_transpose(ctx0, ggml_mul_mat(ctx0, attn, ggml_cont(ctx0, ggml_transpose(ctx0, kbeta_gexp)))));
cb(k_cumdecay, "k_cumdecay", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs)
ggml_tensor * attn_kq = ggml_mul_mat(ctx0, k, q);
attn_kq = ggml_mul(ctx0, attn_kq, decay_mask);
attn_kq = ggml_mul(ctx0, attn_kq, diag_mask);
cb(attn_kq, "attn_kq", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs)
// vectorized calculation of key_gdiff
// improved from the chunked version:
// g_last = torch.clamp(g_cum[:, :, -1], max=50.0).exp().unsqueeze(-1).unsqueeze(-1)
// g_diff = torch.clamp(g_cum[:, :, -1:] - g_cum, max=50.0).exp()
// key_gdiff = key * g_diff.unsqueeze(-1)
// kgdmulvnew = (key_gdiff).transpose(-1, -2) @ v_new
// last_recurrent_state = last_recurrent_state * g_last + kgdmulvnew
// get last element in g_cumsum along chunk_size dimension (ne0)
// example: [[x, y, z, ..., last], ...] -> [[last], ...]
ggml_tensor * g_last = ggml_view_4d(ctx0, g_cumsum, 1, 1, g_cumsum->ne[2], g_cumsum->ne[3],
g_cumsum->nb[1], g_cumsum->nb[2], g_cumsum->nb[3],
(g_cumsum->ne[0] - 1) * ggml_element_size(g_cumsum));
g_last = ggml_cont(ctx0, g_last);
cb(g_last, "g_last", il); // shape: (1, 1, n_chunks, H_v * n_seqs)
ggml_tensor * g_last_exp = ggml_exp(ctx0, g_last);
cb(g_last_exp, "g_last_exp", il); // shape: (1, 1, n_chunks, H_v * n_seqs)
ggml_tensor * g_diff = ggml_neg(ctx0, ggml_sub(ctx0, g_cumsum, g_last));
cb(g_diff, "g_diff", il); // shape: (chunk_size, 1, n_chunks, H_v * n_seqs)
ggml_tensor * g_diff_exp = ggml_exp(ctx0, g_diff);
ggml_tensor * g_diff_exp_t = ggml_reshape_4d(ctx0, g_diff_exp,
1, chunk_size, n_chunks, g_diff_exp->ne[3]);
ggml_tensor * key_gdiff = ggml_mul(ctx0, k, g_diff_exp_t);
cb(key_gdiff, "key_gdiff", il); // shape: (S_k, chunk_size, n_chunks, H_v * n_seqs)
ggml_tensor * key_gdiff_t = ggml_cont(ctx0, ggml_transpose(ctx0, key_gdiff));
cb(key_gdiff_t, "key_gdiff_t", il); // shape: (chunk_size, S_k, n_chunks, H_v * n_seqs)
// state to be updated per chunk
ggml_tensor * new_state = state; // ggml_dup(ctx0, state);
cb(new_state, "new_state", il); // shape: (S_v, S_v, H_v, n_seqs)
// shape after loop of chunks: (S_v, chunk_size, n_chunks, H_v * n_seqs)
ggml_tensor * core_attn_out = nullptr;
for (int64_t chunk = 0; chunk < n_chunks; chunk++) {
// shape: (S_k, chunk_size, 1, H_k * n_seqs)
ggml_tensor * q_chunk = get_slice_2d(ctx0, q, chunk); // (no cont), next op: ggml_mul
// shape: (S_v, chunk_size, 1, H_v * n_seqs)
ggml_tensor * v_chunk = get_slice_2d(ctx0, v, chunk); // (no cont), next op: ggml_repeat
// shape: (chunk_size, 1, n_chunks, H_v * n_seqs)
ggml_tensor * gexp_chunk = get_slice_2d(ctx0, gexp, chunk); // (no cont), next op: ggml_mul
// shape: (chunk_size, 1, H_v * n_seqs)
ggml_tensor * k_cumdecay_chunk = get_slice_2d(ctx0, k_cumdecay, chunk); // (no cont), next op: ggml_mul_mat
// attn = (q_i @ k_i.transpose(-1, -2) * decay_mask[:, :, i]).masked_fill_(mask, 0)
// replaced by precomputed attn_kq
ggml_tensor * attn_chunk = get_slice_2d(ctx0, attn_kq, chunk);
cb(attn_chunk, "attn_chunk", il);
ggml_tensor * state_t = ggml_cont_4d(ctx0, ggml_permute(ctx0, new_state, 1, 0, 2, 3), S_v, S_v, 1, H_v * n_seqs);
// v_prime = (k_cumdecay[:, :, i]) @ last_recurrent_state
ggml_tensor * v_prime = ggml_mul_mat(ctx0, state_t, k_cumdecay_chunk);
cb(v_prime, "v_prime_chunk", il); // shape: (S_v, 1, H_v * n_seqs)
// v_new = v_i - v_prime
ggml_tensor * v_new = ggml_sub(ctx0, ggml_repeat(ctx0, v_chunk, v_prime), v_prime);
ggml_tensor * v_new_t = ggml_cont(ctx0, ggml_transpose(ctx0, v_new));
cb(v_new, "v_new_chunk", il);
// attn_inter = (q_i * g[:, :, i, :, None].exp()) @ last_recurrent_state
ggml_tensor * q_g_exp = ggml_mul(ctx0, q_chunk, gexp_chunk);
ggml_tensor * attn_inter = ggml_mul_mat(ctx0, state_t, q_g_exp);
cb(attn_inter, "attn_inter_chunk", il);
// core_attn_out[:, :, i] = attn_inter + attn @ v_new
ggml_tensor * v_attn = ggml_mul_mat(ctx0, v_new_t, attn_chunk);
cb(v_attn, "v_attn_chunk", il);
ggml_tensor * core_attn_out_chunk = ggml_add(ctx0, attn_inter, v_attn);
cb(core_attn_out_chunk, "core_attn_out_chunk", il); // shape: (S_v, chunk_size, 1, H_v * n_seqs)
core_attn_out = core_attn_out == nullptr
? core_attn_out_chunk
: ggml_concat(ctx0, core_attn_out, core_attn_out_chunk, 2);
// kgdmulvnew = (key_gdiff).transpose(-1, -2) @ v_new
ggml_tensor * k_gdiff_t = get_slice_2d(ctx0, key_gdiff_t, chunk);
//ggml_tensor * kgdmulvnew = ggml_mul_mat(ctx0, k_gdiff, v_new); // this is slower on metal, why?
ggml_tensor * kgdmulvnew = ggml_mul_mat(ctx0, v_new_t, k_gdiff_t);
// last_recurrent_state = last_recurrent_state * g_last + kgdmulvnew
ggml_tensor * gexp_last_chunk = ggml_cont(ctx0, get_slice_2d(ctx0, g_last_exp, chunk));
new_state = ggml_add(ctx0,
ggml_mul(ctx0, new_state, ggml_reshape_4d(ctx0, gexp_last_chunk, gexp_last_chunk->ne[0], gexp_last_chunk->ne[1], H_v, n_seqs)),
ggml_reshape_4d(ctx0, kgdmulvnew, kgdmulvnew->ne[0], kgdmulvnew->ne[1], H_v, n_seqs));
}
// truncate padded tokens
ggml_tensor * output_tokens = ggml_view_4d(ctx0, core_attn_out,
S_v, n_tokens, H_v, n_seqs,
ggml_row_size(core_attn_out->type, S_v),
ggml_row_size(core_attn_out->type, S_v * chunk_size * n_chunks),
ggml_row_size(core_attn_out->type, S_v * chunk_size * n_chunks * H_v), 0);
output_tokens = ggml_cont(ctx0, output_tokens);
cb(output_tokens, "output_tokens", il);
// permute back to (S_v, H_v, n_tokens, n_seqs)
output_tokens = ggml_permute(ctx0, output_tokens, 0, 2, 1, 3);
output_tokens = ggml_cont(ctx0, output_tokens);
return {output_tokens, new_state};
}
std::pair<ggml_tensor *, ggml_tensor *> llm_build_qwen35moe::build_delta_net_autoregressive(
ggml_tensor * q,
ggml_tensor * k,
ggml_tensor * v,
ggml_tensor * g,
ggml_tensor * beta,
ggml_tensor * state,
int il) {
const int64_t S_k = q->ne[0];
const int64_t H_k = q->ne[1];
const int64_t n_tokens = q->ne[2];
const int64_t n_seqs = q->ne[3];
const int64_t S_v = v->ne[0];
const int64_t H_v = v->ne[1];
GGML_ASSERT(n_tokens == 1); // This function is optimized for single token processing
GGML_ASSERT(v->ne[2] == n_tokens);
GGML_ASSERT(k->ne[2] == n_tokens);
GGML_ASSERT(g->ne[0] == H_v && g->ne[1] == n_tokens && g->ne[2] == n_seqs);
GGML_ASSERT(beta->ne[0] == H_v && beta->ne[2] == n_tokens && beta->ne[3] == n_seqs);
GGML_ASSERT(state->ne[0] == S_v && state->ne[1] == S_v * H_v && state->ne[2] == 1 && state->ne[3] == n_seqs);
GGML_ASSERT(q->ne[0] == S_k && q->ne[1] == H_k && q->ne[2] == n_tokens && q->ne[3] == n_seqs);
GGML_ASSERT(k->ne[0] == S_k && k->ne[1] == H_k && k->ne[2] == n_tokens && k->ne[3] == n_seqs);
GGML_ASSERT(H_k == H_v); // we did a repeat to make sure this is the case
const float eps_norm = hparams.f_norm_rms_eps;
q = ggml_l2_norm(ctx0, q, eps_norm);
k = ggml_l2_norm(ctx0, k, eps_norm);
const float scale = 1.0f / sqrtf(S_v);
q = ggml_scale(ctx0, q, scale);
beta = ggml_sigmoid(ctx0, beta);
cb(q, "q_in", il);
cb(k, "k_in", il);
cb(v, "v_in", il);
cb(beta, "beta_in", il);
cb(g, "g_in", il);
state = ggml_reshape_4d(ctx0, state, S_v, S_v, H_v, n_seqs);
ggml_tensor * g_t = ggml_reshape_4d(ctx0, ggml_transpose(ctx0, g), 1, 1, H_k, n_seqs);
ggml_tensor * beta_t = ggml_reshape_4d(ctx0, ggml_transpose(ctx0, beta), 1, 1, H_k, n_seqs);
// Apply exponential to g_t
g_t = ggml_exp(ctx0, g_t);
// Apply the gated delta rule for the single timestep
// last_recurrent_state = last_recurrent_state * g_t
state = ggml_mul(ctx0, state, g_t);
// kv_mem = (last_recurrent_state * k_t.unsqueeze(-1)).sum(dim=-2)
ggml_tensor * k_t_unsqueezed = ggml_reshape_4d(ctx0, k, 1, S_v, H_v, n_seqs);
ggml_tensor * kv_mem = ggml_mul(ctx0, state, k_t_unsqueezed);
// we need to sum over dim=-2, so we transpose, sum, then transpose again
kv_mem = ggml_transpose(ctx0, ggml_sum_rows(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, kv_mem))));
// v_t = v.unsqueeze(2) (we insert the singleton dimension after n_seqs and H_v)
ggml_tensor * v_t = ggml_reshape_4d(ctx0, v, S_v, 1, H_v, n_seqs);
// delta = (v_t - kv_mem) * beta_t
ggml_tensor * v_diff = ggml_sub(ctx0, v_t, kv_mem); // both should be [S_v, 1, H_v, n_seqs]
ggml_tensor * delta = ggml_mul(ctx0, v_diff, beta_t);
// last_recurrent_state = last_recurrent_state + k_t.unsqueeze(-1) * delta
ggml_tensor * k_t_delta = ggml_mul(ctx0, ggml_repeat_4d(ctx0, k_t_unsqueezed, S_v, S_v, H_v, n_seqs), delta);
state = ggml_add(ctx0, state, k_t_delta);
// Compute the attention output
// core_attn_out = (last_recurrent_state * q_t.unsqueeze(-1)).sum(dim=-2)
ggml_tensor * q_t_unsqueezed = ggml_reshape_4d(ctx0, q, 1, S_v, H_v, n_seqs); // unsqueeze q_t
ggml_tensor * state_q = ggml_mul(ctx0, state, q_t_unsqueezed);
// again, since it's over dim = -2, transpose, sum, transpose back
ggml_tensor * core_attn_out =
ggml_transpose(ctx0, ggml_sum_rows(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, state_q))));
// core_attn_out should be [S_v, 1, H_v, n_seqs] after this
cb(core_attn_out, "output_tokens", il);
cb(state, "new_state", il);
return {core_attn_out, state};
}
std::pair<ggml_tensor *, ggml_tensor *> llm_build_qwen35moe::build_qkvz(
ggml_tensor * input,
int il) {
const int64_t n_seqs = ubatch.n_seqs;
const int64_t n_seq_tokens = ubatch.n_seq_tokens;
ggml_tensor * qkv_mixed = build_lora_mm(model.layers[il].wqkv, input);
qkv_mixed = ggml_reshape_3d(ctx0, qkv_mixed, qkv_mixed->ne[0], n_seq_tokens, n_seqs);
cb(qkv_mixed, "linear_attn_qkv_mixed", il);
ggml_tensor * z = build_lora_mm(model.layers[il].wqkv_gate, input);
cb(z, "z", il);
return { qkv_mixed, z };
}
ggml_tensor * llm_build_qwen35moe::build_norm_gated(
ggml_tensor * input,
ggml_tensor * weights,
ggml_tensor * gate,
int layer) {
ggml_tensor * normalized = build_norm(input, weights, nullptr, LLM_NORM_RMS, layer);
ggml_tensor * gated_silu = ggml_silu(ctx0, gate);
return ggml_mul(ctx0, normalized, gated_silu);
}
ggml_tensor * llm_build_qwen35moe ::build_layer_attn(
llm_graph_input_attn_kv * inp,
ggml_tensor * cur,
ggml_tensor * inp_pos,
int * sections,
int il) {
const int64_t n_embd_head = hparams.n_embd_head_v;
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
// Order: joint QG projection, QG split, Q norm, KV projection, K norm, RoPE, attention
// Qwen3Next uses a single Q projection that outputs query + gate
ggml_tensor * Qcur_full = build_lora_mm(model.layers[il].wq, cur); // [ (n_embd_head * 2) * n_head, n_tokens ]
cb(Qcur_full, "Qcur_full", il);
ggml_tensor * Qcur = ggml_view_3d(ctx0, Qcur_full, n_embd_head, n_head, n_tokens,
ggml_element_size(Qcur_full) * n_embd_head * 2,
ggml_element_size(Qcur_full) * n_embd_head * 2 * n_head, 0);
cb(Qcur, "Qcur_reshaped", il);
// Apply Q normalization
Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, nullptr, LLM_NORM_RMS, il);
cb(Qcur, "Qcur_normed", il);
ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
cb(Kcur, "Kcur", il);
ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
cb(Vcur, "Vcur", il);
// Apply K normalization
Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, nullptr, LLM_NORM_RMS, il);
cb(Kcur, "Kcur_normed", il);
ggml_tensor * gate = ggml_view_3d(ctx0, Qcur_full, n_embd_head, n_head, n_tokens,
ggml_element_size(Qcur_full) * n_embd_head * 2,
ggml_element_size(Qcur_full) * n_embd_head * 2 * n_head,
ggml_element_size(Qcur_full) * n_embd_head);
gate = ggml_cont_2d(ctx0, gate, n_embd_head * n_head, n_tokens);
cb(gate, "gate_reshaped", il);
Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
// Apply IMRoPE
Qcur = ggml_rope_multi(
ctx0, Qcur, inp_pos, nullptr,
n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
Kcur = ggml_rope_multi(
ctx0, Kcur, inp_pos, nullptr,
n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
cb(Qcur, "Qcur", il);
cb(Kcur, "Kcur", il);
cb(Vcur, "Vcur", il);
// Attention computation
const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
cur = build_attn(inp,
nullptr, nullptr,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
cb(cur, "attn_pregate", il);
ggml_tensor * gate_sigmoid = ggml_sigmoid(ctx0, gate);
cb(gate_sigmoid, "gate_sigmoid", il);
cur = ggml_mul(ctx0, cur, gate_sigmoid);
cb(cur, "attn_gated", il);
cur = build_lora_mm(model.layers[il].wo, cur);
cb(cur, "attn_output", il);
return cur;
}
ggml_tensor * llm_build_qwen35moe ::build_layer_attn_linear(
llm_graph_input_rs * inp,
ggml_tensor * cur,
ggml_tensor * causal_mask,
ggml_tensor * identity,
ggml_tensor * diag_mask,
int il) {
const auto * mctx_cur = inp->mctx;
const int64_t d_inner = hparams.ssm_d_inner;
const int64_t n_seqs = ubatch.n_seqs;
const int64_t head_k_dim = hparams.ssm_d_state;
const int64_t num_k_heads = hparams.ssm_n_group;
const int64_t num_v_heads = hparams.ssm_dt_rank;
const int64_t head_v_dim = d_inner / num_v_heads;
const int64_t n_seq_tokens = ubatch.n_seq_tokens;
const auto kv_head = mctx_cur->get_head();
GGML_ASSERT(n_seqs != 0);
GGML_ASSERT(ubatch.equal_seqs());
GGML_ASSERT(ubatch.n_tokens == n_seq_tokens * n_seqs);
// Input projections
auto qkvz = build_qkvz(cur, il);
ggml_tensor * qkv_mixed = qkvz.first;
ggml_tensor * z = qkvz.second;
ggml_tensor * beta = build_lora_mm(model.layers[il].ssm_beta, cur);
beta = ggml_reshape_4d(ctx0, beta, num_v_heads, 1, n_seq_tokens, n_seqs);
cb(beta, "beta", il);
ggml_tensor * alpha = build_lora_mm(model.layers[il].ssm_alpha, cur);
alpha = ggml_cont_3d(ctx0, alpha, num_v_heads, n_seq_tokens, n_seqs);
cb(alpha, "alpha", il);
ggml_tensor * alpha_biased = ggml_add(ctx0, alpha, model.layers[il].ssm_dt);
ggml_tensor * alpha_softplus = ggml_softplus(ctx0, alpha_biased);
cb(alpha_softplus, "a_softplus", il);
ggml_tensor * gate = ggml_mul(ctx0, alpha_softplus, model.layers[il].ssm_a); // -A_log.exp() * softplus
cb(gate, "gate", il);
// Get convolution states from cache
ggml_tensor * conv_states_all = mctx_cur->get_r_l(il);
ggml_tensor * ssm_states_all = mctx_cur->get_s_l(il);
// bool use_precomputed_states = n_seq_tokens == 1 && mctx_cur->has_previous_state();
// Build the convolution states tensor
ggml_tensor * conv_states = build_rs(inp, conv_states_all, hparams.n_embd_r(), n_seqs);
cb(conv_states, "conv_states", il);
// Calculate convolution kernel size
ggml_tensor * conv_kernel = model.layers[il].ssm_conv1d;
const int64_t conv_kernel_size = conv_kernel->ne[0];
const int64_t conv_channels = d_inner + 2 * hparams.ssm_n_group * hparams.ssm_d_state;
conv_states = ggml_reshape_3d(ctx0, conv_states, conv_kernel_size - 1, conv_channels, n_seqs);
cb(conv_states, "conv_states_reshaped", il);
qkv_mixed = ggml_permute(ctx0, qkv_mixed, 1, 0, 2, 3);
cb(qkv_mixed, "qkv_mixed_permuted", il);
ggml_tensor * conv_input = ggml_concat(ctx0, conv_states, qkv_mixed, 0);
cb(conv_input, "conv_input", il);
// Update convolution state cache
// Extract the last (conv_kernel_size - 1) states from conv_input
ggml_tensor * last_conv_states =
ggml_view_3d(ctx0, conv_input, conv_kernel_size - 1, conv_channels, n_seqs, conv_input->nb[1],
conv_input->nb[2], (conv_input->ne[0] - conv_states->ne[0]) * ggml_element_size(conv_input));
cb(last_conv_states, "last_conv_states", il);
ggml_tensor * state_update_target =
ggml_view_1d(ctx0, conv_states_all, (conv_kernel_size - 1) * conv_channels * n_seqs,
kv_head * (conv_kernel_size - 1) * conv_channels * ggml_element_size(conv_states_all));
cb(state_update_target, "state_update_target", il);
ggml_build_forward_expand(gf, ggml_cpy(ctx0, last_conv_states, state_update_target));
cb(conv_states_all, "conv_states_updated", il);
// Apply SSM convolution
ggml_tensor * conv_output_proper = ggml_ssm_conv(ctx0, conv_input, conv_kernel);
cb(conv_output_proper, "conv_output_raw", il);
ggml_tensor * conv_output_silu = ggml_silu(ctx0, conv_output_proper);
cb(conv_output_silu, "conv_output_silu", il);
ggml_tensor * conv_qkv_mix = conv_output_silu;
// Calculate the total conv dimension
int64_t qkv_dim = head_k_dim * num_k_heads * 2 + head_v_dim * num_v_heads;
int64_t nb1_qkv = ggml_row_size(conv_qkv_mix->type, qkv_dim);
// Extract the convolved Q, K, V from conv_output
ggml_tensor * q_conv =
ggml_view_2d(ctx0, conv_qkv_mix, head_k_dim * num_k_heads, n_seq_tokens * n_seqs, nb1_qkv, 0);
cb(q_conv, "q_conv", il);
ggml_tensor * k_conv =
ggml_view_2d(ctx0, conv_qkv_mix, head_k_dim * num_k_heads, n_seq_tokens * n_seqs, nb1_qkv,
head_k_dim * num_k_heads * ggml_element_size(conv_qkv_mix));
cb(k_conv, "k_conv", il);
ggml_tensor * v_conv =
ggml_view_2d(ctx0, conv_qkv_mix, head_v_dim * num_v_heads, n_seq_tokens * n_seqs, nb1_qkv,
2 * head_k_dim * num_k_heads * ggml_element_size(conv_qkv_mix));
cb(v_conv, "v_conv", il);
// Unsqueeze them
q_conv = ggml_cont_4d(ctx0, q_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs);
k_conv = ggml_cont_4d(ctx0, k_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs);
v_conv = ggml_cont_4d(ctx0, v_conv, head_v_dim, num_v_heads, n_seq_tokens, n_seqs);
ggml_tensor * state = build_rs(inp, ssm_states_all, hparams.n_embd_s(), n_seqs);
state = ggml_reshape_4d(ctx0, state, head_v_dim, head_v_dim * num_v_heads, 1, n_seqs);
cb(state, "state_predelta", il);
// if head keys and value keys are different, repeat Q/K to match V's head count
// V heads are in tiled order (from conversion), so simple tiled repeat works
if (num_k_heads != num_v_heads) {
GGML_ASSERT(num_v_heads % num_k_heads == 0);
q_conv = ggml_repeat_4d(ctx0, q_conv, head_k_dim, num_v_heads, n_seq_tokens, n_seqs);
k_conv = ggml_repeat_4d(ctx0, k_conv, head_k_dim, num_v_heads, n_seq_tokens, n_seqs);
}
cb(q_conv, "q_conv_predelta", il);
cb(k_conv, "k_conv_predelta", il);
cb(v_conv, "v_conv_predelta", il);
// Choose between build_delta_net_chunking, build_delta_net_recurrent, and build_delta_net_autoregressive based on n_tokens
std::pair<ggml_tensor *, ggml_tensor *> attn_out; // pair of (output, new_state)
if (n_seq_tokens == 1) {
attn_out = build_delta_net_autoregressive(q_conv, k_conv, v_conv, gate, beta, state, il);
} else {
attn_out = build_delta_net_chunking(q_conv, k_conv, v_conv, gate, beta, state, causal_mask, identity, diag_mask, il);
}
ggml_tensor * output = attn_out.first;
ggml_tensor * new_state = attn_out.second;
cb(output, "attn_output", il);
cb(new_state, "new_state", il);
// Update the recurrent states
ggml_build_forward_expand(gf,
ggml_cpy(ctx0, new_state,
ggml_view_1d(ctx0, ssm_states_all, hparams.n_embd_s() * n_seqs,
kv_head * hparams.n_embd_s() * ggml_element_size(ssm_states_all))));
// Reshape both attn_out_final and z to 2D tensors for normalization
// attn_out_final: [head_dim, n_heads, n_tokens, n_seqs] -> [n_heads * n_tokens * n_seqs, head_dim]
ggml_tensor * attn_out_2d_final = ggml_reshape_2d(ctx0, output, head_v_dim, num_v_heads * n_seq_tokens * n_seqs);
// z: [head_dim, n_heads, n_tokens, n_seqs] -> [n_heads * n_tokens * n_seqs, head_dim]
ggml_tensor * z_2d = ggml_reshape_2d(ctx0, z, head_v_dim, num_v_heads * n_seq_tokens * n_seqs);
// Apply gated normalization: self.norm(core_attn_out, z)
ggml_tensor * attn_out_norm = build_norm_gated(attn_out_2d_final, model.layers[il].ssm_norm, z_2d, il);
// Final reshape: [head_dim, n_heads, n_tokens, n_seqs] -> [n_tokens, n_seqs, n_heads * head_dim]
ggml_tensor * final_output = ggml_reshape_3d(ctx0, attn_out_norm, head_v_dim * num_v_heads, n_seq_tokens, n_seqs);
cb(final_output, "final_output", il);
// Output projection
cur = build_lora_mm(model.layers[il].ssm_out, final_output);
cb(cur, "linear_attn_out", il);
// Reshape back to original dimensions
cur = ggml_cont_2d(ctx0, cur, n_embd, n_seq_tokens * n_seqs);
return cur;
}
ggml_tensor * llm_build_qwen35moe ::build_layer_ffn(ggml_tensor * cur, const int il) {
// Check if this is an MoE layer
GGML_ASSERT(model.layers[il].ffn_gate_inp != nullptr);
ggml_tensor * moe_out =
build_moe_ffn(cur,
model.layers[il].ffn_gate_inp, model.layers[il].ffn_up_exps,
model.layers[il].ffn_gate_exps, model.layers[il].ffn_down_exps,
nullptr,
n_expert, n_expert_used, LLM_FFN_SILU,
true, false, 0.0, LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, il);
cb(moe_out, "ffn_moe_out", il);
// Add shared experts if present - following Qwen3Next reference implementation
if (model.layers[il].ffn_up_shexp != nullptr) {
ggml_tensor * ffn_shexp =
build_ffn(cur,
model.layers[il].ffn_up_shexp, NULL, NULL,
model.layers[il].ffn_gate_shexp, NULL, NULL,
model.layers[il].ffn_down_shexp, NULL, NULL,
NULL,
LLM_FFN_SILU, LLM_FFN_PAR, il);
cb(ffn_shexp, "ffn_shexp", il);
// Apply shared expert gating as in the reference implementation
// The shared expert has its own gate that is sigmoided
// Note: ffn_gate_inp_shexp is the shared expert gate (outputs 1 value per token)
ggml_tensor * shared_gate = build_lora_mm(model.layers[il].ffn_gate_inp_shexp, cur);
cb(shared_gate, "shared_expert_gate", il);
// Apply sigmoid to the gate
shared_gate = ggml_sigmoid(ctx0, shared_gate);
cb(shared_gate, "shared_expert_gate_sigmoid", il);
// Apply the gate to the shared expert output
ffn_shexp = ggml_mul(ctx0, ffn_shexp, shared_gate);
cb(ffn_shexp, "ffn_shexp_gated", il);
cur = ggml_add(ctx0, moe_out, ffn_shexp);
cb(cur, "ffn_out", il);
} else {
cur = moe_out;
}
return cur;
}

8
tests/CMakeLists.txt
View File

@ -11,7 +11,9 @@ function(llama_build source)
add_executable(${TEST_TARGET} ${TEST_SOURCES})
target_link_libraries(${TEST_TARGET} PRIVATE common)
install(TARGETS ${TEST_TARGET} RUNTIME)
if (LLAMA_TESTS_INSTALL)
install(TARGETS ${TEST_TARGET} RUNTIME)
endif()
endfunction()
function(llama_test target)
@ -100,7 +102,9 @@ function(llama_build_and_test source)
endif()
add_executable(${TEST_TARGET} ${TEST_SOURCES})
install(TARGETS ${TEST_TARGET} RUNTIME)
if (LLAMA_TESTS_INSTALL)
install(TARGETS ${TEST_TARGET} RUNTIME)
endif()
target_link_libraries(${TEST_TARGET} PRIVATE common)
add_test(

23
tests/test-backend-ops.cpp
View File

@ -5894,33 +5894,36 @@ struct test_pad_ext : public test_case {
const int rp2;
const int lp3;
const int rp3;
const bool v;
const int tfrm; // 0 - none, 1 - non-cont, 2 - perm
const bool circular;
std::string vars() override {
return VARS_TO_STR12(type, ne_a, lp0, rp0, lp1, rp1, lp2, rp2, lp3, rp3, v, circular);
return VARS_TO_STR12(type, ne_a, lp0, rp0, lp1, rp1, lp2, rp2, lp3, rp3, tfrm, circular);
}
test_pad_ext(ggml_type type = GGML_TYPE_F32,
std::array<int64_t, 4> ne_a = {512, 512, 3, 1},
int lp0 = 1, int rp0 = 1, int lp1 = 1, int rp1 = 1,
int lp2 = 1, int rp2 = 1, int lp3 = 1, int rp3 = 1,
bool v = false, bool circular = false)
int tfrm = 0, bool circular = false)
: type(type), ne_a(ne_a), lp0(lp0), rp0(rp0), lp1(lp1), rp1(rp1), lp2(lp2), rp2(rp2), lp3(lp3), rp3(rp3),
v(v), circular(circular) {}
tfrm(tfrm), circular(circular) {}
ggml_tensor * build_graph(ggml_context * ctx) override {
ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne_a.data());
ggml_set_name(a, "a");
if (v) {
if (tfrm == 1) {
a = ggml_view_4d(ctx, a, (a->ne[0] + 1) / 2, (a->ne[1] + 1) / 2, (a->ne[2] + 1) / 2, (a->ne[3] + 1) / 2, a->nb[1], a->nb[2], a->nb[3], 0);
ggml_set_name(a, "view of a");
} else if (tfrm == 2) {
a = ggml_permute(ctx, a, 2, 1, 0, 3);
ggml_set_name(a, "permuted a");
}
ggml_tensor * out = circular
? ggml_pad_ext_circular(ctx, a, lp0, rp0, lp1, rp1, lp2, rp2, lp3, rp3)
: ggml_pad_ext(ctx, a, lp0, rp0, lp1, rp1, lp2, rp2, lp3, rp3);
: ggml_pad_ext (ctx, a, lp0, rp0, lp1, rp1, lp2, rp2, lp3, rp3);
ggml_set_name(out, "out");
return out;
@ -8198,10 +8201,10 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
test_cases.emplace_back(new test_solve_tri(GGML_TYPE_F32, { 64, 64, 4, 4 }, { 200, 64, 4, 4 }));
test_cases.emplace_back(new test_solve_tri(GGML_TYPE_F32, { 64, 64, 4, 4 }, { 384, 64, 4, 4 }));
for (bool v : {false, true}) {
for (int tfrm : {0, 1, 2}) {
for (bool circular : {false, true}) {
test_cases.emplace_back(new test_pad_ext(GGML_TYPE_F32, {512, 512, 1, 1}, 0, 1, 0, 1, 0, 0, 0, 0, v, circular));
test_cases.emplace_back(new test_pad_ext(GGML_TYPE_F32, {11, 22, 33, 44}, 1, 2, 3, 4, 5, 6, 7, 8, v, circular));
test_cases.emplace_back(new test_pad_ext(GGML_TYPE_F32, {512, 512, 1, 1}, 0, 1, 0, 1, 0, 0, 0, 0, tfrm, circular));
test_cases.emplace_back(new test_pad_ext(GGML_TYPE_F32, {11, 22, 33, 44}, 1, 2, 3, 4, 5, 6, 7, 8, tfrm, circular));
}
}
@ -8520,7 +8523,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_perf() {
test_cases.emplace_back(new test_rope(type, { 80, 32, 512, 1}, 20, GGML_ROPE_TYPE_NEOX, 512, 1.0f, 0.0f, 1.0f, ff, v, fw)); // neox (stablelm)
test_cases.emplace_back(new test_rope(type, { 64, 8, 512, 1}, 64, GGML_ROPE_TYPE_NEOX, 512, 1.0f, 0.0f, 1.0f, ff, v, fw)); // neox (falcon 40B)
test_cases.emplace_back(new test_rope(type, {128, 12, 512, 1}, 128, GGML_ROPE_TYPE_MROPE, 512, 1.0f, 0.0f, 1.0f, ff, v, fw)); // rope_multi,m-rope (qwen2vl 2B)
test_cases.emplace_back(new test_rope(type, {128, 12, 2, 1}, 128, GGML_ROPE_TYPE_IMROPE, 512, 1.0f, 0.0f, 1.0f, ff, v, fw)); // rope_multi,imrope (qwen3vl 2B)
test_cases.emplace_back(new test_rope(type, {128, 12, 512, 1}, 128, GGML_ROPE_TYPE_IMROPE, 512, 1.0f, 0.0f, 1.0f, ff, v, fw)); // rope_multi,imrope (qwen3vl 2B)
test_cases.emplace_back(new test_rope(type, { 80, 16, 2, 1}, 80, GGML_ROPE_TYPE_VISION, 512, 1.0f, 0.0f, 1.0f, ff, v, fw)); // rope_multi,m-rope (qwen2vl ViT)
}
}

136
tools/mtmd/clip.cpp
View File

@ -10,6 +10,7 @@
#include "ggml-backend.h"
#include "gguf.h"
#include <algorithm>
#include <cassert>
#include <cmath>
#include <cstdlib>
@ -1116,9 +1117,8 @@ struct clip_model_loader {
case PROJECTOR_TYPE_LFM2:
{
get_u32(KEY_PROJ_SCALE_FACTOR, hparams.n_merge, false);
// ref: https://huggingface.co/LiquidAI/LFM2-VL-3B/blob/main/preprocessor_config.json
// config above specifies number of tokens after downsampling, while here it is before, relax lowerbound to 64
hparams.set_limit_image_tokens(64, 1024);
// ref: https://huggingface.co/LiquidAI/LFM2.5-VL-1.6B/blob/main/processor_config.json
hparams.set_limit_image_tokens(64, 256);
} break;
case PROJECTOR_TYPE_PIXTRAL:
case PROJECTOR_TYPE_LIGHTONOCR:
@ -2807,6 +2807,119 @@ private:
}
};
// ref: https://github.com/huggingface/transformers/blob/v5.1.0/src/transformers/models/lfm2_vl/image_processing_lfm2_vl_fast.py
// some of the logic is similar to llava_uhd, but with different hyperparameters and some logic is unique (e.g. grid layout)
struct lfm2_vl_image_processor {
// ref: https://huggingface.co/LiquidAI/LFM2.5-VL-1.6B/blob/main/processor_config.json
static constexpr int min_tiles = 2;
static constexpr int max_tiles = 10;
static constexpr float max_pixels_tolerance = 2.0f;
static constexpr int tile_size = 512;
static llava_uhd::slice_instructions get_slice_instructions(struct clip_ctx * ctx, const clip_image_size & original_size) {
llava_uhd::slice_instructions inst;
const auto & params = ctx->model.hparams;
const int align_size = params.patch_size * params.n_merge;
inst.interpolation_overview = img_tool::RESIZE_ALGO_BILINEAR;
inst.interpolation_refined = img_tool::RESIZE_ALGO_BILINEAR;
inst.overview_size = img_tool::calc_size_preserved_ratio(original_size, align_size, params.image_min_pixels, params.image_max_pixels);
// tile if either dimension exceeds tile_size with tolerance
const bool needs_tiling = original_size.width > tile_size * max_pixels_tolerance || original_size.height > tile_size * max_pixels_tolerance;
if (!needs_tiling) {
inst.refined_size = clip_image_size{0, 0};
inst.grid_size = clip_image_size{0, 0};
return inst;
}
const clip_image_size grid = get_grid_layout(original_size.height, original_size.width);
inst.grid_size = grid;
inst.refined_size = clip_image_size{tile_size * grid.width, tile_size * grid.height};
LOG_DBG("%s: original size: %d x %d, overview size: %d x %d, refined size: %d x %d, grid size: %d x %d\n",
__func__,
original_size.width, original_size.height,
inst.overview_size.width, inst.overview_size.height,
inst.refined_size.width, inst.refined_size.height,
grid.width, grid.height);
for (int row = 0; row < grid.height; row++) {
for (int col = 0; col < grid.width; col++) {
llava_uhd::slice_coordinates slice;
slice.x = col * tile_size;
slice.y = row * tile_size;
slice.size = clip_image_size{tile_size, tile_size};
inst.slices.push_back(slice);
LOG_DBG("%s: slice %d: x=%d, y=%d, size=%d x %d\n",
__func__, (int)inst.slices.size() - 1,
slice.x, slice.y, slice.size.width, slice.size.height);
}
}
return inst;
}
private:
static clip_image_size find_closest_aspect_ratio(
float aspect_ratio,
const std::vector<clip_image_size> & target_ratios,
int width, int height) {
float best_ratio_diff = std::numeric_limits<float>::max();
clip_image_size best_ratio = {1, 1};
const float area = static_cast<float>(width * height);
for (const auto & ratio : target_ratios) {
const float target_aspect_ratio = static_cast<float>(ratio.width) / ratio.height;
const float ratio_diff = std::abs(aspect_ratio - target_aspect_ratio);
if (ratio_diff < best_ratio_diff) {
best_ratio_diff = ratio_diff;
best_ratio = ratio;
} else if (ratio_diff == best_ratio_diff) {
const float target_area = static_cast<float>(tile_size * tile_size * ratio.width * ratio.height);
if (area > 0.5f * target_area) {
best_ratio = ratio;
}
}
}
return best_ratio;
}
static std::vector<clip_image_size> get_target_ratios() {
std::vector<clip_image_size> ratios;
for (int n = min_tiles; n <= max_tiles; n++) {
for (int w = 1; w <= n; w++) {
for (int h = 1; h <= n; h++) {
if (w * h >= min_tiles && w * h <= max_tiles) {
bool found = false;
for (const auto & r : ratios) {
if (r.width == w && r.height == h) {
found = true;
break;
}
}
if (!found) {
ratios.push_back({w, h});
}
}
}
}
}
std::sort(ratios.begin(), ratios.end(), [](const clip_image_size & a, const clip_image_size & b) {
return a.width * a.height < b.width * b.height;
});
return ratios;
}
static clip_image_size get_grid_layout(int height, int width) {
const float aspect_ratio = static_cast<float>(width) / height;
const auto ratios = get_target_ratios();
return find_closest_aspect_ratio(aspect_ratio, ratios, width, height);
}
};
// returns the normalized float tensor for llava-1.5, for spatial_unpad with anyres processing for llava-1.6 it returns the normalized image patch tensors as a vector
// res_imgs memory is being allocated here, previous allocations will be freed if found
bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, struct clip_image_f32_batch * res_imgs) {
@ -3021,6 +3134,20 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, str
} break;
case PROJECTOR_TYPE_LFM2:
{
auto const inst = lfm2_vl_image_processor::get_slice_instructions(ctx, original_size);
std::vector<clip_image_u8_ptr> imgs = llava_uhd::slice_image(img, inst);
for (size_t i = 0; i < imgs.size(); ++i) {
clip_image_f32_ptr res(clip_image_f32_init());
normalize_image_u8_to_f32(*imgs[i], *res, params.image_mean, params.image_std);
res_imgs->entries.push_back(std::move(res));
}
res_imgs->grid_x = inst.grid_size.width;
res_imgs->grid_y = inst.grid_size.height;
} break;
case PROJECTOR_TYPE_KIMIVL:
{
GGML_ASSERT(params.image_min_pixels > 0 && params.image_max_pixels > 0);
@ -3032,8 +3159,7 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, str
const std::array<uint8_t, 3> pad_color = {122, 116, 104};
clip_image_u8 resized_img;
const bool pad = (ctx->proj_type() != PROJECTOR_TYPE_LFM2);
img_tool::resize(*img, resized_img, target_size, img_tool::RESIZE_ALGO_BILINEAR, pad, pad_color);
img_tool::resize(*img, resized_img, target_size, img_tool::RESIZE_ALGO_BILINEAR, true, pad_color);
clip_image_f32_ptr res(clip_image_f32_init());
normalize_image_u8_to_f32(resized_img, *res, params.image_mean, params.image_std);
res_imgs->entries.push_back(std::move(res));

4
tools/mtmd/models/qwen3vl.cpp
View File

@ -182,7 +182,9 @@ ggml_cgraph * clip_graph_qwen3vl::build() {
model.mm_1_w, model.mm_1_b,
ffn_op_type::FFN_GELU, -1);
embeddings = ggml_concat(ctx0, embeddings, deepstack_features, 0); // concat along the feature dimension
if (deepstack_features) {
embeddings = ggml_concat(ctx0, embeddings, deepstack_features, 0);
} // concat along the feature dimension
// build the graph
ggml_build_forward_expand(gf, embeddings);

19
tools/mtmd/mtmd.cpp
View File

@ -85,6 +85,7 @@ enum mtmd_slice_tmpl {
MTMD_SLICE_TMPL_MINICPMV_2_6,
MTMD_SLICE_TMPL_LLAMA4,
MTMD_SLICE_TMPL_IDEFICS3,
MTMD_SLICE_TMPL_LFM2,
};
const char * mtmd_default_marker() {
@ -307,9 +308,19 @@ struct mtmd_context {
img_end = "<|im_end|>";
} else if (proj == PROJECTOR_TYPE_LFM2) {
img_beg = "<|image_start|>";
img_end = "<|image_end|>";
// multi-tile:
// <|image_start|>
// <|img_row_1_col_1|> (tile) <|img_row_1_col_2|> (tile) ...
// <|img_thumbnail|> (thumbnail)
// <|image_end|>
// single-tile:
// <|image_start|> (image) <|image_end|>
img_beg = "<|image_start|>";
img_end = "<|image_end|>";
slice_tmpl = MTMD_SLICE_TMPL_LFM2;
sli_img_start_tmpl = "<|img_row_%d_col_%d|>";
tok_ov_img_start = {lookup_token("<|img_thumbnail|>")};
ov_img_first = false;
} else if (proj == PROJECTOR_TYPE_GLM4V) {
img_beg = "<|begin_of_image|>";
img_end = "<|end_of_image|>";
@ -562,11 +573,13 @@ struct mtmd_tokenizer {
}
// handle llava-uhd style preprocessing
const bool has_tiling_grid = batch_f32.grid_x > 0 && batch_f32.grid_y > 0;
if (
ctx->slice_tmpl == MTMD_SLICE_TMPL_MINICPMV_2_5
|| ctx->slice_tmpl == MTMD_SLICE_TMPL_MINICPMV_2_6
|| ctx->slice_tmpl == MTMD_SLICE_TMPL_LLAMA4
|| ctx->slice_tmpl == MTMD_SLICE_TMPL_IDEFICS3
|| (ctx->slice_tmpl == MTMD_SLICE_TMPL_LFM2 && has_tiling_grid)
) {
const int n_col = batch_f32.grid_x;
const int n_row = batch_f32.grid_y;

6
tools/server/server-context.cpp
View File

@ -3584,6 +3584,8 @@ void server_routes::init_routes() {
auto res = create_response();
std::vector<raw_buffer> files;
json body = convert_responses_to_chatcmpl(json::parse(req.body));
SRV_DBG("%s\n", "Request converted: OpenAI Responses -> OpenAI Chat Completions");
SRV_DBG("converted request: %s\n", body.dump().c_str());
json body_parsed = oaicompat_chat_params_parse(
body,
meta->chat_params,
@ -3600,6 +3602,8 @@ void server_routes::init_routes() {
auto res = create_response();
std::vector<raw_buffer> files;
json body = convert_anthropic_to_oai(json::parse(req.body));
SRV_DBG("%s\n", "Request converted: Anthropic -> OpenAI Chat Completions");
SRV_DBG("converted request: %s\n", body.dump().c_str());
json body_parsed = oaicompat_chat_params_parse(
body,
meta->chat_params,
@ -3616,6 +3620,8 @@ void server_routes::init_routes() {
auto res = create_response();
std::vector<raw_buffer> files;
json body = convert_anthropic_to_oai(json::parse(req.body));
SRV_DBG("%s\n", "Request converted: Anthropic -> OpenAI Chat Completions");
SRV_DBG("converted request: %s\n", body.dump().c_str());
json body_parsed = oaicompat_chat_params_parse(
body,
meta->chat_params,

4
tools/server/server-task.cpp
View File

@ -80,7 +80,6 @@ json task_params::to_json(bool only_metrics) const {
{"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},
@ -144,7 +143,6 @@ json task_params::to_json(bool only_metrics) const {
{"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},
@ -257,12 +255,10 @@ task_params server_task::params_from_json_cmpl(
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

4
tools/tts/README.md
View File

@ -34,7 +34,7 @@ $ build/bin/llama-quantize models/outetts-0.2-0.5B-f16.gguf \
```
The quantized model will be `models/outetts-0.2-0.5B-q8_0.gguf`.
Next we do something simlar for the audio decoder. First download or checkout
Next we do something similar for the audio decoder. First download or checkout
the model for the voice decoder:
```console
$ pushd models
@ -42,7 +42,7 @@ $ git clone --branch main --single-branch --depth 1 https://huggingface.co/novat
$ cd WavTokenizer-large-speech-75token && git lfs install && git lfs pull
$ popd
```
This model file is PyTorch checkpoint (.ckpt) and we first need to convert it to
This model file is a PyTorch checkpoint (.ckpt) and we first need to convert it to
huggingface format:
```console
(venv) python tools/tts/convert_pt_to_hf.py \