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 dev-refactoring 

# Conflicts:
#	ggml/src/ggml-backend.cpp
#	src/llama.cpp
This commit is contained in:
hongruichen 2024-10-28 13:31:54 +08:00
parent 4abaf7d87e 8841ce3f43
commit c42433cf76
91 changed files with 10453 additions and 3733 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
CMakeLists.txt
View File

@ -88,6 +88,10 @@ if (NOT DEFINED GGML_LLAMAFILE)
set(GGML_LLAMAFILE_DEFAULT ON)
endif()
if (NOT DEFINED GGML_AMX)
set(GGML_AMX ON)
endif()
if (NOT DEFINED GGML_CUDA_GRAPHS)
set(GGML_CUDA_GRAPHS_DEFAULT ON)
endif()

24
Makefile
View File

@ -93,11 +93,6 @@ GGML_METAL := 1
DEPRECATE_WARNING := 1
endif
ifdef LLAMA_OPENMP
GGML_OPENMP := 1
DEPRECATE_WARNING := 1
endif
ifdef LLAMA_RPC
GGML_RPC := 1
DEPRECATE_WARNING := 1
@ -584,6 +579,11 @@ ifndef GGML_NO_LLAMAFILE
OBJ_GGML += ggml/src/llamafile/sgemm.o
endif
ifndef GGML_NO_AMX
MK_CPPFLAGS += -DGGML_USE_AMX
OBJ_GGML += ggml/src/ggml-amx.o ggml/src/ggml-amx/mmq.o
endif
ifdef GGML_RPC
MK_CPPFLAGS += -DGGML_USE_RPC
OBJ_GGML += ggml/src/ggml-rpc.o
@ -1087,6 +1087,19 @@ ggml/src/llamafile/sgemm.o: \
$(CXX) $(CXXFLAGS) -c $< -o $@
endif # GGML_NO_LLAMAFILE
ifndef GGML_NO_AMX
ggml/src/ggml-amx.o: \
ggml/src/ggml-amx.cpp \
ggml/include/ggml-amx.h
$(CXX) $(CXXFLAGS) -c $< -o $@
ggml/src/ggml-amx/mmq.o: \
ggml/src/ggml-amx/mmq.cpp \
ggml/src/ggml-amx/mmq.h \
ggml/include/ggml.h
$(CXX) $(CXXFLAGS) -c $< -o $@
endif
ifdef GGML_RPC
ggml/src/ggml-rpc.o: \
ggml/src/ggml-rpc.cpp \
@ -1238,6 +1251,7 @@ clean:
rm -vrf ggml/src/ggml-metal-embed.metal
rm -vrf ggml/src/ggml-cuda/*.o
rm -vrf ggml/src/ggml-cuda/template-instances/*.o
rm -vrf ggml/src/ggml-amx/*.o
rm -rvf $(BUILD_TARGETS)
rm -rvf $(TEST_TARGETS)
rm -f vulkan-shaders-gen ggml/src/ggml-vulkan-shaders.hpp ggml/src/ggml-vulkan-shaders.cpp

12
README.md
View File

@ -29,9 +29,9 @@ variety of hardware - locally and in the cloud.
- Plain C/C++ implementation without any dependencies
- Apple silicon is a first-class citizen - optimized via ARM NEON, Accelerate and Metal frameworks
- AVX, AVX2 and AVX512 support for x86 architectures
- AVX, AVX2, AVX512 and AMX support for x86 architectures
- 1.5-bit, 2-bit, 3-bit, 4-bit, 5-bit, 6-bit, and 8-bit integer quantization for faster inference and reduced memory use
- Custom CUDA kernels for running LLMs on NVIDIA GPUs (support for AMD GPUs via HIP)
- Custom CUDA kernels for running LLMs on NVIDIA GPUs (support for AMD GPUs via HIP and Moore Threads MTT GPUs via MUSA)
- Vulkan and SYCL backend support
- CPU+GPU hybrid inference to partially accelerate models larger than the total VRAM capacity
@ -93,6 +93,7 @@ Typically finetunes of the base models below are supported as well.
- [x] [FalconMamba Models](https://huggingface.co/collections/tiiuae/falconmamba-7b-66b9a580324dd1598b0f6d4a)
- [x] [Jais](https://huggingface.co/inceptionai/jais-13b-chat)
- [x] [Bielik-11B-v2.3](https://huggingface.co/collections/speakleash/bielik-11b-v23-66ee813238d9b526a072408a)
- [x] [RWKV-6](https://github.com/BlinkDL/RWKV-LM)
(instructions for supporting more models: [HOWTO-add-model.md](./docs/development/HOWTO-add-model.md))
@ -122,6 +123,7 @@ Typically finetunes of the base models below are supported as well.
- Rust (nicer API): [mdrokz/rust-llama.cpp](https://github.com/mdrokz/rust-llama.cpp)
- Rust (more direct bindings): [utilityai/llama-cpp-rs](https://github.com/utilityai/llama-cpp-rs)
- C#/.NET: [SciSharp/LLamaSharp](https://github.com/SciSharp/LLamaSharp)
- C#/VB.NET (more features - community license): [LM-Kit.NET](https://docs.lm-kit.com/lm-kit-net/index.html)
- Scala 3: [donderom/llm4s](https://github.com/donderom/llm4s)
- Clojure: [phronmophobic/llama.clj](https://github.com/phronmophobic/llama.clj)
- React Native: [mybigday/llama.rn](https://github.com/mybigday/llama.rn)
@ -130,6 +132,8 @@ Typically finetunes of the base models below are supported as well.
- Flutter/Dart: [netdur/llama_cpp_dart](https://github.com/netdur/llama_cpp_dart)
- PHP (API bindings and features built on top of llama.cpp): [distantmagic/resonance](https://github.com/distantmagic/resonance) [(more info)](https://github.com/ggerganov/llama.cpp/pull/6326)
- Guile Scheme: [guile_llama_cpp](https://savannah.nongnu.org/projects/guile-llama-cpp)
- Swift [srgtuszy/llama-cpp-swift](https://github.com/srgtuszy/llama-cpp-swift)
- Swift [ShenghaiWang/SwiftLlama](https://github.com/ShenghaiWang/SwiftLlama)
**UI:**
@ -170,6 +174,7 @@ Unless otherwise noted these projects are open-source with permissive licensing:
- [LARS - The LLM & Advanced Referencing Solution](https://github.com/abgulati/LARS) (AGPL)
- [LLMUnity](https://github.com/undreamai/LLMUnity) (MIT)
- [Llama Assistant](https://github.com/vietanhdev/llama-assistant) (GPL)
- [PocketPal AI - An iOS and Android App](https://github.com/a-ghorbani/pocketpal-ai) (MIT)
*(to have a project listed here, it should clearly state that it depends on `llama.cpp`)*
@ -185,6 +190,7 @@ Unless otherwise noted these projects are open-source with permissive licensing:
- [Paddler](https://github.com/distantmagic/paddler) - Stateful load balancer custom-tailored for llama.cpp
- [GPUStack](https://github.com/gpustack/gpustack) - Manage GPU clusters for running LLMs
- [llama_cpp_canister](https://github.com/onicai/llama_cpp_canister) - llama.cpp as a smart contract on the Internet Computer, using WebAssembly
**Games:**
- [Lucy's Labyrinth](https://github.com/MorganRO8/Lucys_Labyrinth) - A simple maze game where agents controlled by an AI model will try to trick you.
@ -413,7 +419,7 @@ Please refer to [Build llama.cpp locally](./docs/build.md)
| [BLAS](./docs/build.md#blas-build) | All |
| [BLIS](./docs/backend/BLIS.md) | All |
| [SYCL](./docs/backend/SYCL.md) | Intel and Nvidia GPU |
| [MUSA](./docs/build.md#musa) | Moore Threads GPU |
| [MUSA](./docs/build.md#musa) | Moore Threads MTT GPU |
| [CUDA](./docs/build.md#cuda) | Nvidia GPU |
| [hipBLAS](./docs/build.md#hipblas) | AMD GPU |
| [Vulkan](./docs/build.md#vulkan) | GPU |

2
ci/run.sh
View File

@ -53,7 +53,7 @@ if [ ! -z ${GG_BUILD_SYCL} ]; then
exit 1
fi
CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_SYCL=1 DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON"
CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_SYCL=1 -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON"
fi
if [ ! -z ${GG_BUILD_VULKAN} ]; then

359
common/arg.cpp
View File

File diff suppressed because it is too large Load Diff

41
common/common.cpp
View File

@ -12,6 +12,7 @@
#include <algorithm>
#include <cinttypes>
#include <climits>
#include <cmath>
#include <codecvt>
#include <cstdarg>
@ -23,10 +24,10 @@
#include <regex>
#include <sstream>
#include <string>
#include <thread>
#include <unordered_map>
#include <unordered_set>
#include <vector>
#include <thread>
#if defined(__APPLE__) && defined(__MACH__)
#include <sys/types.h>
@ -400,17 +401,19 @@ std::string common_params_get_system_info(const common_params & params) {
// String utils
//
std::vector<std::string> string_split(std::string input, char separator) {
std::vector<std::string> parts;
size_t separator_pos = input.find(separator);
while (separator_pos != std::string::npos) {
std::string part = input.substr(0, separator_pos);
parts.emplace_back(part);
input = input.substr(separator_pos + 1);
separator_pos = input.find(separator);
}
parts.emplace_back(input);
return parts;
std::string string_format(const char * fmt, ...) {
va_list ap;
va_list ap2;
va_start(ap, fmt);
va_copy(ap2, ap);
int size = vsnprintf(NULL, 0, fmt, ap);
GGML_ASSERT(size >= 0 && size < INT_MAX); // NOLINT
std::vector<char> buf(size + 1);
int size2 = vsnprintf(buf.data(), size + 1, fmt, ap2);
GGML_ASSERT(size2 == size);
va_end(ap2);
va_end(ap);
return std::string(buf.data(), size);
}
std::string string_strip(const std::string & str) {
@ -939,7 +942,7 @@ struct common_init_result common_init_from_params(common_params & params) {
}
if (llama_model_has_encoder(model)) {
llama_encode(lctx, llama_batch_get_one(tmp.data(), tmp.size(), 0, 0));
llama_encode(lctx, llama_batch_get_one(tmp.data(), tmp.size()));
llama_token decoder_start_token_id = llama_model_decoder_start_token(model);
if (decoder_start_token_id == -1) {
decoder_start_token_id = bos;
@ -948,7 +951,7 @@ struct common_init_result common_init_from_params(common_params & params) {
tmp.push_back(decoder_start_token_id);
}
if (llama_model_has_decoder(model)) {
llama_decode(lctx, llama_batch_get_one(tmp.data(), std::min(tmp.size(), (size_t) params.n_batch), 0, 0));
llama_decode(lctx, llama_batch_get_one(tmp.data(), std::min(tmp.size(), (size_t) params.n_batch)));
}
llama_kv_cache_clear(lctx);
llama_synchronize(lctx);
@ -1019,7 +1022,7 @@ static ggml_type kv_cache_type_from_str(const std::string & s) {
return GGML_TYPE_Q5_1;
}
throw std::runtime_error("Invalid cache type: " + s);
throw std::runtime_error("Unsupported cache type: " + s);
}
struct llama_context_params common_context_params_to_llama(const common_params & params) {
@ -1031,7 +1034,7 @@ struct llama_context_params common_context_params_to_llama(const common_params &
cparams.n_ubatch = params.n_ubatch;
cparams.n_threads = params.cpuparams.n_threads;
cparams.n_threads_batch = params.cpuparams_batch.n_threads == -1 ?
params.cpuparams.n_threads : params.cpuparams_batch.n_threads;
params.cpuparams.n_threads : params.cpuparams_batch.n_threads;
cparams.logits_all = params.logits_all;
cparams.embeddings = params.embedding;
cparams.rope_scaling_type = params.rope_scaling_type;
@ -2003,6 +2006,10 @@ void yaml_dump_non_result_info(FILE * stream, const common_params & params, cons
fprintf(stream, "chunks: %d # default: -1 (unlimited)\n", params.n_chunks);
fprintf(stream, "color: %s # default: false\n", params.use_color ? "true" : "false");
fprintf(stream, "ctx_size: %d # default: 512\n", params.n_ctx);
fprintf(stream, "dry_allowed_length: %d # default: 2\n", sparams.dry_allowed_length);
fprintf(stream, "dry_base: %.2f # default: 1.75\n", sparams.dry_base);
fprintf(stream, "dry_multiplier: %.1f # default: 0.0\n", sparams.dry_multiplier);
fprintf(stream, "dry_penalty_last_n: %d # default: -1 (0 = disable, -1 = context size)\n", sparams.dry_penalty_last_n);
fprintf(stream, "escape: %s # default: false\n", params.escape ? "true" : "false");
fprintf(stream, "file: # never logged, see prompt instead. Can still be specified for input.\n");
fprintf(stream, "frequency_penalty: %f # default: 0.0 \n", sparams.penalty_freq);
@ -2088,6 +2095,8 @@ void yaml_dump_non_result_info(FILE * stream, const common_params & params, cons
fprintf(stream, "top_k: %d # default: 40\n", sparams.top_k);
fprintf(stream, "top_p: %f # default: 0.95\n", sparams.top_p);
fprintf(stream, "min_p: %f # default: 0.0\n", sparams.min_p);
fprintf(stream, "xtc_probability: %f # default: 0.0\n", sparams.xtc_probability);
fprintf(stream, "xtc_threshold: %f # default: 0.1\n", sparams.xtc_threshold);
fprintf(stream, "typ_p: %f # default: 1.0\n", sparams.typ_p);
fprintf(stream, "verbose_prompt: %s # default: false\n", params.verbose_prompt ? "true" : "false");
fprintf(stream, "display_prompt: %s # default: true\n", params.display_prompt ? "true" : "false");

114
common/common.h
View File

@ -84,12 +84,15 @@ enum llama_example {
enum common_sampler_type {
COMMON_SAMPLER_TYPE_NONE = 0,
COMMON_SAMPLER_TYPE_TOP_K = 1,
COMMON_SAMPLER_TYPE_TOP_P = 2,
COMMON_SAMPLER_TYPE_MIN_P = 3,
COMMON_SAMPLER_TYPE_TFS_Z = 4,
COMMON_SAMPLER_TYPE_TYPICAL_P = 5,
COMMON_SAMPLER_TYPE_TEMPERATURE = 6,
COMMON_SAMPLER_TYPE_DRY = 1,
COMMON_SAMPLER_TYPE_TOP_K = 2,
COMMON_SAMPLER_TYPE_TOP_P = 3,
COMMON_SAMPLER_TYPE_MIN_P = 4,
COMMON_SAMPLER_TYPE_TFS_Z = 5,
COMMON_SAMPLER_TYPE_TYPICAL_P = 6,
COMMON_SAMPLER_TYPE_TEMPERATURE = 7,
COMMON_SAMPLER_TYPE_XTC = 8,
COMMON_SAMPLER_TYPE_INFILL = 9,
};
// dimensionality reduction methods, used by cvector-generator
@ -102,35 +105,46 @@ enum dimre_method {
struct common_sampler_params {
uint32_t seed = LLAMA_DEFAULT_SEED; // the seed used to initialize llama_sampler
int32_t n_prev = 64; // number of previous tokens to remember
int32_t n_probs = 0; // if greater than 0, output the probabilities of top n_probs tokens.
int32_t min_keep = 0; // 0 = disabled, otherwise samplers should return at least min_keep tokens
int32_t top_k = 40; // <= 0 to use vocab size
float top_p = 0.95f; // 1.0 = disabled
float min_p = 0.05f; // 0.0 = disabled
float tfs_z = 1.00f; // 1.0 = disabled
float typ_p = 1.00f; // typical_p, 1.0 = disabled
float temp = 0.80f; // <= 0.0 to sample greedily, 0.0 to not output probabilities
float dynatemp_range = 0.00f; // 0.0 = disabled
float dynatemp_exponent = 1.00f; // controls how entropy maps to temperature in dynamic temperature sampler
int32_t penalty_last_n = 64; // last n tokens to penalize (0 = disable penalty, -1 = context size)
float penalty_repeat = 1.00f; // 1.0 = disabled
float penalty_freq = 0.00f; // 0.0 = disabled
float penalty_present = 0.00f; // 0.0 = disabled
int32_t mirostat = 0; // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
float mirostat_tau = 5.00f; // target entropy
float mirostat_eta = 0.10f; // learning rate
bool penalize_nl = false; // consider newlines as a repeatable token
bool ignore_eos = false;
bool no_perf = false; // disable performance metrics
int32_t n_prev = 64; // number of previous tokens to remember
int32_t n_probs = 0; // if greater than 0, output the probabilities of top n_probs tokens.
int32_t min_keep = 0; // 0 = disabled, otherwise samplers should return at least min_keep tokens
int32_t top_k = 40; // <= 0 to use vocab size
float top_p = 0.95f; // 1.0 = disabled
float min_p = 0.05f; // 0.0 = disabled
float xtc_probability = 0.00f; // 0.0 = disabled
float xtc_threshold = 0.10f; // > 0.5 disables XTC
float tfs_z = 1.00f; // 1.0 = disabled
float typ_p = 1.00f; // typical_p, 1.0 = disabled
float temp = 0.80f; // <= 0.0 to sample greedily, 0.0 to not output probabilities
float dynatemp_range = 0.00f; // 0.0 = disabled
float dynatemp_exponent = 1.00f; // controls how entropy maps to temperature in dynamic temperature sampler
int32_t penalty_last_n = 64; // last n tokens to penalize (0 = disable penalty, -1 = context size)
float penalty_repeat = 1.00f; // 1.0 = disabled
float penalty_freq = 0.00f; // 0.0 = disabled
float penalty_present = 0.00f; // 0.0 = disabled
float dry_multiplier = 0.0f; // 0.0 = disabled; DRY repetition penalty for tokens extending repetition:
float dry_base = 1.75f; // 0.0 = disabled; multiplier * base ^ (length of sequence before token - allowed length)
int32_t dry_allowed_length = 2; // tokens extending repetitions beyond this receive penalty
int32_t dry_penalty_last_n = -1; // how many tokens to scan for repetitions (0 = disable penalty, -1 = context size)
int32_t mirostat = 0; // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
float mirostat_tau = 5.00f; // target entropy
float mirostat_eta = 0.10f; // learning rate
bool penalize_nl = false; // consider newlines as a repeatable token
bool ignore_eos = false;
bool no_perf = false; // disable performance metrics
std::vector<std::string> dry_sequence_breakers = {"\n", ":", "\"", "*"}; // default sequence breakers for DRY
std::vector<enum common_sampler_type> samplers = {
COMMON_SAMPLER_TYPE_DRY,
COMMON_SAMPLER_TYPE_TOP_K,
COMMON_SAMPLER_TYPE_TFS_Z,
COMMON_SAMPLER_TYPE_TYPICAL_P,
COMMON_SAMPLER_TYPE_TOP_P,
COMMON_SAMPLER_TYPE_MIN_P,
COMMON_SAMPLER_TYPE_TEMPERATURE
COMMON_SAMPLER_TYPE_XTC,
COMMON_SAMPLER_TYPE_TEMPERATURE,
};
std::string grammar; // optional BNF-like grammar to constrain sampling
@ -268,21 +282,21 @@ struct common_params {
// embedding
bool embedding = false; // get only sentence embedding
int32_t embd_normalize = 2; // normalisation for embendings (-1=none, 0=max absolute int16, 1=taxicab, 2=euclidean, >2=p-norm)
int32_t embd_normalize = 2; // normalisation for embeddings (-1=none, 0=max absolute int16, 1=taxicab, 2=euclidean, >2=p-norm)
std::string embd_out = ""; // empty = default, "array" = [[],[]...], "json" = openai style, "json+" = same "json" + cosine similarity matrix
std::string embd_sep = "\n"; // separator of embendings
std::string embd_sep = "\n"; // separator of embeddings
bool reranking = false; // enable reranking support on server
// server params
int32_t port = 8080; // server listens on this network port
int32_t timeout_read = 600; // http read timeout in seconds
int32_t timeout_write = timeout_read; // http write timeout in seconds
int n_threads_http = -1; // number of threads to process HTTP requests (TODO: support threadpool)
int32_t n_threads_http = -1; // number of threads to process HTTP requests (TODO: support threadpool)
int32_t n_cache_reuse = 0; // min chunk size to reuse from the cache via KV shifting
std::string hostname = "127.0.0.1";
std::string public_path = ""; // NOLINT
std::string chat_template = ""; // NOLINT
std::string system_prompt = ""; // NOLINT
bool enable_chat_template = true;
std::vector<std::string> api_keys;
@ -352,16 +366,27 @@ void common_init();
std::string common_params_get_system_info(const common_params & params);
bool parse_cpu_range(const std::string& range, bool(&boolmask)[GGML_MAX_N_THREADS]);
bool parse_cpu_mask(const std::string& mask, bool(&boolmask)[GGML_MAX_N_THREADS]);
void postprocess_cpu_params(cpu_params& cpuparams, const cpu_params* role_model = nullptr);
bool parse_cpu_range(const std::string & range, bool(&boolmask)[GGML_MAX_N_THREADS]);
bool parse_cpu_mask(const std::string & mask, bool(&boolmask)[GGML_MAX_N_THREADS]);
void postprocess_cpu_params(cpu_params & cpuparams, const cpu_params * role_model = nullptr);
bool set_process_priority(enum ggml_sched_priority prio);
//
// String utils
//
std::vector<std::string> string_split(std::string input, char separator);
#ifdef __GNUC__
#ifdef __MINGW32__
#define LLAMA_COMMON_ATTRIBUTE_FORMAT(...) __attribute__((format(gnu_printf, __VA_ARGS__)))
#else
#define LLAMA_COMMON_ATTRIBUTE_FORMAT(...) __attribute__((format(printf, __VA_ARGS__)))
#endif
#else
#define LLAMA_COMMON_ATTRIBUTE_FORMAT(...)
#endif
LLAMA_COMMON_ATTRIBUTE_FORMAT(1, 2)
std::string string_format(const char * fmt, ...);
std::string string_strip(const std::string & str);
std::string string_get_sortable_timestamp();
@ -370,6 +395,7 @@ void string_replace_all(std::string & s, const std::string & search, const std::
template<class T>
static std::vector<T> string_split(const std::string & str, char delim) {
static_assert(!std::is_same<T, std::string>::value, "Please use the specialized version for std::string");
std::vector<T> values;
std::istringstream str_stream(str);
std::string token;
@ -382,6 +408,22 @@ static std::vector<T> string_split(const std::string & str, char delim) {
return values;
}
template<>
std::vector<std::string> string_split<std::string>(const std::string & input, char separator)
{
std::vector<std::string> parts;
size_t begin_pos = 0;
size_t separator_pos = input.find(separator);
while (separator_pos != std::string::npos) {
std::string part = input.substr(begin_pos, separator_pos - begin_pos);
parts.emplace_back(part);
begin_pos = separator_pos + 1;
separator_pos = input.find(separator, begin_pos);
}
parts.emplace_back(input.substr(begin_pos, separator_pos - begin_pos));
return parts;
}
bool string_parse_kv_override(const char * data, std::vector<llama_model_kv_override> & overrides);
void string_process_escapes(std::string & input);

2
common/json-schema-to-grammar.cpp
View File

@ -611,7 +611,7 @@ private:
}
return join_seq();
};
return _add_rule(name, "\"\\\"\" " + to_rule(transform()) + " \"\\\"\" space");
return _add_rule(name, "\"\\\"\" (" + to_rule(transform()) + ") \"\\\"\" space");
}
/*

111
common/sampling.cpp
View File

@ -130,10 +130,12 @@ std::string common_sampler_params::print() const {
snprintf(result, sizeof(result),
"\trepeat_last_n = %d, repeat_penalty = %.3f, frequency_penalty = %.3f, presence_penalty = %.3f\n"
"\ttop_k = %d, tfs_z = %.3f, top_p = %.3f, min_p = %.3f, typical_p = %.3f, temp = %.3f\n"
"\tdry_multiplier = %.3f, dry_base = %.3f, dry_allowed_length = %d, dry_penalty_last_n = %d\n"
"\ttop_k = %d, tfs_z = %.3f, top_p = %.3f, min_p = %.3f, xtc_probability = %.3f, xtc_threshold = %.3f, typical_p = %.3f, temp = %.3f\n"
"\tmirostat = %d, mirostat_lr = %.3f, mirostat_ent = %.3f",
penalty_last_n, penalty_repeat, penalty_freq, penalty_present,
top_k, tfs_z, top_p, min_p, typ_p, temp,
dry_multiplier, dry_base, dry_allowed_length, dry_penalty_last_n,
top_k, tfs_z, top_p, min_p, xtc_probability, xtc_threshold, typ_p, temp,
mirostat, mirostat_eta, mirostat_tau);
return std::string(result);
@ -171,54 +173,57 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
params.penalize_nl,
params.ignore_eos));
if (params.temp > 0.0f) {
if (params.mirostat == 0) {
for (const auto & cnstr : params.samplers) {
switch (cnstr) {
case COMMON_SAMPLER_TYPE_TOP_K:
llama_sampler_chain_add(result->chain, llama_sampler_init_top_k (params.top_k));
if (params.mirostat == 0) {
for (const auto & cnstr : params.samplers) {
switch (cnstr) {
case COMMON_SAMPLER_TYPE_DRY:
{
std::vector<const char*> c_breakers;
c_breakers.reserve(params.dry_sequence_breakers.size());
for (const auto& str : params.dry_sequence_breakers) {
c_breakers.push_back(str.c_str());
}
llama_sampler_chain_add(result->chain, llama_sampler_init_dry (model, params.dry_multiplier, params.dry_base, params.dry_allowed_length, params.dry_penalty_last_n, c_breakers.data(), c_breakers.size()));
}
break;
case COMMON_SAMPLER_TYPE_TOP_P:
llama_sampler_chain_add(result->chain, llama_sampler_init_top_p (params.top_p, params.min_keep));
break;
case COMMON_SAMPLER_TYPE_MIN_P:
llama_sampler_chain_add(result->chain, llama_sampler_init_min_p (params.min_p, params.min_keep));
break;
case COMMON_SAMPLER_TYPE_TFS_Z:
llama_sampler_chain_add(result->chain, llama_sampler_init_tail_free(params.tfs_z, params.min_keep));
break;
case COMMON_SAMPLER_TYPE_TYPICAL_P:
llama_sampler_chain_add(result->chain, llama_sampler_init_typical (params.typ_p, params.min_keep));
break;
case COMMON_SAMPLER_TYPE_TEMPERATURE:
llama_sampler_chain_add(result->chain, llama_sampler_init_temp_ext (params.temp, params.dynatemp_range, params.dynatemp_exponent));
break;
default:
GGML_ASSERT(false && "unknown sampler type");
}
case COMMON_SAMPLER_TYPE_TOP_K:
llama_sampler_chain_add(result->chain, llama_sampler_init_top_k (params.top_k));
break;
case COMMON_SAMPLER_TYPE_TOP_P:
llama_sampler_chain_add(result->chain, llama_sampler_init_top_p (params.top_p, params.min_keep));
break;
case COMMON_SAMPLER_TYPE_MIN_P:
llama_sampler_chain_add(result->chain, llama_sampler_init_min_p (params.min_p, params.min_keep));
break;
case COMMON_SAMPLER_TYPE_XTC:
llama_sampler_chain_add(result->chain, llama_sampler_init_xtc (params.xtc_probability, params.xtc_threshold, params.min_keep, params.seed));
break;
case COMMON_SAMPLER_TYPE_TFS_Z:
llama_sampler_chain_add(result->chain, llama_sampler_init_tail_free(params.tfs_z, params.min_keep));
break;
case COMMON_SAMPLER_TYPE_TYPICAL_P:
llama_sampler_chain_add(result->chain, llama_sampler_init_typical (params.typ_p, params.min_keep));
break;
case COMMON_SAMPLER_TYPE_TEMPERATURE:
llama_sampler_chain_add(result->chain, llama_sampler_init_temp_ext (params.temp, params.dynatemp_range, params.dynatemp_exponent));
break;
case COMMON_SAMPLER_TYPE_INFILL:
llama_sampler_chain_add(result->chain, llama_sampler_init_infill (model));
break;
default:
GGML_ASSERT(false && "unknown sampler type");
}
llama_sampler_chain_add(result->chain, llama_sampler_init_softmax());
llama_sampler_chain_add(result->chain, llama_sampler_init_dist(params.seed));
} else if (params.mirostat == 1) {
llama_sampler_chain_add(result->chain, llama_sampler_init_temp(params.temp));
llama_sampler_chain_add(result->chain, llama_sampler_init_mirostat(llama_n_vocab(model), params.seed, params.mirostat_tau, params.mirostat_eta, 100));
} else if (params.mirostat == 2) {
llama_sampler_chain_add(result->chain, llama_sampler_init_temp(params.temp));
llama_sampler_chain_add(result->chain, llama_sampler_init_mirostat_v2(params.seed, params.mirostat_tau, params.mirostat_eta));
} else {
GGML_ASSERT(false && "unknown mirostat version");
}
llama_sampler_chain_add(result->chain, llama_sampler_init_dist(params.seed));
} else if (params.mirostat == 1) {
llama_sampler_chain_add(result->chain, llama_sampler_init_temp(params.temp));
llama_sampler_chain_add(result->chain, llama_sampler_init_mirostat(llama_n_vocab(model), params.seed, params.mirostat_tau, params.mirostat_eta, 100));
} else if (params.mirostat == 2) {
llama_sampler_chain_add(result->chain, llama_sampler_init_temp(params.temp));
llama_sampler_chain_add(result->chain, llama_sampler_init_mirostat_v2(params.seed, params.mirostat_tau, params.mirostat_eta));
} else {
if (params.n_probs > 0) {
// some use cases require to sample greedily, but still obtain the probabilities of the top tokens
// ref: https://github.com/ggerganov/llama.cpp/pull/9605
//
// the following will not produce exactly the same probs as applyging softmax to the full vocabulary, but
// it is much faster, since we avoid sorting all tokens and should give a good approximation
llama_sampler_chain_add(result->chain, llama_sampler_init_top_k(params.n_probs));
llama_sampler_chain_add(result->chain, llama_sampler_init_softmax());
}
llama_sampler_chain_add(result->chain, llama_sampler_init_greedy());
GGML_ASSERT(false && "unknown mirostat version");
}
return result;
@ -366,36 +371,45 @@ std::string common_sampler_prev_str(common_sampler * gsmpl, llama_context * ctx_
char common_sampler_type_to_chr(enum common_sampler_type cnstr) {
switch (cnstr) {
case COMMON_SAMPLER_TYPE_DRY: return 'd';
case COMMON_SAMPLER_TYPE_TOP_K: return 'k';
case COMMON_SAMPLER_TYPE_TFS_Z: return 'f';
case COMMON_SAMPLER_TYPE_TYPICAL_P: return 'y';
case COMMON_SAMPLER_TYPE_TOP_P: return 'p';
case COMMON_SAMPLER_TYPE_MIN_P: return 'm';
case COMMON_SAMPLER_TYPE_TEMPERATURE: return 't';
case COMMON_SAMPLER_TYPE_XTC: return 'x';
case COMMON_SAMPLER_TYPE_INFILL: return 'i';
default : return '?';
}
}
std::string common_sampler_type_to_str(enum common_sampler_type cnstr) {
switch (cnstr) {
case COMMON_SAMPLER_TYPE_DRY: return "dry";
case COMMON_SAMPLER_TYPE_TOP_K: return "top_k";
case COMMON_SAMPLER_TYPE_TFS_Z: return "tfs_z";
case COMMON_SAMPLER_TYPE_TYPICAL_P: return "typ_p";
case COMMON_SAMPLER_TYPE_TOP_P: return "top_p";
case COMMON_SAMPLER_TYPE_MIN_P: return "min_p";
case COMMON_SAMPLER_TYPE_TEMPERATURE: return "temperature";
case COMMON_SAMPLER_TYPE_XTC: return "xtc";
case COMMON_SAMPLER_TYPE_INFILL: return "infill";
default : return "";
}
}
std::vector<common_sampler_type> common_sampler_types_from_names(const std::vector<std::string> & names, bool allow_alt_names) {
std::unordered_map<std::string, common_sampler_type> sampler_canonical_name_map {
{ "dry", COMMON_SAMPLER_TYPE_DRY },
{ "top_k", COMMON_SAMPLER_TYPE_TOP_K },
{ "top_p", COMMON_SAMPLER_TYPE_TOP_P },
{ "typ_p", COMMON_SAMPLER_TYPE_TYPICAL_P },
{ "min_p", COMMON_SAMPLER_TYPE_MIN_P },
{ "tfs_z", COMMON_SAMPLER_TYPE_TFS_Z },
{ "temperature", COMMON_SAMPLER_TYPE_TEMPERATURE },
{ "xtc", COMMON_SAMPLER_TYPE_XTC },
{ "infill", COMMON_SAMPLER_TYPE_INFILL },
};
// since samplers names are written multiple ways
@ -436,12 +450,15 @@ std::vector<common_sampler_type> common_sampler_types_from_names(const std::vect
std::vector<common_sampler_type> common_sampler_types_from_chars(const std::string & chars) {
std::unordered_map<char, common_sampler_type> sampler_name_map = {
{ common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_DRY), COMMON_SAMPLER_TYPE_DRY },
{ common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TOP_K), COMMON_SAMPLER_TYPE_TOP_K },
{ common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TFS_Z), COMMON_SAMPLER_TYPE_TFS_Z },
{ common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TYPICAL_P), COMMON_SAMPLER_TYPE_TYPICAL_P },
{ common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TOP_P), COMMON_SAMPLER_TYPE_TOP_P },
{ common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_MIN_P), COMMON_SAMPLER_TYPE_MIN_P },
{ common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TEMPERATURE), COMMON_SAMPLER_TYPE_TEMPERATURE }
{ common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TEMPERATURE), COMMON_SAMPLER_TYPE_TEMPERATURE },
{ common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_XTC), COMMON_SAMPLER_TYPE_XTC },
{ common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_INFILL), COMMON_SAMPLER_TYPE_INFILL },
};
std::vector<common_sampler_type> samplers;

6
convert_hf_to_gguf.py
View File

@ -573,6 +573,9 @@ class Model:
if chkhsh == "0876d13b50744004aa9aeae05e7b0647eac9d801b5ba4668afc01e709c15e19f":
# ref: https://huggingface.co/BAAI/bge-small-en-v1.5
res = "bert-bge"
if chkhsh == "8e62295832751ca1e8f92f2226f403dea30dc5165e448b5bfa05af5340c64ec7":
# ref: https://huggingface.co/BAAI/bge-large-zh-v1.5
res = "bert-bge-large"
if chkhsh == "b6dc8df998e1cfbdc4eac8243701a65afe638679230920b50d6f17d81c098166":
# ref: https://huggingface.co/mosaicml/mpt-7b
res = "mpt"
@ -2864,6 +2867,9 @@ class Rwkv6Model(Model):
self.gguf_writer.add_token_list(tokens)
self.gguf_writer.add_token_types(toktypes)
special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=False)
special_vocab.chat_template = "rwkv-world"
# hack: Add '\n\n' as the EOT token to make it chat normally
special_vocab._set_special_token("eot", 261)
special_vocab.add_to_gguf(self.gguf_writer)
def set_gguf_parameters(self):

1
convert_hf_to_gguf_update.py
View File

@ -72,6 +72,7 @@ models = [
{"name": "deepseek-coder", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/deepseek-ai/deepseek-coder-6.7b-base", },
{"name": "falcon", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tiiuae/falcon-7b", },
{"name": "bert-bge", "tokt": TOKENIZER_TYPE.WPM, "repo": "https://huggingface.co/BAAI/bge-small-en-v1.5", },
{"name": "bert-bge-large", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/BAAI/bge-large-zh-v1.5", },
{"name": "mpt", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/mosaicml/mpt-7b", },
{"name": "starcoder", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/bigcode/starcoder2-3b", },
{"name": "gpt-2", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/openai-community/gpt2", },

3
convert_lora_to_gguf.py
View File

@ -348,6 +348,9 @@ if __name__ == '__main__':
if ".base_layer.weight" in name:
continue
logger.error(f"Unexpected name '{name}': Not a lora_A or lora_B tensor")
if ".embed_tokens.weight" in name or ".lm_head.weight" in name:
logger.error("Embeddings is present in the adapter. This can be due to new tokens added during fine tuning")
logger.error("Hint: if you are using TRL, make sure not to call setup_chat_format()")
sys.exit(1)
if base_name in tensor_map:

8
docs/build.md
View File

@ -198,6 +198,8 @@ The following compilation options are also available to tweak performance:
### MUSA
This provides GPU acceleration using the MUSA cores of your Moore Threads MTT GPU. Make sure to have the MUSA SDK installed. You can download it from here: [MUSA SDK](https://developer.mthreads.com/sdk/download/musa).
- Using `make`:
```bash
make GGML_MUSA=1
@ -209,6 +211,12 @@ The following compilation options are also available to tweak performance:
cmake --build build --config Release
```
The environment variable [`MUSA_VISIBLE_DEVICES`](https://docs.mthreads.com/musa-sdk/musa-sdk-doc-online/programming_guide/Z%E9%99%84%E5%BD%95/) can be used to specify which GPU(s) will be used.
The environment variable `GGML_CUDA_ENABLE_UNIFIED_MEMORY=1` can be used to enable unified memory in Linux. This allows swapping to system RAM instead of crashing when the GPU VRAM is exhausted.
Most of the compilation options available for CUDA should also be available for MUSA, though they haven't been thoroughly tested yet.
### hipBLAS
This provides BLAS acceleration on HIP-supported AMD GPUs.

1
examples/batched-bench/batched-bench.cpp
View File

@ -74,7 +74,6 @@ int main(int argc, char ** argv) {
batch.n_seq_id + i,
batch.seq_id + i,
batch.logits + i,
0, 0, 0, // unused
};
const int ret = llama_decode(ctx, batch_view);

2
examples/cvector-generator/cvector-generator.cpp
View File

@ -339,7 +339,7 @@ static bool cb_eval(struct ggml_tensor * t, bool ask, void * user_data) {
static bool get_hidden_layers(llama_context * ctx, std::vector<llama_token> & tokens) {
llama_kv_cache_clear(ctx);
if (llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size(), 0, 0))) {
if (llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size()))) {
fprintf(stderr, "%s : failed to eval\n", __func__);
return false;
}

2
examples/eval-callback/eval-callback.cpp
View File

@ -131,7 +131,7 @@ static bool run(llama_context * ctx, const common_params & params) {
std::vector<llama_token> tokens = common_tokenize(ctx, params.prompt, add_bos);
if (llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size(), 0, 0))) {
if (llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size()))) {
LOG_ERR("%s : failed to eval\n", __func__);
return false;
}

13
examples/imatrix/imatrix.cpp
View File

@ -496,6 +496,8 @@ static bool compute_imatrix(llama_context * ctx, const common_params & params) {
// clear the KV cache
llama_kv_cache_clear(ctx);
llama_batch batch = llama_batch_init(n_batch, 0, 1);
for (int j = 0; j < num_batches; ++j) {
const int batch_start = start + j * n_batch;
const int batch_size = std::min(end - batch_start, n_batch);
@ -508,9 +510,14 @@ static bool compute_imatrix(llama_context * ctx, const common_params & params) {
tokens[batch_start] = llama_token_bos(llama_get_model(ctx));
}
// TODO: use batch.logits to save computations instead of relying on logits_all == true
if (llama_decode(ctx, llama_batch_get_one(tokens.data() + batch_start, batch_size, j * n_batch, 0))) {
common_batch_clear(batch);
for (int i = 0; i < batch_size; i++) {
common_batch_add(batch, tokens[batch_start + i], j*n_batch + i, {0}, true);
}
if (llama_decode(ctx, batch)) {
LOG_ERR("%s : failed to eval\n", __func__);
llama_batch_free(batch);
return false;
}
@ -523,6 +530,8 @@ static bool compute_imatrix(llama_context * ctx, const common_params & params) {
}
}
llama_batch_free(batch);
const auto t_end = std::chrono::high_resolution_clock::now();
if (i == 0) {

16
examples/infill/infill.cpp
View File

@ -205,11 +205,11 @@ int main(int argc, char ** argv) {
std::vector<llama_token> inp_pfx = common_tokenize(ctx, params.input_prefix, false);
std::vector<llama_token> inp_sfx = common_tokenize(ctx, params.input_suffix, false);
GGML_ASSERT(llama_token_prefix(model) >= 0);
GGML_ASSERT(llama_token_suffix(model) >= 0);
GGML_ASSERT(llama_token_fim_pre(model) >= 0);
GGML_ASSERT(llama_token_fim_suf(model) >= 0);
inp_pfx.insert(inp_pfx.begin(), llama_token_prefix(model));
inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(model));
inp_pfx.insert(inp_pfx.begin(), llama_token_fim_pre(model));
inp_sfx.insert(inp_sfx.begin(), llama_token_fim_suf(model));
embd_inp = params.spm_infill ? inp_sfx : inp_pfx;
embd_end = params.spm_infill ? inp_pfx : inp_sfx;
@ -218,7 +218,7 @@ int main(int argc, char ** argv) {
}
embd_inp.insert(embd_inp.end(), embd_end.begin(), embd_end.end());
const llama_token middle_token = llama_token_middle(model);
const llama_token middle_token = llama_token_fim_mid(model);
if (middle_token >= 0) {
embd_inp.push_back(middle_token);
}
@ -396,7 +396,7 @@ int main(int argc, char ** argv) {
LOG_DBG("eval: %s\n", string_from(ctx, embd).c_str());
if (llama_decode(ctx, llama_batch_get_one(&embd[i], n_eval, n_past, 0))) {
if (llama_decode(ctx, llama_batch_get_one(&embd[i], n_eval))) {
LOG_ERR("%s : failed to eval\n", __func__);
return 1;
}
@ -508,8 +508,8 @@ int main(int argc, char ** argv) {
std::vector<llama_token> inp_pfx = common_tokenize(ctx, params.input_prefix, false);
std::vector<llama_token> inp_sfx = common_tokenize(ctx, params.input_suffix, false);
inp_pfx.insert(inp_pfx.begin(), llama_token_prefix(model));
inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(model));
inp_pfx.insert(inp_pfx.begin(), llama_token_fim_pre(model));
inp_sfx.insert(inp_sfx.begin(), llama_token_fim_suf(model));
embd_inp = params.spm_infill ? inp_sfx : inp_pfx;
embd_end = params.spm_infill ? inp_pfx : inp_sfx;

2
examples/json_schema_to_grammar.py
View File

@ -540,7 +540,7 @@ class SchemaConverter:
return self._add_rule(
name,
to_rule(transform()) if self._raw_pattern \
else "\"\\\"\" " + to_rule(transform()) + " \"\\\"\" space")
else "\"\\\"\" (" + to_rule(transform()) + ") \"\\\"\" space")
def _resolve_ref(self, ref):

18
examples/llama-bench/llama-bench.cpp
View File

@ -151,7 +151,7 @@ static std::string get_gpu_info() {
int count = ggml_backend_sycl_get_device_count();
for (int i = 0; i < count; i++) {
char buf[128];
ggml_sycl_get_device_description(i, buf, sizeof(buf));
ggml_backend_sycl_get_device_description(i, buf, sizeof(buf));
id += buf;
if (i < count - 1) {
id += "/";
@ -1428,7 +1428,7 @@ struct sql_printer : public printer {
}
};
static void test_prompt(llama_context * ctx, int n_prompt, int n_past, int n_batch, int n_threads) {
static void test_prompt(llama_context * ctx, int n_prompt, int n_batch, int n_threads) {
llama_set_n_threads(ctx, n_threads, n_threads);
const llama_model * model = llama_get_model(ctx);
@ -1444,14 +1444,14 @@ static void test_prompt(llama_context * ctx, int n_prompt, int n_past, int n_bat
for (int i = 1; i < n_tokens; i++) {
tokens[i] = std::rand() % n_vocab;
}
llama_decode(ctx, llama_batch_get_one(tokens.data(), n_tokens, n_past + n_processed, 0));
llama_decode(ctx, llama_batch_get_one(tokens.data(), n_tokens));
n_processed += n_tokens;
}
llama_synchronize(ctx);
}
static void test_gen(llama_context * ctx, int n_gen, int n_past, int n_threads) {
static void test_gen(llama_context * ctx, int n_gen, int n_threads) {
llama_set_n_threads(ctx, n_threads, n_threads);
const llama_model * model = llama_get_model(ctx);
@ -1460,7 +1460,7 @@ static void test_gen(llama_context * ctx, int n_gen, int n_past, int n_threads)
llama_token token = llama_add_bos_token(model) ? llama_token_bos(model) : std::rand() % n_vocab;
for (int i = 0; i < n_gen; i++) {
llama_decode(ctx, llama_batch_get_one(&token, 1, n_past + i, 0));
llama_decode(ctx, llama_batch_get_one(&token, 1));
llama_synchronize(ctx);
token = std::rand() % n_vocab;
}
@ -1596,13 +1596,13 @@ int main(int argc, char ** argv) {
fprintf(stderr, "llama-bench: benchmark %d/%ld: warmup prompt run\n", params_idx, params_count);
}
//test_prompt(ctx, std::min(t.n_batch, std::min(t.n_prompt, 32)), 0, t.n_batch, t.n_threads);
test_prompt(ctx, t.n_prompt, 0, t.n_batch, t.n_threads);
test_prompt(ctx, t.n_prompt, t.n_batch, t.n_threads);
}
if (t.n_gen > 0) {
if (params.progress) {
fprintf(stderr, "llama-bench: benchmark %d/%ld: warmup generation run\n", params_idx, params_count);
}
test_gen(ctx, 1, 0, t.n_threads);
test_gen(ctx, 1, t.n_threads);
}
for (int i = 0; i < params.reps; i++) {
@ -1614,13 +1614,13 @@ int main(int argc, char ** argv) {
if (params.progress) {
fprintf(stderr, "llama-bench: benchmark %d/%ld: prompt run %d/%d\n", params_idx, params_count, i + 1, params.reps);
}
test_prompt(ctx, t.n_prompt, 0, t.n_batch, t.n_threads);
test_prompt(ctx, t.n_prompt, t.n_batch, t.n_threads);
}
if (t.n_gen > 0) {
if (params.progress) {
fprintf(stderr, "llama-bench: benchmark %d/%ld: generation run %d/%d\n", params_idx, params_count, i + 1, params.reps);
}
test_gen(ctx, t.n_gen, t.n_prompt, t.n_threads);
test_gen(ctx, t.n_gen, t.n_threads);
}
uint64_t t_ns = get_time_ns() - t_start;

3
examples/llama.android/llama/src/main/cpp/llama-android.cpp
View File

@ -283,9 +283,6 @@ Java_android_llama_cpp_LLamaAndroid_new_1batch(JNIEnv *, jobject, jint n_tokens,
nullptr,
nullptr,
nullptr,
0,
0,
0,
};
if (embd) {

1
examples/llama.swiftui/llama.cpp.swift/LibLlama.swift
View File

@ -46,7 +46,6 @@ actor LlamaContext {
let sparams = llama_sampler_chain_default_params()
self.sampling = llama_sampler_chain_init(sparams)
llama_sampler_chain_add(self.sampling, llama_sampler_init_temp(0.4))
llama_sampler_chain_add(self.sampling, llama_sampler_init_softmax())
llama_sampler_chain_add(self.sampling, llama_sampler_init_dist(1234))
}

783
examples/llama.vim Normal file
View File

@ -0,0 +1,783 @@
" LLM-based text completion using llama.cpp
"
" requires:
"
" - neovim or vim
" - curl
" - llama.cpp server instance
" - FIM-compatible model
"
" sample config:
"
" - Tab - accept the current suggestion
" - Shift+Tab - accept just the first line of the suggestion
" - Ctrl+F - toggle FIM completion manually
"
" make symlink or copy this file to ~/.config/nvim/autoload/llama.vim
"
" start the llama.cpp server with a FIM-compatible model. for example:
"
" $ llama-server -m {model.gguf} --port 8012 -ngl 99 -fa -dt 0.1 --ubatch-size 512 --batch-size 1024 --cache-reuse 256
"
" --batch-size [512, model max context]
"
" adjust the batch size to control how much of the provided local context will be used during the inference
" lower values will use smaller part of the context around the cursor, which will result in faster processing
"
" --ubatch-size [64, 2048]
"
" chunks the batch into smaller chunks for faster processing
" depends on the specific hardware. use llama-bench to profile and determine the best size
"
" --cache-reuse (ge:llama_config.n_predict, 1024]
"
" this should be either 0 (disabled) or strictly larger than g:llama_config.n_predict
" using non-zero value enables context reuse on the server side which dramatically improves the performance at
" large contexts. a value of 256 should be good for all cases
"
" run this once to initialise llama.vim:
"
" :call llama#init()
"
" more info: https://github.com/ggerganov/llama.cpp/pull/9787
"
" colors (adjust to your liking)
highlight llama_hl_hint guifg=#ff772f ctermfg=202
highlight llama_hl_info guifg=#77ff2f ctermfg=119
" general parameters:
"
" endpoint: llama.cpp server endpoint
" n_prefix: number of lines before the cursor location to include in the local prefix
" n_suffix: number of lines after the cursor location to include in the local suffix
" n_predict: max number of tokens to predict
" t_max_prompt_ms: max alloted time for the prompt processing (TODO: not yet supported)
" t_max_predict_ms: max alloted time for the prediction
" show_info: show extra info about the inference (0 - disabled, 1 - statusline, 2 - inline)
" auto_fim: trigger FIM completion automatically on cursor movement
" max_line_suffix: do not auto-trigger FIM completion if there are more than this number of characters to the right of the cursor
"
" ring buffer of chunks, accumulated with time upon:
"
" - completion request
" - yank
" - entering a buffer
" - leaving a buffer
" - writing a file
"
" parameters for the ring-buffer with extra context:
"
" ring_n_chunks: max number of chunks to pass as extra context to the server (0 to disable)
" ring_chunk_size: max size of the chunks (in number of lines)
" note: adjust these numbers so that you don't overrun your context
" at ring_n_chunks = 64 and ring_chunk_size = 64 you need ~32k context
" ring_scope: the range around the cursor position (in number of lines) for gathering chunks after FIM
" ring_update_ms: how often to process queued chunks in normal mode
"
let s:default_config = {
\ 'endpoint': 'http://127.0.0.1:8012/infill',
\ 'n_prefix': 256,
\ 'n_suffix': 64,
\ 'n_predict': 128,
\ 't_max_prompt_ms': 500,
\ 't_max_predict_ms': 3000,
\ 'show_info': 2,
\ 'auto_fim': v:true,
\ 'max_line_suffix': 8,
\ 'ring_n_chunks': 64,
\ 'ring_chunk_size': 64,
\ 'ring_scope': 1024,
\ 'ring_update_ms': 1000,
\ }
let g:llama_config = get(g:, 'llama_config', s:default_config)
function! s:get_indent(str)
let l:count = 0
for i in range(len(a:str))
if a:str[i] == "\t"
let l:count += &tabstop - 1
else
break
endif
endfor
return l:count
endfunction
function! s:rand(i0, i1) abort
return a:i0 + rand() % (a:i1 - a:i0 + 1)
endfunction
function! llama#init()
if !executable('curl')
echohl WarningMsg
echo 'llama.vim requires the "curl" command to be available'
echohl None
return
endif
let s:pos_x = 0 " cursor position upon start of completion
let s:pos_y = 0
let s:line_cur = ''
let s:line_cur_prefix = ''
let s:line_cur_suffix = ''
let s:ring_chunks = [] " current set of chunks used as extra context
let s:ring_queued = [] " chunks that are queued to be sent for processing
let s:ring_n_evict = 0
let s:hint_shown = v:false
let s:pos_y_pick = -9999 " last y where we picked a chunk
let s:pos_dx = 0
let s:content = []
let s:can_accept = v:false
let s:timer_fim = -1
let s:t_fim_start = reltime() " used to measure total FIM time
let s:t_last_move = reltime() " last time the cursor moved
let s:current_job = v:null
let s:ghost_text_nvim = exists('*nvim_buf_get_mark')
let s:ghost_text_vim = has('textprop')
if s:ghost_text_vim
let s:hlgroup_hint = 'llama_hl_hint'
let s:hlgroup_info = 'llama_hl_info'
if empty(prop_type_get(s:hlgroup_hint))
call prop_type_add(s:hlgroup_hint, {'highlight': s:hlgroup_hint})
endif
if empty(prop_type_get(s:hlgroup_info))
call prop_type_add(s:hlgroup_info, {'highlight': s:hlgroup_info})
endif
endif
augroup llama
autocmd!
autocmd InsertEnter * inoremap <expr> <silent> <C-F> llama#fim_inline(v:false)
autocmd InsertLeavePre * call llama#fim_cancel()
autocmd CursorMoved * call s:on_move()
autocmd CursorMovedI * call s:on_move()
autocmd CompleteChanged * call llama#fim_cancel()
if g:llama_config.auto_fim
autocmd CursorMovedI * call llama#fim(v:true)
endif
" gather chunks upon yanking
autocmd TextYankPost * if v:event.operator ==# 'y' | call s:pick_chunk(v:event.regcontents, v:false, v:true) | endif
" gather chunks upon entering/leaving a buffer
autocmd BufEnter * call timer_start(100, {-> s:pick_chunk(getline(max([1, line('.') - g:llama_config.ring_chunk_size/2]), min([line('.') + g:llama_config.ring_chunk_size/2, line('$')])), v:true, v:true)})
autocmd BufLeave * call s:pick_chunk(getline(max([1, line('.') - g:llama_config.ring_chunk_size/2]), min([line('.') + g:llama_config.ring_chunk_size/2, line('$')])), v:true, v:true)
" gather chunk upon saving the file
autocmd BufWritePost * call s:pick_chunk(getline(max([1, line('.') - g:llama_config.ring_chunk_size/2]), min([line('.') + g:llama_config.ring_chunk_size/2, line('$')])), v:true, v:true)
augroup END
silent! call llama#fim_cancel()
" init background update of the ring buffer
if g:llama_config.ring_n_chunks > 0
call s:ring_update()
endif
endfunction
" compute how similar two chunks of text are
" 0 - no similarity, 1 - high similarity
" TODO: figure out something better
function! s:chunk_sim(c0, c1)
let l:lines0 = len(a:c0)
let l:lines1 = len(a:c1)
let l:common = 0
for l:line0 in a:c0
for l:line1 in a:c1
if l:line0 == l:line1
let l:common += 1
break
endif
endfor
endfor
return 2.0 * l:common / (l:lines0 + l:lines1)
endfunction
" pick a random chunk of size g:llama_config.ring_chunk_size from the provided text and queue it for processing
"
" no_mod - do not pick chunks from buffers with pending changes
" do_evict - evict chunks that are very similar to the new one
"
function! s:pick_chunk(text, no_mod, do_evict)
" do not pick chunks from buffers with pending changes or buffers that are not files
if a:no_mod && (getbufvar(bufnr('%'), '&modified') || !buflisted(bufnr('%')) || !filereadable(expand('%')))
return
endif
" if the extra context option is disabled - do nothing
if g:llama_config.ring_n_chunks <= 0
return
endif
" don't pick very small chunks
if len(a:text) < 3
return
endif
if len(a:text) + 1 < g:llama_config.ring_chunk_size
let l:chunk = a:text
else
let l:l0 = s:rand(0, max([0, len(a:text) - g:llama_config.ring_chunk_size/2]))
let l:l1 = min([l:l0 + g:llama_config.ring_chunk_size/2, len(a:text)])
let l:chunk = a:text[l:l0:l:l1]
endif
let l:chunk_str = join(l:chunk, "\n") . "\n"
" check if this chunk is already added
let l:exist = v:false
for i in range(len(s:ring_chunks))
if s:ring_chunks[i].data == l:chunk
let l:exist = v:true
break
endif
endfor
for i in range(len(s:ring_queued))
if s:ring_queued[i].data == l:chunk
let l:exist = v:true
break
endif
endfor
if l:exist
return
endif
" evict queued chunks that are very similar to the new one
for i in range(len(s:ring_queued) - 1, 0, -1)
if s:chunk_sim(s:ring_queued[i].data, l:chunk) > 0.9
if a:do_evict
call remove(s:ring_queued, i)
let s:ring_n_evict += 1
else
return
endif
endif
endfor
" also from s:ring_chunks
for i in range(len(s:ring_chunks) - 1, 0, -1)
if s:chunk_sim(s:ring_chunks[i].data, l:chunk) > 0.9
if a:do_evict
call remove(s:ring_chunks, i)
let s:ring_n_evict += 1
else
return
endif
endif
endfor
" TODO: become parameter ?
if len(s:ring_queued) == 16
call remove(s:ring_queued, 0)
endif
call add(s:ring_queued, {'data': l:chunk, 'str': l:chunk_str, 'time': reltime(), 'filename': expand('%')})
"let &statusline = 'extra context: ' . len(s:ring_chunks) . ' / ' . len(s:ring_queued)
endfunction
" picks a queued chunk, sends it for processing and adds it to s:ring_chunks
" called every g:llama_config.ring_update_ms
function! s:ring_update()
call timer_start(g:llama_config.ring_update_ms, {-> s:ring_update()})
" update only if in normal mode or if the cursor hasn't moved for a while
if mode() !=# 'n' && reltimefloat(reltime(s:t_last_move)) < 3.0
return
endif
if len(s:ring_queued) == 0
return
endif
" move the first queued chunk to the ring buffer
if len(s:ring_chunks) == g:llama_config.ring_n_chunks
call remove(s:ring_chunks, 0)
endif
call add(s:ring_chunks, remove(s:ring_queued, 0))
"let &statusline = 'updated context: ' . len(s:ring_chunks) . ' / ' . len(s:ring_queued)
" send asynchronous job with the new extra context so that it is ready for the next FIM
let l:extra_context = []
for l:chunk in s:ring_chunks
call add(l:extra_context, {
\ 'text': l:chunk.str,
\ 'time': l:chunk.time,
\ 'filename': l:chunk.filename
\ })
endfor
" no samplers needed here
let l:request = json_encode({
\ 'input_prefix': "",
\ 'input_suffix': "",
\ 'input_extra': l:extra_context,
\ 'prompt': "",
\ 'n_predict': 1,
\ 'temperature': 0.0,
\ 'stream': v:false,
\ 'samplers': ["temperature"],
\ 'cache_prompt': v:true,
\ 't_max_prompt_ms': 1,
\ 't_max_predict_ms': 1
\ })
let l:curl_command = [
\ "curl",
\ "--silent",
\ "--no-buffer",
\ "--request", "POST",
\ "--url", g:llama_config.endpoint,
\ "--header", "Content-Type: application/json",
\ "--data", l:request
\ ]
" no callbacks because we don't need to process the response
if s:ghost_text_nvim
call jobstart(l:curl_command, {})
elseif s:ghost_text_vim
call job_start(l:curl_command, {})
endif
endfunction
" necessary for 'inoremap <expr>'
function! llama#fim_inline(is_auto) abort
call llama#fim(a:is_auto)
return ''
endfunction
" the main FIM call
" takes local context around the cursor and sends it together with the extra context to the server for completion
function! llama#fim(is_auto) abort
" we already have a suggestion for the current cursor position
if s:hint_shown && !a:is_auto
call llama#fim_cancel()
return
endif
call llama#fim_cancel()
" avoid sending repeated requests too fast
if reltimefloat(reltime(s:t_fim_start)) < 0.6
if s:timer_fim != -1
call timer_stop(s:timer_fim)
let s:timer_fim = -1
endif
let s:t_fim_start = reltime()
let s:timer_fim = timer_start(600, {-> llama#fim(v:true)})
return
endif
let s:t_fim_start = reltime()
let s:content = []
let s:can_accept = v:false
let s:pos_x = col('.') - 1
let s:pos_y = line('.')
let l:max_y = line('$')
let l:lines_prefix = getline(max([1, s:pos_y - g:llama_config.n_prefix]), s:pos_y - 1)
let l:lines_suffix = getline(s:pos_y + 1, min([l:max_y, s:pos_y + g:llama_config.n_suffix]))
let s:line_cur = getline('.')
let s:line_cur_prefix = strpart(s:line_cur, 0, s:pos_x)
let s:line_cur_suffix = strpart(s:line_cur, s:pos_x)
if a:is_auto && len(s:line_cur_suffix) > g:llama_config.max_line_suffix
return
endif
let l:prefix = ""
\ . join(l:lines_prefix, "\n")
\ . "\n"
let l:prompt = ""
\ . s:line_cur_prefix
let l:suffix = ""
\ . s:line_cur_suffix
\ . "\n"
\ . join(l:lines_suffix, "\n")
\ . "\n"
" prepare the extra context data
let l:extra_context = []
for l:chunk in s:ring_chunks
call add(l:extra_context, {
\ 'text': l:chunk.str,
\ 'time': l:chunk.time,
\ 'filename': l:chunk.filename
\ })
endfor
" the indentation of the current line
let l:indent = strlen(matchstr(s:line_cur_prefix, '^\s*'))
let l:request = json_encode({
\ 'input_prefix': l:prefix,
\ 'input_suffix': l:suffix,
\ 'input_extra': l:extra_context,
\ 'prompt': l:prompt,
\ 'n_predict': g:llama_config.n_predict,
\ 'n_indent': l:indent,
\ 'top_k': 40,
\ 'top_p': 0.99,
\ 'stream': v:false,
\ 'samplers': ["top_k", "top_p", "infill"],
\ 'cache_prompt': v:true,
\ 't_max_prompt_ms': g:llama_config.t_max_prompt_ms,
\ 't_max_predict_ms': g:llama_config.t_max_predict_ms
\ })
let l:curl_command = [
\ "curl",
\ "--silent",
\ "--no-buffer",
\ "--request", "POST",
\ "--url", g:llama_config.endpoint,
\ "--header", "Content-Type: application/json",
\ "--data", l:request
\ ]
if s:current_job != v:null
if s:ghost_text_nvim
call jobstop(s:current_job)
elseif s:ghost_text_vim
call job_stop(s:current_job)
endif
endif
" send the request asynchronously
if s:ghost_text_nvim
let s:current_job = jobstart(l:curl_command, {
\ 'on_stdout': function('s:fim_on_stdout', [s:pos_x, s:pos_y, a:is_auto]),
\ 'on_exit': function('s:fim_on_exit'),
\ 'stdout_buffered': v:true
\ })
elseif s:ghost_text_vim
let s:current_job = job_start(l:curl_command, {
\ 'out_cb': function('s:fim_on_stdout', [s:pos_x, s:pos_y, a:is_auto]),
\ 'exit_cb': function('s:fim_on_exit')
\ })
endif
" TODO: per-file location
let l:delta_y = abs(s:pos_y - s:pos_y_pick)
" gather some extra context nearby and process it in the background
" only gather chunks if the cursor has moved a lot
" TODO: something more clever? reranking?
if a:is_auto && l:delta_y > 32
" expand the prefix even further
call s:pick_chunk(getline(max([1, s:pos_y - g:llama_config.ring_scope]), max([1, s:pos_y - g:llama_config.n_prefix])), v:false, v:false)
" pick a suffix chunk
call s:pick_chunk(getline(min([l:max_y, s:pos_y + g:llama_config.n_suffix]), min([l:max_y, s:pos_y + g:llama_config.n_suffix + g:llama_config.ring_chunk_size])), v:false, v:false)
let s:pos_y_pick = s:pos_y
endif
endfunction
" if first_line == v:true accept only the first line of the response
function! llama#fim_accept(first_line)
" insert the suggestion at the cursor location
if s:can_accept && len(s:content) > 0
call setline(s:pos_y, s:line_cur[:(s:pos_x - 1)] . s:content[0])
if len(s:content) > 1
if !a:first_line
call append(s:pos_y, s:content[1:-1])
endif
endif
" move the cursor to the end of the accepted text
if !a:first_line && len(s:content) > 1
call cursor(s:pos_y + len(s:content) - 1, s:pos_x + s:pos_dx + 1)
else
call cursor(s:pos_y, s:pos_x + len(s:content[0]))
endif
endif
call llama#fim_cancel()
endfunction
function! llama#fim_cancel()
let s:hint_shown = v:false
" clear the virtual text
let l:bufnr = bufnr('%')
if s:ghost_text_nvim
let l:id_vt_fim = nvim_create_namespace('vt_fim')
call nvim_buf_clear_namespace(l:bufnr, l:id_vt_fim, 0, -1)
elseif s:ghost_text_vim
call prop_remove({'type': s:hlgroup_hint, 'all': v:true})
call prop_remove({'type': s:hlgroup_info, 'all': v:true})
endif
" remove the mappings
silent! iunmap <buffer> <Tab>
silent! iunmap <buffer> <S-Tab>
silent! iunmap <buffer> <Esc>
endfunction
function! s:on_move()
let s:t_last_move = reltime()
call llama#fim_cancel()
endfunction
" callback that processes the FIM result from the server and displays the suggestion
function! s:fim_on_stdout(pos_x, pos_y, is_auto, job_id, data, event = v:null)
if s:ghost_text_nvim
let l:raw = join(a:data, "\n")
elseif s:ghost_text_vim
let l:raw = a:data
endif
if len(l:raw) == 0
return
endif
if a:pos_x != col('.') - 1 || a:pos_y != line('.')
return
endif
" show the suggestion only in insert mode
if mode() !=# 'i'
return
endif
let s:pos_x = a:pos_x
let s:pos_y = a:pos_y
let s:can_accept = v:true
let l:has_info = v:false
if s:can_accept && v:shell_error
if !a:is_auto
call add(s:content, "<| curl error: is the server on? |>")
endif
let s:can_accept = v:false
endif
let l:n_prompt = 0
let l:t_prompt_ms = 1.0
let l:s_prompt = 0
let l:n_predict = 0
let l:t_predict_ms = 1.0
let l:s_predict = 0
" get the generated suggestion
if s:can_accept
let l:response = json_decode(l:raw)
for l:part in split(get(l:response, 'content', ''), "\n", 1)
call add(s:content, l:part)
endfor
" remove trailing new lines
while len(s:content) > 0 && s:content[-1] == ""
call remove(s:content, -1)
endwhile
let l:generation_settings = get(l:response, 'generation_settings', {})
let l:n_ctx = get(l:generation_settings, 'n_ctx', 0)
let l:n_cached = get(l:response, 'tokens_cached', 0)
let l:truncated = get(l:response, 'truncated', v:false)
" if response.timings is available
if len(get(l:response, 'timings', {})) > 0
let l:has_info = v:true
let l:timings = get(l:response, 'timings', {})
let l:n_prompt = get(l:timings, 'prompt_n', 0)
let l:t_prompt_ms = get(l:timings, 'prompt_ms', 1)
let l:s_prompt = get(l:timings, 'prompt_per_second', 0)
let l:n_predict = get(l:timings, 'predicted_n', 0)
let l:t_predict_ms = get(l:timings, 'predicted_ms', 1)
let l:s_predict = get(l:timings, 'predicted_per_second', 0)
endif
endif
if len(s:content) == 0
call add(s:content, "")
let s:can_accept = v:false
endif
if len(s:content) == 0
return
endif
" NOTE: the following is logic for discarding predictions that repeat existing text
" the code is quite ugly and there is very likely a simpler and more canonical way to implement this
"
" still, I wonder if there is some better way that avoids having to do these special hacks?
" on one hand, the LLM 'sees' the contents of the file before we start editing, so it is normal that it would
" start generating whatever we have given it via the extra context. but on the other hand, it's not very
" helpful to re-generate the same code that is already there
" truncate the suggestion if the first line is empty
if len(s:content) == 1 && s:content[0] == ""
let s:content = [""]
endif
" ... and the next lines are repeated
if len(s:content) > 1 && s:content[0] == "" && s:content[1:] == getline(s:pos_y + 1, s:pos_y + len(s:content) - 1)
let s:content = [""]
endif
" truncate the suggestion if it repeats the suffix
if len(s:content) == 1 && s:content[0] == s:line_cur_suffix
let s:content = [""]
endif
" find the first non-empty line (strip whitespace)
let l:cmp_y = s:pos_y + 1
while l:cmp_y < line('$') && getline(l:cmp_y) =~? '^\s*$'
let l:cmp_y += 1
endwhile
if (s:line_cur_prefix . s:content[0]) == getline(l:cmp_y)
" truncate the suggestion if it repeats the next line
if len(s:content) == 1
let s:content = [""]
endif
" ... or if the second line of the suggestion is the prefix of line l:cmp_y + 1
if len(s:content) == 2 && s:content[-1] == getline(l:cmp_y + 1)[:len(s:content[-1]) - 1]
let s:content = [""]
endif
" ... or if the middle chunk of lines of the suggestion is the same as [l:cmp_y + 1, l:cmp_y + len(s:content) - 1)
if len(s:content) > 2 && join(s:content[1:-1], "\n") == join(getline(l:cmp_y + 1, l:cmp_y + len(s:content) - 1), "\n")
let s:content = [""]
endif
endif
" keep only lines that have the same or larger whitespace prefix as s:line_cur_prefix
"let l:indent = strlen(matchstr(s:line_cur_prefix, '^\s*'))
"for i in range(1, len(s:content) - 1)
" if strlen(matchstr(s:content[i], '^\s*')) < l:indent
" let s:content = s:content[:i - 1]
" break
" endif
"endfor
let s:pos_dx = len(s:content[-1])
let s:content[-1] .= s:line_cur_suffix
call llama#fim_cancel()
" display virtual text with the suggestion
let l:bufnr = bufnr('%')
if s:ghost_text_nvim
let l:id_vt_fim = nvim_create_namespace('vt_fim')
endif
" construct the info message
if g:llama_config.show_info > 0 && l:has_info
let l:prefix = ' '
if l:truncated
let l:info = printf("%s | WARNING: the context is full: %d / %d, increase the server context size or reduce g:llama_config.ring_n_chunks",
\ g:llama_config.show_info == 2 ? l:prefix : 'llama.vim',
\ l:n_cached, l:n_ctx
\ )
else
let l:info = printf("%s | c: %d / %d, r: %d / %d, e: %d, q: %d / 16 | p: %d (%.2f ms, %.2f t/s) | g: %d (%.2f ms, %.2f t/s) | t: %.2f ms",
\ g:llama_config.show_info == 2 ? l:prefix : 'llama.vim',
\ l:n_cached, l:n_ctx, len(s:ring_chunks), g:llama_config.ring_n_chunks, s:ring_n_evict, len(s:ring_queued),
\ l:n_prompt, l:t_prompt_ms, l:s_prompt,
\ l:n_predict, l:t_predict_ms, l:s_predict,
\ 1000.0 * reltimefloat(reltime(s:t_fim_start))
\ )
endif
if g:llama_config.show_info == 1
" display the info in the statusline
let &statusline = l:info
let l:info = ''
endif
endif
" display the suggestion and append the info to the end of the first line
if s:ghost_text_nvim
call nvim_buf_set_extmark(l:bufnr, l:id_vt_fim, s:pos_y - 1, s:pos_x - 1, {
\ 'virt_text': [[s:content[0], 'llama_hl_hint'], [l:info, 'llama_hl_info']],
\ 'virt_text_win_col': virtcol('.') - 1
\ })
call nvim_buf_set_extmark(l:bufnr, l:id_vt_fim, s:pos_y - 1, 0, {
\ 'virt_lines': map(s:content[1:], {idx, val -> [[val, 'llama_hl_hint']]}),
\ 'virt_text_win_col': virtcol('.')
\ })
elseif s:ghost_text_vim
let l:new_suffix = s:content[0]
if !empty(l:new_suffix)
call prop_add(s:pos_y, s:pos_x + 1, {
\ 'type': s:hlgroup_hint,
\ 'text': l:new_suffix
\ })
endif
for line in s:content[1:]
call prop_add(s:pos_y, 0, {
\ 'type': s:hlgroup_hint,
\ 'text': line,
\ 'text_padding_left': s:get_indent(line),
\ 'text_align': 'below'
\ })
endfor
if !empty(l:info)
call prop_add(s:pos_y, 0, {
\ 'type': s:hlgroup_info,
\ 'text': l:info,
\ 'text_padding_left': col('$'),
\ 'text_wrap': 'truncate'
\ })
endif
endif
" setup accept shortcuts
inoremap <buffer> <Tab> <C-O>:call llama#fim_accept(v:false)<CR>
inoremap <buffer> <S-Tab> <C-O>:call llama#fim_accept(v:true)<CR>
let s:hint_shown = v:true
endfunction
function! s:fim_on_exit(job_id, exit_code, event = v:null)
if a:exit_code != 0
echom "Job failed with exit code: " . a:exit_code
endif
let s:current_job = v:null
endfunction

2
examples/llava/llava-cli.cpp
View File

@ -20,7 +20,7 @@ static bool eval_tokens(struct llama_context * ctx_llama, std::vector<llama_toke
if (n_eval > n_batch) {
n_eval = n_batch;
}
if (llama_decode(ctx_llama, llama_batch_get_one(&tokens[i], n_eval, *n_past, 0))) {
if (llama_decode(ctx_llama, llama_batch_get_one(&tokens[i], n_eval))) {
LOG_ERR("%s : failed to eval. token %d/%d (batch size %d, n_past %d)\n", __func__, i, N, n_batch, *n_past);
return false;
}

40
examples/llava/llava.cpp
View File

@ -401,6 +401,39 @@ bool llava_image_embed_make_with_clip_img(clip_ctx * ctx_clip, int n_threads, co
return true;
}
struct llava_embd_batch {
std::vector<llama_pos> pos;
std::vector<int32_t> n_seq_id;
std::vector<llama_seq_id> seq_id_0;
std::vector<llama_seq_id *> seq_ids;
std::vector<int8_t> logits;
llama_batch batch;
llava_embd_batch(float * embd, int32_t n_tokens, llama_pos pos_0, llama_seq_id seq_id) {
pos .resize(n_tokens);
n_seq_id.resize(n_tokens);
seq_ids .resize(n_tokens + 1);
logits .resize(n_tokens);
seq_id_0.resize(1);
seq_id_0[0] = seq_id;
seq_ids [n_tokens] = nullptr;
batch = {
/*n_tokens =*/ n_tokens,
/*tokens =*/ nullptr,
/*embd =*/ embd,
/*pos =*/ pos.data(),
/*n_seq_id =*/ n_seq_id.data(),
/*seq_id =*/ seq_ids.data(),
/*logits =*/ logits.data(),
};
for (int i = 0; i < n_tokens; i++) {
batch.pos [i] = pos_0 + i;
batch.n_seq_id[i] = 1;
batch.seq_id [i] = seq_id_0.data();
batch.logits [i] = false;
}
}
};
bool llava_eval_image_embed(llama_context * ctx_llama, const struct llava_image_embed * image_embed, int n_batch, int * n_past) {
int n_embd = llama_n_embd(llama_get_model(ctx_llama));
@ -409,8 +442,9 @@ bool llava_eval_image_embed(llama_context * ctx_llama, const struct llava_image_
if (n_eval > n_batch) {
n_eval = n_batch;
}
llama_batch batch = {int32_t(n_eval), nullptr, (image_embed->embed+i*n_embd), nullptr, nullptr, nullptr, nullptr, *n_past, 1, 0, };
if (llama_decode(ctx_llama, batch)) {
float * embd = image_embed->embed+i*n_embd;
llava_embd_batch llava_batch = llava_embd_batch(embd, n_eval, *n_past, 0);
if (llama_decode(ctx_llama, llava_batch.batch)) {
LOG_ERR("%s : failed to eval\n", __func__);
return false;
}
@ -432,7 +466,7 @@ struct llava_image_embed * llava_image_embed_make_with_bytes(struct clip_ctx * c
bool image_embed_result = llava_image_embed_make_with_clip_img(ctx_clip, n_threads, img, &image_embed, &n_image_pos);
if (!image_embed_result) {
clip_image_u8_free(img);
LOG_ERR("%s: coulnd't embed the image\n", __func__);
LOG_ERR("%s: couldn't embed the image\n", __func__);
return NULL;
}

2
examples/llava/minicpmv-cli.cpp
View File

@ -97,7 +97,7 @@ static bool eval_tokens(struct llama_context * ctx_llama, std::vector<llama_toke
if (n_eval > n_batch) {
n_eval = n_batch;
}
if (llama_decode(ctx_llama, llama_batch_get_one(&tokens[i], n_eval, *n_past, 0))) {
if (llama_decode(ctx_llama, llama_batch_get_one(&tokens[i], n_eval))) {
LOG_ERR("%s : failed to eval. token %d/%d (batch size %d, n_past %d)\n", __func__, i, N, n_batch, *n_past);
return false;
}

4
examples/lookahead/lookahead.cpp
View File

@ -89,8 +89,8 @@ int main(int argc, char ** argv) {
const auto t_enc_start = ggml_time_us();
// eval the prompt
llama_decode(ctx, llama_batch_get_one( inp.data(), n_input - 1, 0, 0));
llama_decode(ctx, llama_batch_get_one(&inp.back(), 1, n_input - 1, 0));
llama_decode(ctx, llama_batch_get_one( inp.data(), n_input - 1));
llama_decode(ctx, llama_batch_get_one(&inp.back(), 1));
for (int s = 1; s < W + G + 1; ++s) {
llama_kv_cache_seq_cp(ctx, 0, s, -1, -1);

4
examples/lookup/lookup.cpp
View File

@ -89,8 +89,8 @@ int main(int argc, char ** argv){
const auto t_enc_start = ggml_time_us();
llama_decode(ctx, llama_batch_get_one( inp.data(), n_input - 1, 0, 0));
llama_decode(ctx, llama_batch_get_one(&inp.back(), 1, n_input - 1, 0));
llama_decode(ctx, llama_batch_get_one( inp.data(), n_input - 1));
llama_decode(ctx, llama_batch_get_one(&inp.back(), 1));
const auto t_enc_end = ggml_time_us();

41
examples/main/README.md
View File

@ -187,6 +187,30 @@ Use the `--no-penalize-nl` option to disable newline penalization when applying
Example usage: `--repeat-penalty 1.15 --repeat-last-n 128 --no-penalize-nl`
### DRY Repetition Penalty
DRY (Don't Repeat Yourself) sampling is an effective technique for reducing repetition in generated text even across long contexts by penalizing tokens based on their recent usage patterns (original [PR link](https://github.com/oobabooga/text-generation-webui/pull/5677)).
- `--dry-multiplier N`: Set the DRY sampling multiplier (default: 0.0, 0.0 = disabled).
- `--dry-base N`: Set the DRY sampling base value (default: 1.75).
- `--dry-allowed-length N`: Set the allowed length for DRY sampling (default: 2).
- `--dry-penalty-last-n N`: Set DRY penalty for the last n tokens (default: -1, 0 = disable, -1 = context size).
- `--dry-sequence-breaker STRING`: Add a sequence breaker for DRY sampling. Can be used more than once to add multiple sequence breakers. Using this clears out the default breakers, which consist of: `['\n', ':', '"', '*']`. If the string `"none"` is supplied, no sequence breakers are used.
The `dry-multiplier` option controls the strength of the DRY sampling effect. A value of 0.0 disables DRY sampling, while higher values increase its influence. A typical recommended value is 0.8.
The `dry-base` option sets the base value for the exponential penalty calculation in DRY sampling. Higher values lead to more aggressive penalization of repetitions.
The `dry-allowed-length` option sets the maximum length of repeated sequences that will not be penalized. Repetitions shorter than or equal to this length are not penalized, allowing for natural repetitions of short phrases or common words.
The `dry-penalty-last-n` option controls how many recent tokens to consider when applying the DRY penalty. A value of -1 considers the entire context. Use a positive value to limit the consideration to a specific number of recent tokens.
The `dry-sequence-breaker` option adds a single sequence breaker and can be used more than once to specify multiple sequence breakers. Sequence breakers interrupt sequence matching and break the input into parts where matching can be applied.
DRY sampling provides more nuanced control over text generation, particularly for reducing long-range repetitions and maintaining global coherence.
Example usage: `--dry-multiplier 0.8 --dry-base 1.75 --dry-allowed-length 2 --dry-penalty-last-n -1 --dry-sequence-breaker "—" --dry-sequence-breaker "##"`
### Top-K Sampling
- `--top-k N`: Limit the next token selection to the K most probable tokens (default: 40).
@ -241,6 +265,19 @@ The `--mirostat-ent` option sets the Mirostat target entropy (tau), which repres
Example usage: `--mirostat 2 --mirostat-lr 0.05 --mirostat-ent 3.0`
### XTC Sampling
- `--xtc-probability N`: Sets the chance for token removal (checked once on sampler start) (default: 0.0).
- `--xtc-threshold N`: Sets a minimum probability threshold for tokens to be removed (default: 0.1).
Exclude Top Choices (XTC) is a unique sampler that is designed to remove top tokens from consideration and avoid more obvious and repetitive outputs. With a chance of `xtc-probability` it searches for tokens with probabilities of `xtc-threshold` and above, then removes all such tokens except the least probable one.
By removing top tokens XTC can improve the variety of answers, break writing clichés and inhibit repition, since clichés and repeated phrases are usually more likely to appear. By keeping the last token above the threshold, XTC ensures that the answer is still coherent. XTC is meant to be used for creative tasks, but feel free to experiment with different settings for different models.
Being experimental and unique, XTC is disabled by default. The recommended combination of samplers is Min-P followed by XTC on its default settings: `--sampling-seq mx --min-p 0.02 --xtc-probability 0.5`.
Example usage: `--xtc-probability 0.5 --xtc-threshold 0.1`
### Logit Bias
- `-l TOKEN_ID(+/-)BIAS, --logit-bias TOKEN_ID(+/-)BIAS`: Modify the likelihood of a token appearing in the generated text completion.
@ -284,10 +321,6 @@ These options help improve the performance and memory usage of the LLaMA models.
These flags attempt optimizations that help on some systems with non-uniform memory access. This currently consists of one of the above strategies, and disabling prefetch and readahead for mmap. The latter causes mapped pages to be faulted in on first access instead of all at once, and in combination with pinning threads to NUMA nodes, more of the pages end up on the NUMA node where they are used. Note that if the model is already in the system page cache, for example because of a previous run without this option, this will have little effect unless you drop the page cache first. This can be done by rebooting the system or on Linux by writing '3' to '/proc/sys/vm/drop_caches' as root.
### Memory Float 32
- `--memory-f32`: Use 32-bit floats instead of 16-bit floats for memory key+value. This doubles the context memory requirement and cached prompt file size but does not appear to increase generation quality in a measurable way. Not recommended.
### Batch Size
- `-b N, --batch-size N`: Set the batch size for prompt processing (default: `2048`). This large batch size benefits users who have BLAS installed and enabled it during the build. If you don't have BLAS enabled ("BLAS=0"), you can use a smaller number, such as 8, to see the prompt progress as it's evaluated in some situations.

50
examples/main/main.cpp
View File

@ -528,7 +528,7 @@ int main(int argc, char ** argv) {
int enc_input_size = embd_inp.size();
llama_token * enc_input_buf = embd_inp.data();
if (llama_encode(ctx, llama_batch_get_one(enc_input_buf, enc_input_size, 0, 0))) {
if (llama_encode(ctx, llama_batch_get_one(enc_input_buf, enc_input_size))) {
LOG_ERR("%s : failed to eval\n", __func__);
return 1;
}
@ -569,30 +569,30 @@ int main(int argc, char ** argv) {
if (!params.ctx_shift){
LOG_DBG("\n\n%s: context full and context shift is disabled => stopping\n", __func__);
break;
} else {
if (params.n_predict == -2) {
LOG_DBG("\n\n%s: context full and n_predict == -%d => stopping\n", __func__, params.n_predict);
break;
}
const int n_left = n_past - params.n_keep;
const int n_discard = n_left/2;
LOG_DBG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
n_past, n_left, n_ctx, params.n_keep, n_discard);
llama_kv_cache_seq_rm (ctx, 0, params.n_keep , params.n_keep + n_discard);
llama_kv_cache_seq_add(ctx, 0, params.n_keep + n_discard, n_past, -n_discard);
n_past -= n_discard;
LOG_DBG("after swap: n_past = %d\n", n_past);
LOG_DBG("embd: %s\n", string_from(ctx, embd).c_str());
LOG_DBG("clear session path\n");
path_session.clear();
}
if (params.n_predict == -2) {
LOG_DBG("\n\n%s: context full and n_predict == -%d => stopping\n", __func__, params.n_predict);
break;
}
const int n_left = n_past - params.n_keep;
const int n_discard = n_left/2;
LOG_DBG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
n_past, n_left, n_ctx, params.n_keep, n_discard);
llama_kv_cache_seq_rm (ctx, 0, params.n_keep , params.n_keep + n_discard);
llama_kv_cache_seq_add(ctx, 0, params.n_keep + n_discard, n_past, -n_discard);
n_past -= n_discard;
LOG_DBG("after swap: n_past = %d\n", n_past);
LOG_DBG("embd: %s\n", string_from(ctx, embd).c_str());
LOG_DBG("clear session path\n");
path_session.clear();
}
} else {
// context extension via Self-Extend
@ -648,7 +648,7 @@ int main(int argc, char ** argv) {
LOG_DBG("eval: %s\n", string_from(ctx, embd).c_str());
if (llama_decode(ctx, llama_batch_get_one(&embd[i], n_eval, n_past, 0))) {
if (llama_decode(ctx, llama_batch_get_one(&embd[i], n_eval))) {
LOG_ERR("%s : failed to eval\n", __func__);
return 1;
}

1
examples/parallel/parallel.cpp
View File

@ -308,7 +308,6 @@ int main(int argc, char ** argv) {
batch.n_seq_id + i,
batch.seq_id + i,
batch.logits + i,
0, 0, 0, // unused
};
const int ret = llama_decode(ctx, batch_view);

27
examples/perplexity/perplexity.cpp
View File

@ -408,14 +408,21 @@ static results_perplexity perplexity_v2(llama_context * ctx, const common_params
// clear the KV cache
llama_kv_cache_clear(ctx);
llama_batch batch = llama_batch_init(n_batch, 0, 1);
for (int j = 0; j < num_batches; ++j) {
const int batch_start = start + j * n_batch;
const int batch_size = std::min(end - batch_start, n_batch);
common_batch_clear(batch);
for (int i = 0; i < batch_size; i++) {
common_batch_add(batch, tokens[batch_start + i], j*n_batch + i, {0}, true);
}
//LOG_DBG(" Batch %d: starts at %d, size is %d, n_past is %d\n",j,batch_start,batch_size,j * n_batch);
// TODO: use llama_batch.logits instead of relying on logits_all == true
if (llama_decode(ctx, llama_batch_get_one(tokens.data() + batch_start, batch_size, j * n_batch, 0))) {
if (llama_decode(ctx, batch)) {
//LOG_ERR("%s : failed to eval\n", __func__);
llama_batch_free(batch);
return {tokens, -1, logit_history, prob_history};
}
@ -435,6 +442,8 @@ static results_perplexity perplexity_v2(llama_context * ctx, const common_params
}
}
llama_batch_free(batch);
const auto t_end = std::chrono::high_resolution_clock::now();
if (i == 0) {
@ -704,7 +713,6 @@ static bool decode_helper(llama_context * ctx, llama_batch & batch, std::vector<
batch.n_seq_id + i,
batch.seq_id + i,
batch.logits + i,
0, 0, 0, // unused
};
const int ret = llama_decode(ctx, batch_view);
@ -1791,6 +1799,8 @@ static void kl_divergence(llama_context * ctx, const common_params & params) {
// clear the KV cache
llama_kv_cache_clear(ctx);
llama_batch batch = llama_batch_init(n_batch, 0, 1);
for (int j = 0; j < num_batches; ++j) {
const int batch_start = start + j * n_batch;
const int batch_size = std::min(end - batch_start, n_batch);
@ -1803,9 +1813,14 @@ static void kl_divergence(llama_context * ctx, const common_params & params) {
tokens[batch_start] = llama_token_bos(llama_get_model(ctx));
}
// TODO: use llama_batch.logits instead of relying on logits_all == true
if (llama_decode(ctx, llama_batch_get_one(tokens.data() + batch_start, batch_size, j * n_batch, 0))) {
common_batch_clear(batch);
for (int i = 0; i < batch_size; i++) {
common_batch_add(batch, tokens[batch_start + i], j*n_batch + i, {0}, true);
}
if (llama_decode(ctx, batch)) {
LOG_ERR("%s : failed to eval\n", __func__);
llama_batch_free(batch);
return;
}
@ -1818,6 +1833,8 @@ static void kl_divergence(llama_context * ctx, const common_params & params) {
}
}
llama_batch_free(batch);
const auto t_end = std::chrono::high_resolution_clock::now();
if (i == 0) {

33
examples/save-load-state/save-load-state.cpp
View File

@ -42,15 +42,21 @@ int main(int argc, char ** argv) {
llama_sampler * smpl = llama_sampler_chain_init(sparams);
llama_sampler_chain_add(smpl, llama_sampler_init_softmax());
llama_sampler_chain_add(smpl, llama_sampler_init_dist(params.sparams.seed));
// tokenize prompt
auto tokens = common_tokenize(ctx, params.prompt, true);
// prepare the batch
llama_batch batch = llama_batch_init(tokens.size(), 0, 1);
for (size_t i = 0; i < tokens.size(); i++) {
common_batch_add(batch, tokens[i], i, {0}, false);
}
batch.logits[batch.n_tokens - 1] = true; // generate next token
// evaluate prompt
llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size(), n_past, 0));
n_past += tokens.size();
llama_decode(ctx, batch);
n_past += batch.n_tokens;
// save state (rng, logits, embedding and kv_cache) to file
{
@ -77,8 +83,12 @@ int main(int argc, char ** argv) {
printf("%s", next_token_str.c_str());
result0 += next_token_str;
if (llama_decode(ctx, llama_batch_get_one(&next_token, 1, n_past, 0))) {
common_batch_clear(batch);
common_batch_add(batch, next_token, n_past, {0}, true);
if (llama_decode(ctx, batch)) {
fprintf(stderr, "\n%s : failed to evaluate\n", __func__);
llama_batch_free(batch);
llama_free(ctx);
llama_free_model(model);
return 1;
@ -96,7 +106,6 @@ int main(int argc, char ** argv) {
llama_sampler * smpl2 = llama_sampler_chain_init(sparams);
llama_sampler_chain_add(smpl2, llama_sampler_init_softmax());
llama_sampler_chain_add(smpl2, llama_sampler_init_dist(params.sparams.seed));
printf("\nsecond run: %s", params.prompt.c_str());
@ -133,8 +142,12 @@ int main(int argc, char ** argv) {
printf("%s", next_token_str.c_str());
result1 += next_token_str;
if (llama_decode(ctx2, llama_batch_get_one(&next_token, 1, n_past, 0))) {
common_batch_clear(batch);
common_batch_add(batch, next_token, n_past, {0}, true);
if (llama_decode(ctx2, batch)) {
fprintf(stderr, "\n%s : failed to evaluate\n", __func__);
llama_batch_free(batch);
llama_free(ctx2);
llama_free_model(model);
return 1;
@ -156,7 +169,6 @@ int main(int argc, char ** argv) {
llama_sampler * smpl3 = llama_sampler_chain_init(sparams);
llama_sampler_chain_add(smpl3, llama_sampler_init_softmax());
llama_sampler_chain_add(smpl3, llama_sampler_init_dist(params.sparams.seed));
printf("\nsingle seq run: %s", params.prompt.c_str());
@ -221,8 +233,12 @@ int main(int argc, char ** argv) {
printf("%s", next_token_str.c_str());
result2 += next_token_str;
if (llama_decode(ctx3, llama_batch_get_one(&next_token, 1, n_past, 1))) {
common_batch_clear(batch);
common_batch_add(batch, next_token, n_past, {1}, true);
if (llama_decode(ctx3, batch)) {
fprintf(stderr, "\n%s : failed to evaluate\n", __func__);
llama_batch_free(batch);
llama_free(ctx3);
llama_free_model(model);
return 1;
@ -236,6 +252,7 @@ int main(int argc, char ** argv) {
llama_sampler_free(smpl2);
llama_sampler_free(smpl3);
llama_batch_free(batch);
llama_free(ctx3);
llama_free_model(model);

65
examples/server/README.md
View File

@ -60,8 +60,6 @@ The project is under active development, and we are [looking for feedback and co
| `--yarn-attn-factor N` | YaRN: scale sqrt(t) or attention magnitude (default: 1.0)<br/>(env: LLAMA_ARG_YARN_ATTN_FACTOR) |
| `--yarn-beta-slow N` | YaRN: high correction dim or alpha (default: 1.0)<br/>(env: LLAMA_ARG_YARN_BETA_SLOW) |
| `--yarn-beta-fast N` | YaRN: low correction dim or beta (default: 32.0)<br/>(env: LLAMA_ARG_YARN_BETA_FAST) |
| `-gan, --grp-attn-n N` | group-attention factor (default: 1)<br/>(env: LLAMA_ARG_GRP_ATTN_N) |
| `-gaw, --grp-attn-w N` | group-attention width (default: 512.0)<br/>(env: LLAMA_ARG_GRP_ATTN_W) |
| `-dkvc, --dump-kv-cache` | verbose print of the KV cache |
| `-nkvo, --no-kv-offload` | disable KV offload<br/>(env: LLAMA_ARG_NO_KV_OFFLOAD) |
| `-ctk, --cache-type-k TYPE` | KV cache data type for K (default: f16)<br/>(env: LLAMA_ARG_CACHE_TYPE_K) |
@ -116,6 +114,11 @@ The project is under active development, and we are [looking for feedback and co
| `--repeat-penalty N` | penalize repeat sequence of tokens (default: 1.0, 1.0 = disabled) |
| `--presence-penalty N` | repeat alpha presence penalty (default: 0.0, 0.0 = disabled) |
| `--frequency-penalty N` | repeat alpha frequency penalty (default: 0.0, 0.0 = disabled) |
| `--dry-multiplier N` | DRY sampling multiplier (default: 0.0, 0.0 = disabled) |
| `--dry-base N` | DRY sampling base value (default: 1.75) |
| `--dry-allowed-length N` | allowed length for DRY sampling (default: 2) |
| `--dry-penalty-last-n N` | DRY penalty for the last n tokens (default: -1, 0 = disable, -1 = context size) |
| `--dry-sequence-breaker STRING` | add sequence breaker for DRY sampling, clearing out default breakers (`['\n', ':', '"', '*']`) in the process; use `"none"` to not use any sequence breakers
| `--dynatemp-range N` | dynamic temperature range (default: 0.0, 0.0 = disabled) |
| `--dynatemp-exp N` | dynamic temperature exponent (default: 1.0) |
| `--mirostat N` | use Mirostat sampling.<br/>Top K, Nucleus, Tail Free and Locally Typical samplers are ignored if used.<br/>(default: 0, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0) |
@ -149,7 +152,7 @@ The project is under active development, and we are [looking for feedback and co
| `--ssl-cert-file FNAME` | path to file a PEM-encoded SSL certificate<br/>(env: LLAMA_ARG_SSL_CERT_FILE) |
| `-to, --timeout N` | server read/write timeout in seconds (default: 600)<br/>(env: LLAMA_ARG_TIMEOUT) |
| `--threads-http N` | number of threads used to process HTTP requests (default: -1)<br/>(env: LLAMA_ARG_THREADS_HTTP) |
| `-spf, --system-prompt-file FNAME` | set a file to load a system prompt (initial prompt of all slots), this is useful for chat applications |
| `--cache-reuse N` | min chunk size to attempt reusing from the cache via KV shifting (default: 0)<br/>(env: LLAMA_ARG_CACHE_REUSE) |
| `--metrics` | enable prometheus compatible metrics endpoint (default: disabled)<br/>(env: LLAMA_ARG_ENDPOINT_METRICS) |
| `--slots` | enable slots monitoring endpoint (default: disabled)<br/>(env: LLAMA_ARG_ENDPOINT_SLOTS) |
| `--props` | enable changing global properties via POST /props (default: disabled)<br/>(env: LLAMA_ARG_ENDPOINT_PROPS) |
@ -320,7 +323,18 @@ node index.js
- The prompt is a string or an array with the first element given as a string
- The model's `tokenizer.ggml.add_bos_token` metadata is `true`
- The system prompt is empty
These input shapes and data type are allowed for `prompt`:
- Single string: `"string"`
- Single sequence of tokens: `[12, 34, 56]`
- Mixed tokens and strings: `[12, 34, "string", 56, 78]`
Multiple prompts are also supported. In this case, the completion result will be an array.
- Only strings: `["string1", "string2"]`
- Strings and sequences of tokens: `["string1", [12, 34, 56]]`
- Mixed types: `[[12, 34, "string", 56, 78], [12, 34, 56], "string"]`
`temperature`: Adjust the randomness of the generated text. Default: `0.8`
@ -336,6 +350,8 @@ node index.js
`n_predict`: Set the maximum number of tokens to predict when generating text. **Note:** May exceed the set limit slightly if the last token is a partial multibyte character. When 0, no tokens will be generated but the prompt is evaluated into the cache. Default: `-1`, where `-1` is infinity.
`n_indent`: Specify the minimum line indentation for the generated text in number of whitespace characters. Useful for code completion tasks. Default: `0`
`n_keep`: Specify the number of tokens from the prompt to retain when the context size is exceeded and tokens need to be discarded. The number excludes the BOS token.
By default, this value is set to `0`, meaning no tokens are kept. Use `-1` to retain all tokens from the prompt.
@ -358,6 +374,16 @@ node index.js
`frequency_penalty`: Repeat alpha frequency penalty. Default: `0.0`, which is disabled.
`dry_multiplier`: Set the DRY (Don't Repeat Yourself) repetition penalty multiplier. Default: `0.0`, which is disabled.
`dry_base`: Set the DRY repetition penalty base value. Default: `1.75`
`dry_allowed_length`: Tokens that extend repetition beyond this receive exponentially increasing penalty: multiplier * base ^ (length of repeating sequence before token - allowed length). Default: `2`
`dry_penalty_last_n`: How many tokens to scan for repetitions. Default: `-1`, where `0` is disabled and `-1` is context size.
`dry_sequence_breakers`: Specify an array of sequence breakers for DRY sampling. Only a JSON array of strings is accepted. Default: `['\n', ':', '"', '*']`
`mirostat`: Enable Mirostat sampling, controlling perplexity during text generation. Default: `0`, where `0` is disabled, `1` is Mirostat, and `2` is Mirostat 2.0.
`mirostat_tau`: Set the Mirostat target entropy, parameter tau. Default: `5.0`
@ -378,6 +404,8 @@ node index.js
`min_keep`: If greater than 0, force samplers to return N possible tokens at minimum. Default: `0`
`t_max_predict_ms`: Set a time limit in milliseconds for the prediction (a.k.a. text-generation) phase. The timeout will trigger if the generation takes more than the specified time (measured since the first token was generated) and if a new-line character has already been generated. Useful for FIM applications. Default: `0`, which is disabled.
`image_data`: An array of objects to hold base64-encoded image `data` and its `id`s to be reference in `prompt`. You can determine the place of the image in the prompt as in the following: `USER:[img-12]Describe the image in detail.\nASSISTANT:`. In this case, `[img-12]` will be replaced by the embeddings of the image with id `12` in the following `image_data` array: `{..., "image_data": [{"data": "<BASE64_STRING>", "id": 12}]}`. Use `image_data` only with multimodal models, e.g., LLaVA.
`id_slot`: Assign the completion task to an specific slot. If is -1 the task will be assigned to a Idle slot. Default: `-1`
@ -525,8 +553,31 @@ Takes a prefix and a suffix and returns the predicted completion as stream.
- `input_prefix`: Set the prefix of the code to infill.
- `input_suffix`: Set the suffix of the code to infill.
- `input_extra`: Additional context inserted before the FIM prefix.
- `prompt`: Added after the `FIM_MID` token
It also accepts all the options of `/completion` except `stream` and `prompt`.
`input_extra` is array of `{"filename": string, "text": string}` objects.
The endpoint also accepts all the options of `/completion`.
If the model has `FIM_REPO` and `FIM_FILE_SEP` tokens, the [repo-level pattern](https://arxiv.org/pdf/2409.12186) is used:
```txt
<FIM_REP>myproject
<FIM_SEP>{chunk 0 filename}
{chunk 0 text}
<FIM_SEP>{chunk 1 filename}
{chunk 1 text}
...
<FIM_SEP>filename
<FIM_PRE>[input_prefix]<FIM_SUF>[input_suffix]<FIM_MID>[prompt]
```
If the tokens are missing, then the extra context is simply prefixed at the start:
```txt
[input_extra]<FIM_PRE>[input_prefix]<FIM_SUF>[input_suffix]<FIM_MID>[prompt]
```
### **GET** `/props`: Get server global properties.
@ -536,14 +587,12 @@ This endpoint is public (no API key check). By default, it is read-only. To make
```json
{
"system_prompt": "",
"default_generation_settings": { ... },
"total_slots": 1,
"chat_template": ""
}
```
- `system_prompt` - the system prompt (initial prompt of all slots). Please note that this does not take into account the chat template. It will append the prompt at the beginning of formatted prompt.
- `default_generation_settings` - the default generation settings for the `/completion` endpoint, which has the same fields as the `generation_settings` response object from the `/completion` endpoint.
- `total_slots` - the total number of slots for process requests (defined by `--parallel` option)
- `chat_template` - the model's original Jinja2 prompt template
@ -554,7 +603,7 @@ To use this endpoint with POST method, you need to start server with `--props`
*Options:*
- `system_prompt`: Change the system prompt (initial prompt of all slots). Please note that this does not take into account the chat template. It will append the prompt at the beginning of formatted prompt.
- None yet
### POST `/v1/chat/completions`: OpenAI-compatible Chat Completions API

20
examples/server/public/index-new.html
View File

@ -40,9 +40,15 @@
repeat_last_n: 0, // 0 = disable penalty, -1 = context size
repeat_penalty: 1.0, // 1.0 = disabled
penalize_nl: false, // true only useful for infinite completion
dry_multiplier: 0.0, // 0.0 = disabled, 0.8 works well
dry_base: 1.75, // 0.0 = disabled
dry_allowed_length: 2, // tokens extending repetitions beyond this receive penalty, 2 works well
dry_penalty_last_n: -1, // how many tokens to scan for repetitions (0 = disable penalty, -1 = context size)
top_k: 0, // <= 0 to use vocab size
top_p: 1.0, // 1.0 = disabled
min_p: 0.05, // 0 = disabled; recommended for non-english: ~ 0.4
xtc_probability: 0.0, // 0 = disabled;
xtc_threshold: 0.1, // > 0.5 disables XTC;
tfs_z: 1.0, // 1.0 = disabled
typical_p: 1.0, // 1.0 = disabled
presence_penalty: 0.0, // 0.0 = disabled
@ -831,11 +837,17 @@ return html`
<fieldset class="params">
${IntField({ label: "Top-K", title: "Limits the selection of the next token to the K most probable tokens. 1 means no randomness = greedy sampling. If set to 0, it means the entire vocabulary size is considered.", max: 100, min: 0, step: 1, name: "top_k", value: params.value.top_k })}
${IntField({ label: "Penalize Last N", title: "The last n tokens that are taken into account to penalise repetitions. A value of 0 means that this function is deactivated and -1 means that the entire size of the context is taken into account.", max: 2048, min: 0, step: 16, name: "repeat_last_n", value: params.value.repeat_last_n })}
${FloatField({ label: "Top-P", title: "Limits the selection of the next token to a subset of tokens whose combined probability reaches a threshold value P = top-P. If set to 1, it means the entire vocabulary size is considered.", max: 1.0, min: 0.0, name: "top_p", step: 0.01, value: params.value.top_p })}
${FloatField({ label: "Presence Penalty", title: "A penalty that is applied if certain tokens appear repeatedly in the generated text. A higher value leads to fewer repetitions.", max: 1.0, min: 0.0, name: "presence_penalty", step: 0.01, value: params.value.presence_penalty })}
${FloatField({ label: "TFS-Z", title: "Activates tail-free sampling, a method used to limit the prediction of tokens that are too frequent. The parameter z controls the strength of this limitation. A value of 1.0 means that this function is deactivated.", max: 1.0, min: 0.0, name: "tfs_z", step: 0.01, value: params.value.tfs_z })}
${FloatField({ label: "Frequency Penalty", title: "A penalty that is applied based on the frequency with which certain tokens occur in the training data set. A higher value results in rare tokens being favoured.", max: 1.0, min: 0.0, name: "frequency_penalty", step: 0.01, value: params.value.frequency_penalty })}
${FloatField({ label: "Top-P", title: "Limits the selection of the next token to a subset of tokens whose combined probability reaches a threshold value P = top-P. If set to 1, it means the entire vocabulary size is considered.", max: 1.0, min: 0.0, name: "top_p", step: 0.01, value: params.value.top_p })}
${FloatField({ label: "Typical-P", title: "Activates local typical sampling, a method used to limit the prediction of tokens that are atypical in the current context. The parameter p controls the strength of this limitation. A value of 1.0 means that this function is deactivated.", max: 1.0, min: 0.0, name: "typical_p", step: 0.01, value: params.value.typical_p })}
${FloatField({ label: "XTC probability", title: "Sets the chance for token removal (checked once on sampler start)", max: 1.0, min: 0.0, name: "xtc_probability", step: 0.01, value: params.value.xtc_probability })}
${FloatField({ label: "XTC threshold", title: "Sets a minimum probability threshold for tokens to be removed", max: 0.5, min: 0.0, name: "xtc_threshold", step: 0.01, value: params.value.xtc_threshold })}
${FloatField({ label: "DRY Penalty Multiplier", title: "Set the DRY repetition penalty multiplier. Default is 0.0, which disables DRY.", max: 5.0, min: 0.0, name: "dry_multiplier", step: 0.01, value: params.value.dry_multiplier })}
${FloatField({ label: "DRY Base", title: "Set the DRY repetition penalty base value. Default is 1.75", max: 3.0, min: 1.0, name: "dry_base", step: 0.01, value: params.value.dry_base })}
${IntField({ label: "DRY Allowed Length", title: "Tokens that extend repetition beyond this receive exponentially increasing penalty. Default is 2", max: 10, min: 1, step: 1, name: "dry_allowed_length", value: params.value.dry_allowed_length })}
${IntField({ label: "DRY Penalty Last N", title: "How many tokens to scan for repetitions. Default is -1, where 0 is disabled and -1 is context size", max: 2048, min: -1, step: 16, name: "dry_penalty_last_n", value: params.value.dry_penalty_last_n })}
${FloatField({ label: "TFS-Z", title: "Activates tail-free sampling, a method used to limit the prediction of tokens that are too frequent. The parameter z controls the strength of this limitation. A value of 1.0 means that this function is deactivated.", max: 1.0, min: 0.0, name: "tfs_z", step: 0.01, value: params.value.tfs_z })}
${IntField({ label: "Min Keep", title: "If greater than 0, samplers are forced to return N possible tokens at minimum. Default is 0", max: 10, min: 0, name: "min_keep", value: params.value.min_keep })}
</fieldset>
@ -1132,12 +1144,16 @@ document.addEventListener('DOMContentLoaded', (event) => {
const snapSettings = {
temperature: { snapValue: 1.0, snapRangeMultiplier: 6 },
min_p: { snapValue: 0.05, snapRangeMultiplier: 2 },
xtc_probability: { snapValue: 0.0, snapRangeMultiplier: 4 },
xtc_threshold: { snapValue: 0.5, snapRangeMultiplier: 4 },
top_p: { snapValue: 1.0, snapRangeMultiplier: 4 },
tfs_z: { snapValue: 1.0, snapRangeMultiplier: 4 },
typical_p: { snapValue: 1.0, snapRangeMultiplier: 4 },
repeat_penalty: { snapValue: 1.0, snapRangeMultiplier: 4 },
presence_penalty: { snapValue: 0.0, snapRangeMultiplier: 4 },
frequency_penalty: { snapValue: 0.0, snapRangeMultiplier: 4 },
dry_multiplier: { snapValue: 0.0, snapRangeMultiplier: 4 },
dry_base: { snapValue: 1.75, snapRangeMultiplier: 4 },
};
// add an event listener for each slider
Object.keys(snapSettings).forEach(sliderName => {

12
examples/server/public/index.html
View File

@ -304,9 +304,15 @@
repeat_last_n: 256, // 0 = disable penalty, -1 = context size
repeat_penalty: 1.18, // 1.0 = disabled
penalize_nl: false,
dry_multiplier: 0.0, // 0.0 = disabled, 0.8 works well
dry_base: 1.75, // 0.0 = disabled
dry_allowed_length: 2, // tokens extending repetitions beyond this receive penalty, 2 works well
dry_penalty_last_n: -1, // how many tokens to scan for repetitions (0 = disable penalty, -1 = context size)
top_k: 40, // <= 0 to use vocab size
top_p: 0.95, // 1.0 = disabled
min_p: 0.05, // 0 = disabled
xtc_probability: 0.0, // 0 = disabled;
xtc_threshold: 0.1, // > 0.5 disables XTC;
tfs_z: 1.0, // 1.0 = disabled
typical_p: 1.0, // 1.0 = disabled
presence_penalty: 0.0, // 0.0 = disabled
@ -1013,6 +1019,12 @@
${FloatField({ label: "Typical P", max: 1.0, min: 0.0, name: "typical_p", step: 0.01, value: params.value.typical_p })}
${FloatField({ label: "Presence penalty", max: 1.0, min: 0.0, name: "presence_penalty", step: 0.01, value: params.value.presence_penalty })}
${FloatField({ label: "Frequency penalty", max: 1.0, min: 0.0, name: "frequency_penalty", step: 0.01, value: params.value.frequency_penalty })}
${FloatField({ label: "DRY Penalty Multiplier", max: 5.0, min: 0.0, name: "dry_multiplier", step: 0.01, value: params.value.dry_multiplier })}
${FloatField({ label: "DRY Base", max: 3.0, min: 1.0, name: "dry_base", step: 0.01, value: params.value.dry_base })}
${IntField({ label: "DRY Allowed Length", max: 10, min: 2, step: 1, name: "dry_allowed_length", value: params.value.dry_allowed_length })}
${IntField({ label: "DRY Penalty Last N", max: 2048, min: -1, step: 16, name: "dry_penalty_last_n", value: params.value.dry_penalty_last_n })}
${FloatField({ label: "XTC probability", max: 1.0, min: 0.0, name: "xtc_probability", step: 0.01, value: params.value.xtc_probability })}
${FloatField({ label: "XTC threshold", max: 0.5, min: 0.0, name: "xtc_threshold", step: 0.01, value: params.value.xtc_threshold })}
</fieldset>
<hr />
<fieldset class="three">

2
examples/server/public/index.js
View File

File diff suppressed because one or more lines are too long

2
examples/server/public/json-schema-to-grammar.mjs
View File

@ -529,7 +529,7 @@ export class SchemaConverter {
return joinSeq();
};
return this._addRule(name, "\"\\\"\" " + toRule(transform()) + " \"\\\"\" space")
return this._addRule(name, "\"\\\"\" (" + toRule(transform()) + ") \"\\\"\" space")
}
_notStrings(strings) {

0
examples/server/public/style.css Executable file → Normal file
View File

916
examples/server/server.cpp
View File

File diff suppressed because it is too large Load Diff

4
examples/server/tests/features/ctx_shift.feature
View File

@ -13,6 +13,10 @@ Feature: llama.cpp server
And 32 as batch size
And 2 slots
# the prompt is 301 tokens
# the slot context is 256/2 = 128 tokens
# the prompt is truncated to keep the last 109 tokens
# 64 tokens are generated thanks to shifting the context when it gets full
Scenario: Inference with context shift
And 64 server max tokens to predict
Then the server is starting

36
examples/server/tests/features/infill.feature Normal file
View File

@ -0,0 +1,36 @@
@llama.cpp
@infill
Feature: llama.cpp server
# The current model is made by adding FIM tokens to the existing stories260K
# We may want to use a better model in the future, maybe something like SmolLM 360M
Background: Server startup
Given a server listening on localhost:8080
And a model file tinyllamas/stories260K-infill.gguf from HF repo ggml-org/models
And a model file test-model-infill.gguf
And a model alias tinyllama-infill
And 42 as server seed
And 1024 as batch size
And 1024 as ubatch size
And 2048 KV cache size
And 64 max tokens to predict
And 0.0 temperature
Then the server is starting
Then the server is healthy
Scenario: Infill without input_extra
Given a prompt "Complete this"
And an infill input extra none none
And an infill input prefix "#include <cstdio>\n#include \"llama.h\"\n\nint main() {\n int n_threads = llama_"
And an infill input suffix "}\n"
And an infill request with no api error
Then 64 tokens are predicted matching One|day|she|saw|big|scary|bird
Scenario: Infill with input_extra
Given a prompt "Complete this"
And an infill input extra "llama.h" "LLAMA_API int32_t llama_n_threads();\n"
And an infill input prefix "#include <cstdio>\n#include \"llama.h\"\n\nint main() {\n int n_threads = llama_"
And an infill input suffix "}\n"
And an infill request with no api error
Then 64 tokens are predicted matching cuts|Jimmy|mom|came|into|the|room"

46
examples/server/tests/features/steps/steps.py
View File

@ -80,6 +80,11 @@ def step_server_config(context, server_fqdn: str, server_port: str):
context.lora_file = None
context.disable_ctx_shift = False
# infill
context.infill_input_extra = None
context.infill_input_suffix = ''
context.infill_input_prefix = ''
context.tasks_result = []
context.concurrent_tasks = []
context.prompts = []
@ -291,6 +296,28 @@ async def step_request_completion(context, api_error: Literal['raised'] | str):
assert completion == api_error_code, f"completion must be an {api_error_code} status code: {completion}"
@step('an infill request with {api_error} api error')
@async_run_until_complete
async def step_request_completion(context, api_error: Literal['raised'] | str):
if api_error != 'no':
raise ValueError(f'api_error={api_error} is not yet implemented')
payload = {
"prompt": context.prompts[0],
"input_suffix": context.infill_input_suffix,
"input_prefix": context.infill_input_prefix,
"n_predict": context.n_predict,
"seed": context.seed,
"temperature": context.temperature,
}
if context.infill_input_extra is not None:
payload['input_extra'] = context.infill_input_extra
async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
async with session.post(f'{context.base_url}/infill',
json=payload) as response:
assert response.status == 200
context.tasks_result = [await response.json()]
@step('{predicted_n:d} tokens are predicted matching {re_content}')
def step_n_tokens_predicted_with_content(context, predicted_n, re_content):
context.completion = context.tasks_result.pop()
@ -539,6 +566,25 @@ def step_a_prompt_prompt(context, prompt):
context.n_prompts = len(context.prompts)
# TODO: allow this to be repeated
@step('an infill input extra {filename} {text}')
def step_infill_input_extra(context, filename, text):
if filename == 'none':
context.infill_input_extra = None
else:
context.infill_input_extra = [{'filename': filename, 'text': text}]
@step('an infill input suffix {text}')
def step_infill_input_suffix(context, text):
context.infill_input_suffix = text
@step('an infill input prefix {text}')
def step_infill_input_prefix(context, text):
context.infill_input_prefix = text
@step('{num_prompts:d} prompts {prompt} with seed {seed:d}')
def step_many_prompts(context, num_prompts, prompt, seed):
if context.seed is None:

264
examples/server/utils.hpp
View File

@ -24,6 +24,22 @@
#define DEFAULT_OAICOMPAT_MODEL "gpt-3.5-turbo-0613"
using json = nlohmann::ordered_json;
using llama_tokens = std::vector<llama_token>;
#define SLT_INF(slot, fmt, ...) LOG_INF("slot %12.*s: id %2d | task %d | " fmt, 12, __func__, (slot).id, (slot).id_task, __VA_ARGS__)
#define SLT_WRN(slot, fmt, ...) LOG_WRN("slot %12.*s: id %2d | task %d | " fmt, 12, __func__, (slot).id, (slot).id_task, __VA_ARGS__)
#define SLT_ERR(slot, fmt, ...) LOG_ERR("slot %12.*s: id %2d | task %d | " fmt, 12, __func__, (slot).id, (slot).id_task, __VA_ARGS__)
#define SLT_DBG(slot, fmt, ...) LOG_DBG("slot %12.*s: id %2d | task %d | " fmt, 12, __func__, (slot).id, (slot).id_task, __VA_ARGS__)
#define SRV_INF(fmt, ...) LOG_INF("srv %12.*s: " fmt, 12, __func__, __VA_ARGS__)
#define SRV_WRN(fmt, ...) LOG_WRN("srv %12.*s: " fmt, 12, __func__, __VA_ARGS__)
#define SRV_ERR(fmt, ...) LOG_ERR("srv %12.*s: " fmt, 12, __func__, __VA_ARGS__)
#define SRV_DBG(fmt, ...) LOG_DBG("srv %12.*s: " fmt, 12, __func__, __VA_ARGS__)
#define QUE_INF(fmt, ...) LOG_INF("que %12.*s: " fmt, 12, __func__, __VA_ARGS__)
#define QUE_WRN(fmt, ...) LOG_WRN("que %12.*s: " fmt, 12, __func__, __VA_ARGS__)
#define QUE_ERR(fmt, ...) LOG_ERR("que %12.*s: " fmt, 12, __func__, __VA_ARGS__)
#define QUE_DBG(fmt, ...) LOG_DBG("que %12.*s: " fmt, 12, __func__, __VA_ARGS__)
// https://community.openai.com/t/openai-chat-list-of-error-codes-and-types/357791/11
enum error_type {
@ -52,9 +68,235 @@ static T json_value(const json & body, const std::string & key, const T & defaul
}
//
// chat template utils
// tokenizer and input processing utils
//
static bool json_is_array_of_numbers(const json & data) {
if (data.is_array()) {
for (const auto & e : data) {
if (!e.is_number_integer()) {
return false;
}
}
return true;
}
return false;
}
// is array having BOTH numbers & strings?
static bool json_is_array_of_mixed_numbers_strings(const json & data) {
bool seen_string = false;
bool seen_number = false;
if (data.is_array()) {
for (const auto & e : data) {
seen_string |= e.is_string();
seen_number |= e.is_number_integer();
if (seen_number && seen_string) {
return true;
}
}
}
return false;
}
/**
* this handles 2 cases:
* - only string, example: "string"
* - mixed string and tokens, example: [12, 34, "string", 56, 78]
*/
static llama_tokens tokenize_mixed(const llama_context * ctx, const json & json_prompt, bool add_special, bool parse_special) {
// If `add_bos` is true, we only add BOS, when json_prompt is a string,
// or the first element of the json_prompt array is a string.
llama_tokens prompt_tokens;
if (json_prompt.is_array()) {
bool first = true;
for (const auto & p : json_prompt) {
if (p.is_string()) {
auto s = p.template get<std::string>();
llama_tokens p;
if (first) {
p = common_tokenize(ctx, s, add_special, parse_special);
first = false;
} else {
p = common_tokenize(ctx, s, false, parse_special);
}
prompt_tokens.insert(prompt_tokens.end(), p.begin(), p.end());
} else {
if (first) {
first = false;
}
prompt_tokens.push_back(p.template get<llama_token>());
}
}
} else {
auto s = json_prompt.template get<std::string>();
prompt_tokens = common_tokenize(ctx, s, add_special, parse_special);
}
return prompt_tokens;
}
/**
* break the input "prompt" object into multiple prompt if needed, then tokenize them
* this supports these cases:
* - "prompt": "string"
* - "prompt": [12, 34, 56]
* - "prompt": [12, 34, "string", 56, 78]
* and multiple prompts (multi-tasks):
* - "prompt": ["string1", "string2"]
* - "prompt": ["string1", [12, 34, 56]]
* - "prompt": [[12, 34, "string", 56, 78], [12, 34, 56]]
*/
static std::vector<llama_tokens> tokenize_input_prompts(llama_context * ctx, const json & json_prompt, bool add_special, bool parse_special) {
std::vector<llama_tokens> result;
if (json_prompt.is_string() || json_is_array_of_mixed_numbers_strings(json_prompt)) {
// string or mixed
result.push_back(tokenize_mixed(ctx, json_prompt, add_special, parse_special));
} else if (json_is_array_of_numbers(json_prompt)) {
// array of tokens
result.push_back(json_prompt.get<llama_tokens>());
} else if (json_prompt.is_array()) {
// array of prompts
result.reserve(json_prompt.size());
for (const auto & p : json_prompt) {
if (p.is_string() || json_is_array_of_mixed_numbers_strings(p)) {
result.push_back(tokenize_mixed(ctx, p, add_special, parse_special));
} else if (json_is_array_of_numbers(p)) {
// array of tokens
result.push_back(p.get<llama_tokens>());
} else {
throw std::runtime_error("element of \"prompt\" must be a string, an list of tokens, or a list of mixed strings & tokens");
}
}
} else {
throw std::runtime_error("\"prompt\" must be a string, an list of tokens, a list of mixed strings & tokens, or a list of prompts");
}
return result;
}
//
// template utils
//
// format rerank task: [BOS]query[EOS][SEP]doc[EOS]
static llama_tokens format_rerank(const struct llama_model * model, const llama_tokens & query, const llama_tokens & doc) {
llama_tokens result;
result.reserve(doc.size() + query.size() + 4);
result.push_back(llama_token_bos(model));
result.insert(result.end(), query.begin(), query.end());
result.push_back(llama_token_eos(model));
result.push_back(llama_token_sep(model));
result.insert(result.end(), doc.begin(), doc.end());
result.push_back(llama_token_eos(model));
return result;
}
// format infill task
static llama_tokens format_infill(
const llama_context * ctx,
const json & input_prefix,
const json & input_suffix,
const json & input_extra,
const int n_batch,
const int n_predict,
const int n_ctx,
const bool spm_infill,
const llama_tokens & tokens_prompt
) {
// TODO: optimize this block by reducing memory allocations and movement
// use FIM repo-level pattern:
// ref: https://arxiv.org/pdf/2409.12186
//
// [FIM_REP]myproject
// [FIM_SEP]filename0
// extra chunk 0
// [FIM_SEP]filename1
// extra chunk 1
// ...
// [FIM_SEP]filename
// [FIM_PRE]prefix[FIM_SUF]suffix[FIM_MID]prompt
//
llama_tokens extra_tokens;
extra_tokens.reserve(n_ctx);
auto model = llama_get_model(ctx);
auto tokens_prefix = tokenize_mixed(ctx, input_prefix, false, false);
auto tokens_suffix = tokenize_mixed(ctx, input_suffix, false, false);
if (llama_token_fim_rep(model) != LLAMA_TOKEN_NULL) {
// TODO: make project name an input
static const auto k_fim_repo = common_tokenize(ctx, "myproject\n", false, false);
extra_tokens.push_back(llama_token_fim_rep(model));
extra_tokens.insert(extra_tokens.end(), k_fim_repo.begin(), k_fim_repo.end());
}
for (const auto & chunk : input_extra) {
// { "text": string, "filename": string }
const std::string text = json_value(chunk, "text", std::string());
const std::string filename = json_value(chunk, "filename", std::string("tmp"));
if (llama_token_fim_sep(model) != LLAMA_TOKEN_NULL) {
const auto k_fim_file = common_tokenize(ctx, filename + "\n", false, false);
extra_tokens.insert(extra_tokens.end(), llama_token_fim_sep(model));
extra_tokens.insert(extra_tokens.end(), k_fim_file.begin(), k_fim_file.end());
} else {
// chunk separator in binary form to avoid confusing the AI
static const char k_chunk_prefix_str[] = {0x0a, 0x0a, 0x2d, 0x2d, 0x2d, 0x20, 0x73, 0x6e, 0x69, 0x70, 0x70, 0x65, 0x74, 0x20, 0x2d, 0x2d, 0x2d, 0x0a, 0x0a, 0x00};
static const auto k_chunk_prefix_tokens = common_tokenize(ctx, k_chunk_prefix_str, false, false);
extra_tokens.insert(extra_tokens.end(), k_chunk_prefix_tokens.begin(), k_chunk_prefix_tokens.end());
}
const auto chunk_tokens = common_tokenize(ctx, text, false, false);
extra_tokens.insert(extra_tokens.end(), chunk_tokens.begin(), chunk_tokens.end());
}
if (llama_token_fim_sep(model) != LLAMA_TOKEN_NULL) {
// TODO: current filename
static const auto k_fim_file = common_tokenize(ctx, "filename\n", false, false);
extra_tokens.insert(extra_tokens.end(), llama_token_fim_sep(model));
extra_tokens.insert(extra_tokens.end(), k_fim_file.begin(), k_fim_file.end());
}
// for now pick FIM context to fit in a batch (ratio prefix:suffix = 3:1, TODO: configurable?)
const int n_suffix_take = std::min<int>(tokens_suffix.size(), (n_batch/4));
const int n_prefix_take = std::min<int>(tokens_prefix.size(), 3*(n_batch/4) - 3);
// fill the rest of the context with extra chunks
const int n_extra_take = std::min<int>(std::max<int>(0, n_ctx - (n_batch) - 2*n_predict), extra_tokens.size());
tokens_prefix.erase(tokens_prefix.begin(), tokens_prefix.begin() + tokens_prefix.size() - n_prefix_take);
tokens_suffix.resize(n_suffix_take);
tokens_prefix.insert(tokens_prefix.begin(), llama_token_fim_pre(model));
tokens_prefix.insert(tokens_prefix.end(), tokens_prompt.begin(), tokens_prompt.end());
tokens_suffix.insert(tokens_suffix.begin(), llama_token_fim_suf(model));
auto embd_inp = spm_infill ? tokens_suffix : tokens_prefix;
auto embd_end = spm_infill ? tokens_prefix : tokens_suffix;
if (llama_add_bos_token(model)) {
embd_inp.insert(embd_inp.begin(), llama_token_bos(model));
}
SRV_DBG("extra: n_ctx = %d, n_extra_take = %d, n_extra = %d\n", n_ctx, n_extra_take, (int) extra_tokens.size());
// put the extra context before the FIM prefix
embd_inp.insert(embd_inp.begin(), extra_tokens.end() - n_extra_take, extra_tokens.end());
embd_inp.insert(embd_inp.end(), embd_end.begin(), embd_end.end());
embd_inp.push_back(llama_token_fim_mid(model));
return embd_inp;
}
// Format given chat. If tmpl is empty, we take the template from model metadata
inline std::string format_chat(const struct llama_model * model, const std::string & tmpl, const std::vector<json> & messages) {
std::vector<common_chat_msg> chat;
@ -195,14 +437,14 @@ static std::string gen_chatcmplid() {
// other common utils
//
static size_t common_part(const std::vector<llama_token> & a, const std::vector<llama_token> & b) {
static size_t longest_common_prefix(const std::vector<llama_token> & a, const std::vector<llama_token> & b) {
size_t i;
for (i = 0; i < a.size() && i < b.size() && a[i] == b[i]; i++) {}
return i;
}
static size_t common_part(const std::string & a, const std::string & b) {
static size_t longest_common_prefix(const std::string & a, const std::string & b) {
size_t i;
for (i = 0; i < a.size() && i < b.size() && a[i] == b[i]; i++) {}
@ -229,18 +471,6 @@ static size_t find_partial_stop_string(const std::string &stop, const std::strin
return std::string::npos;
}
static bool json_is_array_of_numbers(const json & data) {
if (data.is_array()) {
for (const auto & e : data) {
if (!e.is_number()) {
return false;
}
}
return true;
}
return false;
}
// TODO: reuse llama_detokenize
template <class Iter>
static std::string tokens_to_str(llama_context * ctx, Iter begin, Iter end) {
@ -360,9 +590,9 @@ static json oaicompat_completion_params_parse(
// Handle "logprobs" field
// TODO: The response format of this option is not yet OAI-compatible, but seems like no one really using it; We may need to fix it in the future
if (body.contains("logprobs")) {
if (json_value(body, "logprobs", false)) {
llama_params["n_probs"] = json_value(body, "top_logprobs", 20);
} else if (body.contains("top_logprobs")) {
} else if (body.contains("top_logprobs") && !body.at("top_logprobs").is_null()) {
throw std::runtime_error("top_logprobs requires logprobs to be set to true");
}

4
examples/simple/simple.cpp
View File

@ -138,7 +138,7 @@ int main(int argc, char ** argv) {
// prepare a batch for the prompt
llama_batch batch = llama_batch_get_one(prompt_tokens.data(), prompt_tokens.size(), 0, 0);
llama_batch batch = llama_batch_get_one(prompt_tokens.data(), prompt_tokens.size());
// main loop
@ -175,7 +175,7 @@ int main(int argc, char ** argv) {
fflush(stdout);
// prepare the next batch with the sampled token
batch = llama_batch_get_one(&new_token_id, 1, n_pos, 0);
batch = llama_batch_get_one(&new_token_id, 1);
n_decode += 1;
}

20
examples/speculative/speculative.cpp
View File

@ -39,6 +39,11 @@ int main(int argc, char ** argv) {
return 1;
}
if (params.n_predict < -1) {
LOG_ERR("%s: --n-predict must be >= -1\n", __func__);
return 1;
}
common_init();
if (params.model_draft.empty()) {
@ -155,9 +160,9 @@ int main(int argc, char ** argv) {
const auto t_enc_start = ggml_time_us();
// eval the prompt with both models
llama_decode(ctx_tgt, llama_batch_get_one( inp.data(), n_input - 1, 0, 0));
llama_decode(ctx_tgt, llama_batch_get_one(&inp.back(), 1, n_input - 1, 0));
llama_decode(ctx_dft, llama_batch_get_one( inp.data(), n_input, 0, 0));
llama_decode(ctx_tgt, llama_batch_get_one( inp.data(), n_input - 1));
llama_decode(ctx_tgt, llama_batch_get_one(&inp.back(), 1));
llama_decode(ctx_dft, llama_batch_get_one( inp.data(), n_input));
const auto t_enc_end = ggml_time_us();
@ -180,8 +185,6 @@ int main(int argc, char ** argv) {
// target model sampling context (reuse the llama_context's sampling instance)
struct common_sampler * smpl = common_sampler_init(model_tgt, params.sparams);
struct llama_sampler * softmax = llama_sampler_init_softmax();
// draft sequence data
std::vector<seq_draft> drafts(n_seq_dft);
@ -190,8 +193,8 @@ int main(int argc, char ** argv) {
drafts[s].smpl = common_sampler_init(model_dft, params.sparams);
}
llama_batch batch_dft = llama_batch_init(params.n_ctx, 0, 1);
llama_batch batch_tgt = llama_batch_init(params.n_ctx, 0, n_seq_dft);
llama_batch batch_dft = llama_batch_init(llama_n_batch(ctx_dft), 0, 1);
llama_batch batch_tgt = llama_batch_init(llama_n_batch(ctx_tgt), 0, n_seq_dft);
const auto t_dec_start = ggml_time_us();
@ -441,7 +444,7 @@ int main(int argc, char ** argv) {
++n_past_dft;
}
if (n_predict > params.n_predict || has_eos) {
if ((params.n_predict >= 0 && n_predict > params.n_predict) || has_eos) {
break;
}
@ -624,7 +627,6 @@ int main(int argc, char ** argv) {
common_sampler_free(drafts[s].smpl);
}
llama_sampler_free(softmax);
llama_batch_free(batch_dft);
llama_free(ctx_tgt);

6
flake.lock
View File

@ -20,11 +20,11 @@
},
"nixpkgs": {
"locked": {
"lastModified": 1728018373,
"narHash": "sha256-NOiTvBbRLIOe5F6RbHaAh6++BNjsb149fGZd1T4+KBg=",
"lastModified": 1729256560,
"narHash": "sha256-/uilDXvCIEs3C9l73JTACm4quuHUsIHcns1c+cHUJwA=",
"owner": "NixOS",
"repo": "nixpkgs",
"rev": "bc947f541ae55e999ffdb4013441347d83b00feb",
"rev": "4c2fcb090b1f3e5b47eaa7bd33913b574a11e0a0",
"type": "github"
},
"original": {

4
ggml/CMakeLists.txt
View File

@ -99,6 +99,9 @@ option(GGML_AVX512 "ggml: enable AVX512" OFF)
option(GGML_AVX512_VBMI "ggml: enable AVX512-VBMI" OFF)
option(GGML_AVX512_VNNI "ggml: enable AVX512-VNNI" OFF)
option(GGML_AVX512_BF16 "ggml: enable AVX512-BF16" OFF)
option(GGML_AMX_TILE "ggml: enable AMX-TILE" OFF)
option(GGML_AMX_INT8 "ggml: enable AMX-INT8" OFF)
option(GGML_AMX_BF16 "ggml: enable AMX-BF16" OFF)
option(GGML_FMA "ggml: enable FMA" ${INS_ENB})
if (NOT MSVC)
option(GGML_F16C "ggml: enable F16C" ${INS_ENB}) # in MSVC F16C is implied with AVX2/AVX512
@ -158,6 +161,7 @@ set (GGML_METAL_MACOSX_VERSION_MIN "" CACHE STRING
set (GGML_METAL_STD "" CACHE STRING "ggml: metal standard version (-std flag)")
option(GGML_OPENMP "ggml: use OpenMP" ON)
option(GGML_RPC "ggml: use RPC" OFF)
option(GGML_AMX "ggml: use AMX" OFF)
option(GGML_SYCL "ggml: use SYCL" OFF)
option(GGML_SYCL_F16 "ggml: use 16 bit floats for sycl calculations" OFF)
set (GGML_SYCL_TARGET "INTEL" CACHE STRING

25
ggml/include/ggml-amx.h Normal file
View File

@ -0,0 +1,25 @@
#pragma once
#include "ggml.h"
#include "ggml-backend.h"
#ifdef __cplusplus
extern "C" {
#endif
// buffer_type API
GGML_API ggml_backend_buffer_type_t ggml_backend_amx_buffer_type(void);
GGML_API bool ggml_backend_is_amx(ggml_backend_t backend);
// backend API
GGML_API ggml_backend_t ggml_backend_amx_init(void);
GGML_API void ggml_backend_amx_set_n_threads(ggml_backend_t backend_amx, int n_threads);
GGML_API ggml_backend_reg_t ggml_backend_amx_reg(void);
#ifdef __cplusplus
}
#endif

2
ggml/include/ggml-cann.h
View File

@ -34,6 +34,8 @@ extern "C" {
*/
#define GGML_CANN_MAX_DEVICES 16
GGML_API ggml_backend_reg_t ggml_backend_cann_reg(void);
/**
* @brief Initializes the CANN backend for a specified device.
*

11
ggml/include/ggml-sycl.h
View File

@ -19,6 +19,8 @@ extern "C" {
// backend API
GGML_API ggml_backend_t ggml_backend_sycl_init(int device);
GGML_API bool ggml_backend_is_sycl(ggml_backend_t backend);
// devide buffer
GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int device);
@ -29,14 +31,19 @@ GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_split_buffer_type(const fl
GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_host_buffer_type(void);
GGML_API void ggml_backend_sycl_print_sycl_devices(void);
GGML_API void ggml_sycl_get_gpu_list(int *id_list, int max_len);
GGML_API void ggml_sycl_get_device_description(int device, char *description, size_t description_size);
GGML_API void ggml_backend_sycl_get_gpu_list(int *id_list, int max_len);
GGML_API void ggml_backend_sycl_get_device_description(int device,
char *description,
size_t description_size);
GGML_API int ggml_backend_sycl_get_device_count();
GGML_API void ggml_backend_sycl_get_device_memory(int device, size_t *free, size_t *total);
// SYCL doesn't support registering host memory, keep here for reference
// GGML_API bool ggml_backend_sycl_register_host_buffer(void * buffer, size_t size);
// GGML_API void ggml_backend_sycl_unregister_host_buffer(void * buffer);
GGML_API ggml_backend_reg_t ggml_backend_sycl_reg(void);
#ifdef __cplusplus
}
#endif

2
ggml/include/ggml-vulkan.h
View File

@ -24,6 +24,8 @@ GGML_API ggml_backend_buffer_type_t ggml_backend_vk_buffer_type(size_t dev_num);
// pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
GGML_API ggml_backend_buffer_type_t ggml_backend_vk_host_buffer_type(void);
GGML_API ggml_backend_reg_t ggml_backend_vk_reg(void);
#ifdef __cplusplus
}
#endif

1
ggml/include/ggml.h
View File

@ -2488,6 +2488,7 @@ extern "C" {
GGML_API int ggml_cpu_has_avx512_vbmi(void);
GGML_API int ggml_cpu_has_avx512_vnni(void);
GGML_API int ggml_cpu_has_avx512_bf16(void);
GGML_API int ggml_cpu_has_amx_int8 (void);
GGML_API int ggml_cpu_has_fma (void);
GGML_API int ggml_cpu_has_neon (void);
GGML_API int ggml_cpu_has_sve (void);

42
ggml/src/CMakeLists.txt
View File

@ -267,6 +267,26 @@ if (GGML_LLAMAFILE)
set(GGML_SOURCES_LLAMAFILE llamafile/sgemm.cpp)
endif()
if (GGML_AMX)
if (CMAKE_COMPILER_IS_GNUCC AND CMAKE_CXX_COMPILER_VERSION VERSION_GREATER 11.0)
else()
set(GGML_AMX OFF)
message(WARNING "AMX requires gcc version > 11.0. Turning off GGML_AMX.")
endif()
if (GGML_AMX)
message(STATUS "Using AMX")
list(APPEND GGML_CDEF_PUBLIC GGML_USE_AMX)
file(GLOB GGML_HEADERS_AMX "ggml-amx/*.h")
list(APPEND GGML_HEADERS_AMX "../include/ggml-amx.h")
file(GLOB GGML_SOURCES_AMX "ggml-amx/*.cpp")
list(APPEND GGML_SOURCES_AMX "ggml-amx.cpp")
endif()
endif()
if (GGML_CUDA)
cmake_minimum_required(VERSION 3.18) # for CMAKE_CUDA_ARCHITECTURES
@ -1210,6 +1230,18 @@ elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR CMAKE_GENERATOR_PLATFORM_LW
add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512BF16__>)
add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512BF16__>)
endif()
if (GGML_AMX_TILE)
add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AMX_TILE__>)
add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AMX_TILE__>)
endif()
if (GGML_AMX_INT8)
add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AMX_INT8__>)
add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AMX_INT8__>)
endif()
if (GGML_AMX_BF16)
add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AMX_BF16__>)
add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AMX_BF16__>)
endif()
elseif (GGML_AVX2)
list(APPEND ARCH_FLAGS /arch:AVX2)
elseif (GGML_AVX)
@ -1245,6 +1277,15 @@ elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR CMAKE_GENERATOR_PLATFORM_LW
if (GGML_AVX512_BF16)
list(APPEND ARCH_FLAGS -mavx512bf16)
endif()
if (GGML_AMX_TILE)
list(APPEND ARCH_FLAGS -mamx-tile)
endif()
if (GGML_AMX_INT8)
list(APPEND ARCH_FLAGS -mamx-int8)
endif()
if (GGML_AMX_BF16)
list(APPEND ARCH_FLAGS -mamx-bf16)
endif()
endif()
elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64")
message(STATUS "PowerPC detected")
@ -1370,6 +1411,7 @@ add_library(ggml
${GGML_SOURCES_ROCM} ${GGML_HEADERS_ROCM}
${GGML_SOURCES_BLAS} ${GGML_HEADERS_BLAS}
${GGML_SOURCES_LLAMAFILE} ${GGML_HEADERS_LLAMAFILE}
${GGML_SOURCES_AMX} ${GGML_HEADERS_AMX}
${GGML_SOURCES_CANN} ${GGML_HEADERS_CANN}
${GGML_SOURCES_QNN} ${GGML_HEADERS_QNN}
ggml-aarch64.c ggml-aarch64.h

36
ggml/src/ggml-alloc.c
View File

@ -14,7 +14,7 @@
//#define GGML_ALLOCATOR_DEBUG
//#define AT_PRINTF(...) fprintf(stderr, __VA_ARGS__)
//#define AT_PRINTF(...) GGML_LOG_DEBUG(__VA_ARGS__)
#define AT_PRINTF(...)
@ -89,7 +89,7 @@ void ggml_tallocr_alloc(struct ggml_tallocr * talloc, struct ggml_tensor * tenso
size = GGML_PAD(size, talloc->alignment);
if (talloc->offset + size > ggml_backend_buffer_get_size(talloc->buffer)) {
fprintf(stderr, "%s: not enough space in the buffer to allocate %s (needed %zu, available %zu)\n",
GGML_LOG_ERROR("%s: not enough space in the buffer to allocate %s (needed %zu, available %zu)\n",
__func__, tensor->name, size, ggml_backend_buffer_get_size(talloc->buffer) - talloc->offset);
GGML_ABORT("not enough space in the buffer");
}
@ -172,7 +172,7 @@ static size_t ggml_dyn_tallocr_alloc(struct ggml_dyn_tallocr * alloc, size_t siz
best_fit_block = alloc->n_free_blocks - 1;
} else {
// this should never happen
fprintf(stderr, "%s: not enough space in the buffer to allocate %zu bytes, largest block available %zu bytes\n",
GGML_LOG_ERROR("%s: not enough space in the buffer to allocate %zu bytes, largest block available %zu bytes\n",
__func__, size, max_avail);
GGML_ABORT("not enough space in the buffer");
}
@ -209,16 +209,16 @@ static size_t ggml_dyn_tallocr_alloc(struct ggml_dyn_tallocr * alloc, size_t siz
}
}
}
fprintf(stderr, "max_size = %.2f MB: tensors: ", cur_max / 1024.0 / 1024.0);
GGML_LOG_DEBUG("max_size = %.2f MB: tensors: ", cur_max / 1024.0 / 1024.0);
for (int i = 0; i < 1024; i++) {
if (alloc->allocated_tensors[i].tensor) {
fprintf(stderr, "%s [%zx-%zx] (%.2f MB) ", alloc->allocated_tensors[i].tensor->name,
GGML_LOG_DEBUG("%s [%zx-%zx] (%.2f MB) ", alloc->allocated_tensors[i].tensor->name,
alloc->allocated_tensors[i].offset,
alloc->allocated_tensors[i].offset + ggml_nbytes(alloc->allocated_tensors[i].tensor),
ggml_nbytes(alloc->allocated_tensors[i].tensor) / 1024.0 / 1024.0);
}
}
fprintf(stderr, "\n");
GGML_LOG_DEBUG("\n");
}
#endif
@ -348,7 +348,6 @@ struct tensor_alloc {
};
struct leaf_alloc {
int buffer_id;
struct tensor_alloc leaf;
};
@ -740,7 +739,6 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
for (int i = 0; i < graph->n_leafs; i++) {
struct ggml_tensor * leaf = graph->leafs[i];
struct hash_node * hn = ggml_gallocr_hash_get(galloc, leaf);
galloc->leaf_allocs[i].buffer_id = hn->buffer_id;
if (leaf->view_src || leaf->data) {
galloc->leaf_allocs[i].leaf.buffer_id = -1;
galloc->leaf_allocs[i].leaf.offset = SIZE_MAX;
@ -768,13 +766,13 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
// even if there are no tensors allocated in this buffer, we still need to allocate it to initialize views
if (new_size > cur_size || galloc->buffers[i] == NULL) {
#ifndef NDEBUG
fprintf(stderr, "%s: reallocating %s buffer from size %.02f MiB to %.02f MiB\n", __func__, ggml_backend_buft_name(galloc->bufts[i]), cur_size / 1024.0 / 1024.0, new_size / 1024.0 / 1024.0);
GGML_LOG_DEBUG("%s: reallocating %s buffer from size %.02f MiB to %.02f MiB\n", __func__, ggml_backend_buft_name(galloc->bufts[i]), cur_size / 1024.0 / 1024.0, new_size / 1024.0 / 1024.0);
#endif
ggml_backend_buffer_free(galloc->buffers[i]);
galloc->buffers[i] = ggml_backend_buft_alloc_buffer(galloc->bufts[i], new_size);
if (galloc->buffers[i] == NULL) {
fprintf(stderr, "%s: failed to allocate %s buffer of size %zu\n", __func__, ggml_backend_buft_name(galloc->bufts[i]), new_size);
GGML_LOG_ERROR("%s: failed to allocate %s buffer of size %zu\n", __func__, ggml_backend_buft_name(galloc->bufts[i]), new_size);
return false;
}
ggml_backend_buffer_set_usage(galloc->buffers[i], GGML_BACKEND_BUFFER_USAGE_COMPUTE);
@ -825,14 +823,14 @@ static bool ggml_gallocr_node_needs_realloc(ggml_gallocr_t galloc, struct ggml_t
static bool ggml_gallocr_needs_realloc(ggml_gallocr_t galloc, struct ggml_cgraph * graph) {
if (galloc->n_nodes != graph->n_nodes) {
#ifndef NDEBUG
fprintf(stderr, "%s: graph has different number of nodes\n", __func__);
GGML_LOG_DEBUG("%s: graph has different number of nodes\n", __func__);
#endif
return true;
}
if (galloc->n_leafs != graph->n_leafs) {
#ifndef NDEBUG
fprintf(stderr, "%s: graph has different number of leafs\n", __func__);
GGML_LOG_DEBUG("%s: graph has different number of leafs\n", __func__);
#endif
return true;
}
@ -843,7 +841,7 @@ static bool ggml_gallocr_needs_realloc(ggml_gallocr_t galloc, struct ggml_cgraph
if (!ggml_gallocr_node_needs_realloc(galloc, node, &node_alloc->dst)) {
#ifndef NDEBUG
fprintf(stderr, "%s: node %s is not valid\n", __func__, node->name);
GGML_LOG_DEBUG("%s: node %s is not valid\n", __func__, node->name);
#endif
return true;
}
@ -855,7 +853,7 @@ static bool ggml_gallocr_needs_realloc(ggml_gallocr_t galloc, struct ggml_cgraph
}
if (!ggml_gallocr_node_needs_realloc(galloc, src, &node_alloc->src[j])) {
#ifndef NDEBUG
fprintf(stderr, "%s: src %d (%s) of node %s is not valid\n", __func__, j, src->name, node->name);
GGML_LOG_DEBUG("%s: src %d (%s) of node %s is not valid\n", __func__, j, src->name, node->name);
#endif
return true;
}
@ -869,14 +867,14 @@ bool ggml_gallocr_alloc_graph(ggml_gallocr_t galloc, struct ggml_cgraph * graph)
if (ggml_gallocr_needs_realloc(galloc, graph)) {
if (galloc->n_buffers == 1) {
#ifndef NDEBUG
fprintf(stderr, "%s: reallocating buffers automatically\n", __func__);
GGML_LOG_DEBUG("%s: reallocating buffers automatically\n", __func__);
#endif
if (!ggml_gallocr_reserve(galloc, graph)) {
return false;
}
} else {
#ifndef NDEBUG
fprintf(stderr, "%s: cannot reallocate multi buffer graph automatically, call reserve\n", __func__);
GGML_LOG_DEBUG("%s: cannot reallocate multi buffer graph automatically, call reserve\n", __func__);
#endif
return false;
}
@ -940,7 +938,7 @@ static bool alloc_tensor_range(struct ggml_context * ctx,
ggml_backend_buffer_t buffer = ggml_backend_buft_alloc_buffer(buft, size);
if (buffer == NULL) {
#ifndef NDEBUG
fprintf(stderr, "%s: failed to allocate %s buffer of size %zu\n", __func__, ggml_backend_buft_name(buft), size);
GGML_LOG_DEBUG("%s: failed to allocate %s buffer of size %zu\n", __func__, ggml_backend_buft_name(buft), size);
#endif
for (size_t i = 0; i < *n_buffers; i++) {
ggml_backend_buffer_free((*buffers)[i]);
@ -990,7 +988,7 @@ ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_conte
}
if (this_size > max_size) {
fprintf(stderr, "%s: tensor %s is too large to fit in a %s buffer (tensor size: %zu, max buffer size: %zu)\n",
GGML_LOG_ERROR("%s: tensor %s is too large to fit in a %s buffer (tensor size: %zu, max buffer size: %zu)\n",
__func__, t->name,
ggml_backend_buft_name(buft),
this_size, max_size);
@ -1022,7 +1020,7 @@ ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_conte
if (n_buffers == 0) {
#ifndef NDEBUG
fprintf(stderr, "%s: all tensors in the context are already allocated\n", __func__);
GGML_LOG_DEBUG("%s: all tensors in the context are already allocated\n", __func__);
#endif
return NULL;
}

453
ggml/src/ggml-amx.cpp Normal file
View File

@ -0,0 +1,453 @@
#include "ggml-amx.h"
#include "ggml-amx/common.h"
#include "ggml-amx/mmq.h"
#include "ggml-backend-impl.h"
#include "ggml-impl.h"
#if defined(__gnu_linux__)
#include <sys/syscall.h>
#include <unistd.h>
#endif
#include <cstdlib>
#include <cstring>
#include <memory>
#if defined(__AMX_INT8__)
// AMX buffer interface
static const char * ggml_backend_amx_buffer_get_name(ggml_backend_buffer_t buffer) {
return "AMX";
GGML_UNUSED(buffer);
}
static void ggml_backend_amx_buffer_free_buffer(ggml_backend_buffer_t buffer) {
free(buffer->context);
}
static void * ggml_backend_amx_buffer_get_base(ggml_backend_buffer_t buffer) {
return (void *)(buffer->context);
}
static void ggml_backend_amx_buffer_memset_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) {
memset((char *)tensor->data + offset, value, size);
GGML_UNUSED(buffer);
}
static void ggml_backend_amx_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
if (qtype_has_amx_kernels(tensor->type)) {
ggml_backend_amx_convert_weight(tensor, data, offset, size);
} else {
memcpy((char *)tensor->data + offset, data, size);
}
GGML_UNUSED(buffer);
}
static void ggml_backend_amx_buffer_get_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
GGML_ASSERT(!qtype_has_amx_kernels(tensor->type));
memcpy(data, (const char *)tensor->data + offset, size);
GGML_UNUSED(buffer);
}
static bool ggml_backend_amx_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst) {
if (ggml_backend_buffer_is_host(src->buffer)) {
if (qtype_has_amx_kernels(src->type)) {
ggml_backend_amx_convert_weight(dst, src->data, 0, ggml_backend_amx_get_alloc_size(dst));
} else {
memcpy(dst->data, src->data, ggml_nbytes(src));
}
return true;
}
return false;
GGML_UNUSED(buffer);
}
static void ggml_backend_amx_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
memset(buffer->context, value, buffer->size);
}
static ggml_backend_buffer_i ggml_backend_amx_buffer_interface = {
/* .get_name = */ ggml_backend_amx_buffer_get_name,
/* .free_buffer = */ ggml_backend_amx_buffer_free_buffer,
/* .get_base = */ ggml_backend_amx_buffer_get_base,
/* .init_tensor = */ NULL, // no initialization required
/* .memset_tensor = */ ggml_backend_amx_buffer_memset_tensor,
/* .set_tensor = */ ggml_backend_amx_buffer_set_tensor,
/* .get_tensor = */ ggml_backend_amx_buffer_get_tensor,
/* .cpy_tensor = */ ggml_backend_amx_buffer_cpy_tensor,
/* .clear = */ ggml_backend_amx_buffer_clear,
/* .reset = */ NULL,
};
static const char * ggml_backend_amx_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
return "AMX";
GGML_UNUSED(buft);
}
static ggml_backend_buffer_t ggml_backend_amx_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
void * data = aligned_alloc(TENSOR_ALIGNMENT, size);
if (data == NULL) {
fprintf(stderr, "%s: failed to allocate buffer of size %zu\n", __func__, size);
return NULL;
}
return ggml_backend_buffer_init(buft, ggml_backend_amx_buffer_interface, data, size);
}
static size_t ggml_backend_amx_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
return TENSOR_ALIGNMENT;
GGML_UNUSED(buft);
}
static size_t ggml_backend_amx_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor* tensor) {
return ggml_backend_amx_get_alloc_size(tensor);
GGML_UNUSED(buft);
}
static bool ggml_backend_amx_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
return false;
GGML_UNUSED(buft);
}
ggml_backend_buffer_type_t ggml_backend_amx_buffer_type() {
static struct ggml_backend_buffer_type ggml_backend_buffer_type_amx = {
/* .iface = */ {
/* .get_name = */ ggml_backend_amx_buffer_type_get_name,
/* .alloc_buffer = */ ggml_backend_amx_buffer_type_alloc_buffer,
/* .get_alignment = */ ggml_backend_amx_buffer_type_get_alignment,
/* .get_max_size = */ NULL, // defaults to SIZE_MAX
/* .get_alloc_size = */ ggml_backend_amx_buffer_type_get_alloc_size,
/* .is_host = */ ggml_backend_amx_buffer_type_is_host,
},
/* .device = */ NULL,
/* .context = */ NULL,
};
return &ggml_backend_buffer_type_amx;
}
// backend interface
static const char * ggml_backend_amx_name(ggml_backend_t backend) {
return "AMX";
GGML_UNUSED(backend);
}
static void ggml_backend_amx_free(ggml_backend_t backend) {
ggml_backend_amx_context * ctx = (ggml_backend_amx_context *)backend->context;
delete ctx;
delete backend;
}
static ggml_backend_buffer_type_t ggml_backend_amx_get_default_buffer_type(ggml_backend_t backend) {
return ggml_backend_amx_buffer_type();
GGML_UNUSED(backend);
}
static enum ggml_status ggml_backend_amx_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
ggml_backend_amx_context * ctx = (ggml_backend_amx_context *)backend->context;
for (int i = 0; i < cgraph->n_nodes; i++) {
struct ggml_tensor * node = cgraph->nodes[i];
switch (node->op) {
case GGML_OP_MUL_MAT:
ggml_backend_amx_mul_mat(ctx, node);
break;
case GGML_OP_NONE:
case GGML_OP_RESHAPE:
case GGML_OP_VIEW:
case GGML_OP_PERMUTE:
case GGML_OP_TRANSPOSE:
break;
default:
fprintf(stderr, "%s: unsupported op %s\n", __func__, ggml_op_desc(node));
GGML_ASSERT(false);
}
}
return GGML_STATUS_SUCCESS;
GGML_UNUSED(backend);
}
static struct ggml_backend_i ggml_backend_amx_i = {
/* .get_name = */ ggml_backend_amx_name,
/* .free = */ ggml_backend_amx_free,
/* .get_default_buffer_type = */ ggml_backend_amx_get_default_buffer_type,
/* .set_tensor_async = */ NULL,
/* .get_tensor_async = */ NULL,
/* .cpy_tensor_async = */ NULL,
/* .synchronize = */ NULL,
/* .graph_plan_create = */ NULL,
/* .graph_plan_free = */ NULL,
/* .graph_plan_update = */ NULL,
/* .graph_plan_compute = */ NULL,
/* .graph_compute = */ ggml_backend_amx_graph_compute,
/* .supports_op = */ NULL,
/* .supports_buft = */ NULL,
/* .offload_op = */ NULL,
/* .event_record = */ NULL,
/* .event_wait = */ NULL,
};
static ggml_guid_t ggml_backend_amx_guid() {
static ggml_guid guid = { 0x13, 0xb8, 0xa4, 0xc4, 0xba, 0xfe, 0x51, 0x67, 0x87, 0x44, 0x55, 0x15, 0xb2, 0x35, 0x62, 0x3e };
return &guid;
}
#define ARCH_GET_XCOMP_PERM 0x1022
#define ARCH_REQ_XCOMP_PERM 0x1023
#define XFEATURE_XTILECFG 17
#define XFEATURE_XTILEDATA 18
static bool ggml_amx_init() {
#if defined(__gnu_linux__)
if (syscall(SYS_arch_prctl, ARCH_REQ_XCOMP_PERM, XFEATURE_XTILEDATA)) {
fprintf(stderr, "AMX is not ready to be used!\n");
return false;
}
return true;
#elif defined(_WIN32)
return true;
#endif
}
ggml_backend_t ggml_backend_amx_init() {
// invoke a Linux system call to request access to AMX features
ggml_amx_init();
// backend context
ggml_backend_amx_context * ctx = new ggml_backend_amx_context;
// ggml amx backend
ggml_backend_t backend = new ggml_backend {
/* .guid = */ ggml_backend_amx_guid(),
/* .interface = */ ggml_backend_amx_i,
/* .device = */ ggml_backend_reg_dev_get(ggml_backend_amx_reg(), 0),
/* .context = */ ctx,
};
return backend;
}
bool ggml_backend_is_amx(ggml_backend_t backend) {
return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_amx_guid());
}
void ggml_backend_amx_set_n_threads(ggml_backend_t backend_amx, int n_threads) {
GGML_ASSERT(ggml_backend_is_amx(backend_amx));
ggml_backend_amx_context * ctx = (ggml_backend_amx_context *)backend_amx->context;
ctx->n_threads = n_threads;
}
// device interface
static const char * ggml_backend_amx_device_get_name(ggml_backend_dev_t dev) {
return "AMX";
GGML_UNUSED(dev);
}
static const char * ggml_backend_amx_device_get_description(ggml_backend_dev_t dev) {
return "Intel Advanced Matrix Extensions";
GGML_UNUSED(dev);
}
static void ggml_backend_amx_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
// TODO
*free = 0;
*total = 0;
GGML_UNUSED(dev);
}
static enum ggml_backend_dev_type ggml_backend_amx_device_get_type(ggml_backend_dev_t dev) {
return GGML_BACKEND_DEVICE_TYPE_CPU;
GGML_UNUSED(dev);
}
static void ggml_backend_amx_device_get_props(ggml_backend_dev_t dev, struct ggml_backend_dev_props * props) {
props->name = ggml_backend_amx_device_get_name(dev);
props->description = ggml_backend_amx_device_get_description(dev);
props->type = ggml_backend_amx_device_get_type(dev);
ggml_backend_amx_device_get_memory(dev, &props->memory_free, &props->memory_total);
// `buffer_from_host_ptr` is intended to be used in mmap, when memory layout unchanged
props->caps = {
/* .async = */ false,
/* .host_buffer = */ false,
/* .buffer_from_host_ptr = */ false,
/* .events = */ false,
};
}
static ggml_backend_t ggml_backend_amx_device_init(ggml_backend_dev_t dev, const char * params) {
return ggml_backend_amx_init();
GGML_UNUSED(dev);
GGML_UNUSED(params);
}
static ggml_backend_buffer_type_t ggml_backend_amx_device_get_buffer_type(ggml_backend_dev_t dev) {
return ggml_backend_amx_buffer_type();
GGML_UNUSED(dev);
}
static bool ggml_backend_amx_device_supports_op(ggml_backend_dev_t dev, const struct ggml_tensor * op) {
// handle only 2d gemm for now
auto is_contiguous_2d = [](const struct ggml_tensor * t) {
return ggml_is_contiguous(t) && t->ne[3] == 1 && t->ne[2] == 1;
};
switch (op->op) {
case GGML_OP_NONE:
case GGML_OP_RESHAPE:
case GGML_OP_VIEW:
case GGML_OP_PERMUTE:
case GGML_OP_TRANSPOSE:
return true;
case GGML_OP_MUL_MAT: {
const struct ggml_tensor * src0 = op->src[0];
const struct ggml_tensor * src1 = op->src[1];
const enum ggml_type type = src0->type;
const int64_t ne0 = op->ne[0];
bool is_training = src0->grad || src1->grad;
// amx kernels enables for Q4_0, Q4_1, Q8_0, F16
// Q4_K, Q5_K, Q6_K, IQ4_XS enabled for QK_K = 256
bool has_amx_kernels = qtype_has_amx_kernels(type) || (type == GGML_TYPE_F16);
bool can_use_amx =
is_contiguous_2d(src0) && // src0 must be contiguous
is_contiguous_2d(src1) && // src1 must be contiguous
!is_training && // inference only
src1->type == GGML_TYPE_F32 && // src1 must be float32
has_amx_kernels && // with amx kernel impls
ne0 % (TILE_N * 2) == 0; // out_features is 32x
return can_use_amx;
}
default:
return false;
}
GGML_UNUSED(dev);
}
static bool ggml_backend_amx_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
return buft->iface.get_name == ggml_backend_amx_buffer_type_get_name;
GGML_UNUSED(dev);
}
static const struct ggml_backend_device_i ggml_backend_amx_device_i = {
/* .get_name = */ ggml_backend_amx_device_get_name,
/* .get_description = */ ggml_backend_amx_device_get_description,
/* .get_memory = */ ggml_backend_amx_device_get_memory,
/* .get_type = */ ggml_backend_amx_device_get_type,
/* .get_props = */ ggml_backend_amx_device_get_props,
/* .init_backend = */ ggml_backend_amx_device_init,
/* .get_buffer_type = */ ggml_backend_amx_device_get_buffer_type,
/* .get_host_buffer_type = */ NULL,
/* .buffer_from_host_ptr = */ NULL,
/* .supports_op = */ ggml_backend_amx_device_supports_op,
/* .supports_buft = */ ggml_backend_amx_device_supports_buft,
/* .offload_op = */ NULL,
/* .event_new = */ NULL,
/* .event_free = */ NULL,
/* .event_synchronize = */ NULL,
};
// backend reg interface
static const char * ggml_backend_amx_reg_get_name(ggml_backend_reg_t reg) {
return "AMX";
GGML_UNUSED(reg);
}
static size_t ggml_backend_amx_reg_get_device_count(ggml_backend_reg_t reg) {
return 1;
GGML_UNUSED(reg);
}
static ggml_backend_dev_t ggml_backend_amx_reg_get_device(ggml_backend_reg_t reg, size_t index) {
GGML_ASSERT(index == 0);
static ggml_backend_device ggml_backend_amx_device = {
/* .iface = */ ggml_backend_amx_device_i,
/* .reg = */ reg,
/* .context = */ nullptr,
};
return &ggml_backend_amx_device;
GGML_UNUSED(reg);
GGML_UNUSED(index);
}
static void * ggml_backend_amx_get_proc_address(ggml_backend_reg_t reg, const char * name) {
if (std::strcmp(name, "ggml_backend_set_n_threads") == 0) {
return (void *)ggml_backend_amx_set_n_threads;
}
return NULL;
GGML_UNUSED(reg);
GGML_UNUSED(name);
}
static const struct ggml_backend_reg_i ggml_backend_amx_reg_i = {
/* .get_name = */ ggml_backend_amx_reg_get_name,
/* .get_device_count = */ ggml_backend_amx_reg_get_device_count,
/* .get_device = */ ggml_backend_amx_reg_get_device,
/* .get_proc_address = */ ggml_backend_amx_get_proc_address,
};
ggml_backend_reg_t ggml_backend_amx_reg(void) {
static struct ggml_backend_reg ggml_backend_amx_reg = {
/* .iface = */ ggml_backend_amx_reg_i,
/* .context = */ NULL,
};
return &ggml_backend_amx_reg;
}
#else // if defined(__AMX_INT8__)
ggml_backend_t ggml_backend_amx_init(void) {
fprintf(stderr, "GGML is not compiled with AMX support!\n");
return ggml_backend_t{};
}
void ggml_backend_amx_set_n_threads(ggml_backend_t backend_amx, int n_threads) {
fprintf(stderr, "GGML is not compiled with AMX support!\n");
GGML_UNUSED(backend_amx);
GGML_UNUSED(n_threads);
}
#endif

93
ggml/src/ggml-amx/common.h Normal file
View File

@ -0,0 +1,93 @@
#pragma once
#include "ggml.h"
#include "ggml-cpu-impl.h" // <immintrin.h>
#include <algorithm>
#include <memory>
#include <type_traits>
#if defined(_OPENMP)
#include <omp.h>
#endif
#define TILE_M 16
#define TILE_N 16
#define TILE_K 32
#define VNNI_BLK 4
#define AMX_BLK_SIZE 32
#define TMM0 0
#define TMM1 1
#define TMM2 2
#define TMM3 3
#define TMM4 4
#define TMM5 5
#define TMM6 6
#define TMM7 7
// parallel routines
template <typename T, typename std::enable_if<std::is_integral<T>::value, int>::type = 0>
inline T div_up(T x, T y) { return (x + y - 1) / y; }
template <typename T>
inline void balance211(T n, T nth, T ith, T& n_start, T& n_end) {
#if 0
// onednn partition pattern
T& n_my = n_end;
if (nth <= 1 || n == 0) {
n_start = 0;
n_my = n;
} else {
T n1 = div_up(n, nth);
T n2 = n1 - 1;
T T1 = n - n2 * nth;
n_my = ith < T1 ? n1 : n2;
n_start = ith <= T1 ? ith*n1 : T1 * n1 + (ith - T1) * n2;
}
n_end += n_start;
#else
// pytorch aten partition pattern
T n_my = div_up(n, nth);
n_start = ith * n_my;
n_end = std::min(n_start + n_my, n);
#endif
}
template <typename func_t>
inline void parallel_for(int nth, int n, const func_t& f) {
#if defined(_OPENMP)
#pragma omp parallel num_threads(nth)
{
//int nth = omp_get_num_threads();
int ith = omp_get_thread_num();
int tbegin, tend;
balance211(n, nth, ith, tbegin, tend);
f(tbegin, tend);
}
#else
f(0, n);
GGML_UNUSED(nth);
#endif
}
// quantized types that have AMX support
inline bool qtype_has_amx_kernels(const enum ggml_type type) {
// TODO: fix padding for vnni format
return (type == GGML_TYPE_Q4_0) ||
(type == GGML_TYPE_Q4_1);
//(type == GGML_TYPE_Q8_0) ||
//(type == GGML_TYPE_Q4_K) ||
//(type == GGML_TYPE_Q5_K) ||
//(type == GGML_TYPE_Q6_K) ||
//(type == GGML_TYPE_IQ4_XS);
}
// ggml backend context
struct ggml_backend_amx_context {
int n_threads = GGML_DEFAULT_N_THREADS;
std::unique_ptr<char[]> work_data;
size_t work_size = 0;
};

2509
ggml/src/ggml-amx/mmq.cpp Normal file
View File

File diff suppressed because it is too large Load Diff

17
ggml/src/ggml-amx/mmq.h Normal file
View File

@ -0,0 +1,17 @@
#pragma once
#include "common.h"
#include <stdint.h>
#ifdef __cplusplus
extern "C" {
#endif
size_t ggml_backend_amx_get_alloc_size(const struct ggml_tensor * tensor);
void ggml_backend_amx_convert_weight(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
void ggml_backend_amx_mul_mat(ggml_backend_amx_context * ctx, struct ggml_tensor * dst);
#ifdef __cplusplus
}
#endif

83
ggml/src/ggml-backend.cpp
View File

@ -329,7 +329,6 @@ bool ggml_backend_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type
if (backend->device) {
return ggml_backend_dev_supports_buft(backend->device, buft);
}
return backend->iface.supports_buft(backend, buft);
}
@ -379,7 +378,7 @@ void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst
ggml_backend_tensor_get(src, dst->data, 0, ggml_nbytes(src));
} else if (!ggml_backend_buffer_copy_tensor(src, dst)) {
#ifndef NDEBUG
fprintf(stderr, "%s: warning: slow copy from %s to %s\n", __func__, ggml_backend_buffer_name(src->buffer), ggml_backend_buffer_name(dst->buffer));
GGML_LOG_DEBUG("%s: warning: slow copy from %s to %s\n", __func__, ggml_backend_buffer_name(src->buffer), ggml_backend_buffer_name(dst->buffer));
#endif
size_t nbytes = ggml_nbytes(src);
void * data = malloc(nbytes);
@ -538,6 +537,14 @@ void * ggml_backend_reg_get_proc_address(ggml_backend_reg_t reg, const char * na
#include "ggml-metal.h"
#endif
#ifdef GGML_USE_SYCL
#include "ggml-sycl.h"
#endif
#ifdef GGML_USE_VULKAN
#include "ggml-vulkan.h"
#endif
#ifdef GGML_USE_BLAS
#include "ggml-blas.h"
#endif
@ -546,6 +553,18 @@ void * ggml_backend_reg_get_proc_address(ggml_backend_reg_t reg, const char * na
#include "ggml-rpc.h"
#endif
#ifndef __AMX_INT8__
#undef GGML_USE_AMX
#endif
#ifdef GGML_USE_AMX
# include "ggml-amx.h"
#endif
#ifdef GGML_USE_CANN
#include "ggml-cann.h"
#endif
#ifdef GGML_USE_QNN
#include "ggml-qnn.h"
#endif
@ -561,14 +580,26 @@ struct ggml_backend_registry {
#ifdef GGML_USE_METAL
register_backend(ggml_backend_metal_reg());
#endif
#ifdef GGML_USE_SYCL
register_backend(ggml_backend_sycl_reg());
#endif
#ifdef GGML_USE_VULKAN
register_backend(ggml_backend_vk_reg());
#endif
#ifdef GGML_USE_BLAS
register_backend(ggml_backend_blas_reg());
#endif
#ifdef GGML_USE_RPC
register_backend(ggml_backend_rpc_reg());
#endif
#ifdef GGML_USE_AMX
register_backend(ggml_backend_amx_reg());
#endif
#ifdef GGML_USE_CANN
register_backend(ggml_backend_cann_reg());
#endif
// TODO: sycl, vulkan, kompute, cann
// TODO: kompute
#ifdef GGML_USE_QNN
register_backend(ggml_backend_qnn_reg());
@ -579,7 +610,7 @@ struct ggml_backend_registry {
void register_backend(ggml_backend_reg_t reg) {
#ifndef NDEBUG
fprintf(stderr, "%s: registered backend %s (%zu devices)\n",
GGML_LOG_DEBUG("%s: registered backend %s (%zu devices)\n",
__func__, ggml_backend_reg_name(reg), ggml_backend_reg_dev_count(reg));
#endif
backends.push_back(reg);
@ -590,7 +621,7 @@ struct ggml_backend_registry {
void register_device(ggml_backend_dev_t device) {
#ifndef NDEBUG
fprintf(stderr, "%s: registered device %s (%s)\n", __func__, ggml_backend_dev_name(device), ggml_backend_dev_description(device));
GGML_LOG_DEBUG("%s: registered device %s (%s)\n", __func__, ggml_backend_dev_name(device), ggml_backend_dev_description(device));
#endif
devices.push_back(device);
}
@ -690,8 +721,6 @@ ggml_backend_t ggml_backend_init_best(void) {
// backend CPU
static const size_t TENSOR_ALIGNMENT = 32; // required for mmap as gguf only guarantees 32-byte alignment
static const char * ggml_backend_cpu_buffer_get_name(ggml_backend_buffer_t buffer) {
return "CPU";
@ -710,7 +739,7 @@ static void * ggml_backend_cpu_buffer_get_base(ggml_backend_buffer_t buffer) {
}
static void ggml_backend_cpu_buffer_free_buffer(ggml_backend_buffer_t buffer) {
free(buffer->context);
ggml_aligned_free(buffer->context, buffer->size);
}
static void ggml_backend_cpu_buffer_memset_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) {
@ -778,14 +807,19 @@ static const char * ggml_backend_cpu_buffer_type_get_name(ggml_backend_buffer_ty
}
static ggml_backend_buffer_t ggml_backend_cpu_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
size += TENSOR_ALIGNMENT; // malloc may return an address that is not aligned
void * data = malloc(size); // TODO: use GGML_ALIGNED_MALLOC (move to ggml-impl.h)
auto alloc_size = size;
if (alloc_size == 0) {
alloc_size = 1;
}
void * data = ggml_aligned_malloc(alloc_size);
if (data == NULL) {
fprintf(stderr, "%s: failed to allocate buffer of size %zu\n", __func__, size);
GGML_LOG_ERROR("%s: failed to allocate buffer of size %zu\n", __func__, alloc_size);
return NULL;
}
return ggml_backend_buffer_init(buft, ggml_backend_cpu_buffer_i, data, size);
return ggml_backend_buffer_init(buft, ggml_backend_cpu_buffer_i, data, alloc_size);
}
static size_t ggml_backend_cpu_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
@ -844,7 +878,7 @@ static ggml_backend_buffer_t ggml_backend_cpu_hbm_buffer_type_alloc_buffer(ggml_
void * ptr;
int result = hbw_posix_memalign(&ptr, ggml_backend_cpu_buffer_type_get_alignment(buft), size);
if (result != 0) {
fprintf(stderr, "failed to allocate HBM buffer of size %zu\n", size);
GGML_LOG_ERROR("failed to allocate HBM buffer of size %zu\n", size);
return NULL;
}
@ -1467,7 +1501,7 @@ static int ggml_backend_sched_backend_from_buffer(ggml_backend_sched_t sched, co
}
#ifndef NDEBUG
fprintf(stderr, "%s: warning: no backend supports op %s with a weight with buffer type %s used in tensor %s, the weight will need to be copied\n",
GGML_LOG_DEBUG("%s: warning: no backend supports op %s with a weight with buffer type %s used in tensor %s, the weight will need to be copied\n",
__func__, ggml_op_desc(tensor), ggml_backend_buffer_name(buffer), tensor->name);
#endif
@ -1556,13 +1590,13 @@ static void ggml_backend_sched_print_assignments(ggml_backend_sched_t sched, str
for (int i = 0; i < graph->n_nodes; i++) {
if (cur_split < sched->n_splits && i == sched->splits[cur_split].i_start) {
ggml_backend_t split_backend = sched->backends[sched->splits[cur_split].backend_id];
fprintf(stderr, "\n## SPLIT #%d: %s # %d inputs: ", cur_split, ggml_backend_name(split_backend),
GGML_LOG_DEBUG("\n## SPLIT #%d: %s # %d inputs: ", cur_split, ggml_backend_name(split_backend),
sched->splits[cur_split].n_inputs);
for (int j = 0; j < sched->splits[cur_split].n_inputs; j++) {
fprintf(stderr, "[%s (%5.5s)] ", sched->splits[cur_split].inputs[j]->name,
GGML_LOG_DEBUG("[%s (%5.5s)] ", sched->splits[cur_split].inputs[j]->name,
fmt_size(ggml_nbytes(sched->splits[cur_split].inputs[j])));
}
fprintf(stderr, "\n");
GGML_LOG_DEBUG("\n");
cur_split++;
}
struct ggml_tensor * node = graph->nodes[i];
@ -1570,7 +1604,7 @@ static void ggml_backend_sched_print_assignments(ggml_backend_sched_t sched, str
continue;
}
ggml_backend_t tensor_backend = ggml_backend_sched_get_tensor_backend(sched, node);
fprintf(stderr, "node #%3d (%10.10s): %20.20s (%5.5s) [%5.5s %8.8s]:", i, ggml_op_name(node->op), node->name,
GGML_LOG_DEBUG("node #%3d (%10.10s): %20.20s (%5.5s) [%5.5s %8.8s]:", i, ggml_op_name(node->op), node->name,
fmt_size(ggml_nbytes(node)), tensor_backend ? ggml_backend_name(tensor_backend) : "NULL", GET_CAUSE(node));
for (int j = 0; j < GGML_MAX_SRC; j++) {
struct ggml_tensor * src = node->src[j];
@ -1578,10 +1612,10 @@ static void ggml_backend_sched_print_assignments(ggml_backend_sched_t sched, str
continue;
}
ggml_backend_t src_backend = ggml_backend_sched_get_tensor_backend(sched, src);
fprintf(stderr, " %20.20s (%5.5s) [%5.5s %8.8s]", src->name,
GGML_LOG_DEBUG(" %20.20s (%5.5s) [%5.5s %8.8s]", src->name,
fmt_size(ggml_nbytes(src)), src_backend ? ggml_backend_name(src_backend) : "NULL", GET_CAUSE(src));
}
fprintf(stderr, "\n");
GGML_LOG_DEBUG("\n");
}
}
@ -2095,11 +2129,11 @@ static bool ggml_backend_sched_alloc_splits(ggml_backend_sched_t sched) {
// the re-allocation may cause the split inputs to be moved to a different address
ggml_backend_sched_synchronize(sched);
#ifndef NDEBUG
fprintf(stderr, "%s: failed to allocate graph, reserving (backend_ids_changed = %d)\n", __func__, backend_ids_changed);
GGML_LOG_DEBUG("%s: failed to allocate graph, reserving (backend_ids_changed = %d)\n", __func__, backend_ids_changed);
#endif
ggml_gallocr_reserve_n(sched->galloc, &sched->graph, sched->node_backend_ids, sched->leaf_backend_ids);
if (!ggml_gallocr_alloc_graph(sched->galloc, &sched->graph)) {
fprintf(stderr, "%s: failed to allocate graph\n", __func__);
GGML_LOG_ERROR("%s: failed to allocate graph\n", __func__);
return false;
}
}
@ -2242,6 +2276,7 @@ ggml_backend_sched_t ggml_backend_sched_new(
sched->backends[b] = backends[b];
sched->bufts[b] = bufts ? bufts[b] : ggml_backend_get_default_buffer_type(backends[b]);
GGML_ASSERT(ggml_backend_supports_buft(backends[b], sched->bufts[b]));
if (sched->n_copies > 1) {
for (int c = 0; c < sched->n_copies; c++) {
sched->events[b][c] = ggml_backend_event_new(backends[b]->device);
@ -2493,7 +2528,7 @@ struct ggml_backend_graph_copy ggml_backend_graph_copy(ggml_backend_t backend, s
struct ggml_context * ctx_unallocated = ggml_init(params);
if (ctx_allocated == NULL || ctx_unallocated == NULL) {
fprintf(stderr, "failed to allocate context for graph copy\n");
GGML_LOG_ERROR("%s: failed to allocate context for graph copy\n", __func__);
ggml_hash_set_free(&hash_set);
free(node_copies);
free(node_init);
@ -2516,7 +2551,7 @@ struct ggml_backend_graph_copy ggml_backend_graph_copy(ggml_backend_t backend, s
// allocate nodes
ggml_backend_buffer_t buffer = ggml_backend_alloc_ctx_tensors(ctx_allocated, backend);
if (buffer == NULL) {
fprintf(stderr, "failed to allocate buffer for graph copy\n");
GGML_LOG_ERROR("%s: failed to allocate buffer for graph copy\n", __func__);
ggml_hash_set_free(&hash_set);
free(node_copies);
free(node_init);

8
ggml/src/ggml-blas.cpp
View File

@ -297,14 +297,14 @@ ggml_backend_t ggml_backend_blas_init(void) {
/* .context = */ ctx,
};
#if !defined(NDEBUG) && defined(OPENBLAS_VERSION) && defined(GGML_USE_OPENMP)
#if defined(OPENBLAS_VERSION) && defined(GGML_USE_OPENMP)
if (openblas_get_parallel() != OPENBLAS_OPENMP) {
fprintf(stderr, "%s: warning: ggml is using OpenMP, but OpenBLAS was compiled without OpenMP support\n", __func__);
GGML_LOG_DEBUG("%s: warning: ggml is using OpenMP, but OpenBLAS was compiled without OpenMP support\n", __func__);
}
#endif
#if !defined(NDEBUG) && defined(BLIS_ENABLE_CBLAS) && defined(GGML_USE_OPENMP) && !defined(BLIS_ENABLE_OPENMP)
fprintf(stderr, "%s: warning: ggml is using OpenMP, but BLIS was compiled without OpenMP support\n", __func__);
#if defined(BLIS_ENABLE_CBLAS) && defined(GGML_USE_OPENMP) && !defined(BLIS_ENABLE_OPENMP)
GGML_LOG_DEBUG("%s: warning: ggml is using OpenMP, but BLIS was compiled without OpenMP support\n", __func__);
#endif
return backend;

361
ggml/src/ggml-cann.cpp
View File

@ -39,6 +39,8 @@
#include "ggml-common.h"
#define GGML_CANN_NAME "CANN"
/**
* @brief Handles CANN errors by printing an error message and aborting.
*
@ -851,13 +853,6 @@ static void ggml_backend_cann_buffer_set_tensor(
void *transform_buffer = malloc(size);
ggml_backend_cann_transform(tensor, data, transform_buffer);
#ifndef NDEBUG
void *check_buffer = malloc(size);
ggml_backend_cann_transform_back(tensor, transform_buffer,
check_buffer);
GGML_ASSERT(memcmp(data, check_buffer, size) == 0);
free(check_buffer);
#endif
ACL_CHECK(aclrtMemcpy((char *)tensor->data + offset, size,
transform_buffer, size,
ACL_MEMCPY_HOST_TO_DEVICE));
@ -969,7 +964,7 @@ static void ggml_backend_cann_buffer_clear(
* This structure defines function pointers to operations that can be performed
* on a CANN buffer within the backend.
*/
static ggml_backend_buffer_i ggml_backend_cann_buffer_interface = {
static const ggml_backend_buffer_i ggml_backend_cann_buffer_interface = {
/* .get_name = */ ggml_backend_cann_buffer_get_name,
/* .free_buffer = */ ggml_backend_cann_buffer_free_buffer,
/* .get_base = */ ggml_backend_cann_buffer_get_base,
@ -1105,19 +1100,25 @@ static size_t ggml_backend_cann_buffer_type_get_alloc_size(
GGML_UNUSED(buft);
}
static bool ggml_backend_cann_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
return false;
GGML_UNUSED(buft);
}
/**
* @brief Interface for managing CANN buffer types in the GGML backend.
*
* Provides function pointers for allocating, querying properties, and managing
* memory for CANN buffer types in the GGML backend.
*/
static ggml_backend_buffer_type_i ggml_backend_cann_buffer_type_interface = {
static const ggml_backend_buffer_type_i ggml_backend_cann_buffer_type_interface = {
/* .get_name = */ ggml_backend_cann_buffer_type_name,
/* .alloc_buffer = */ ggml_backend_cann_buffer_type_alloc_buffer,
/* .get_alignment = */ ggml_backend_cann_buffer_type_get_alignment,
/* .get_max_size = */ NULL, // defaults to SIZE_MAX
/* .get_alloc_size = */ ggml_backend_cann_buffer_type_get_alloc_size,
/* .is_host = */ NULL,
/* .is_host = */ ggml_backend_cann_buffer_type_is_host,
};
/**
@ -1148,6 +1149,7 @@ ggml_backend_cann_buffer_type(int32_t device) {
for (int32_t i = 0; i < GGML_CANN_MAX_DEVICES; i++) {
ggml_backend_cann_buffer_types[i] = {
/* .iface = */ ggml_backend_cann_buffer_type_interface,
/* .device = */ ggml_backend_reg_dev_get(ggml_backend_cann_reg(), device),
/* .context = */
new ggml_backend_cann_buffer_type_context{
i, "CANN" + std::to_string(i)},
@ -1263,7 +1265,7 @@ ggml_backend_buffer_type_t ggml_backend_cann_host_buffer_type() {
/* .get_alloc_size = */ ggml_backend_cpu_buffer_type()->iface.get_alloc_size,
/* .is_host = */ ggml_backend_cpu_buffer_type()->iface.is_host,
},
/* .device = */ nullptr,
/* .device = */ ggml_backend_reg_dev_get(ggml_backend_cann_reg(), 0),
/* .context = */ nullptr,
};
@ -1510,13 +1512,6 @@ static void ggml_backend_cann_set_tensor_async(ggml_backend_t backend,
void *transform_buffer = malloc(size);
ggml_backend_cann_transform(tensor, data, transform_buffer);
#ifndef NDEBUG
void *check_buffer = malloc(size);
ggml_backend_cann_transform_back(tensor, transform_buffer,
check_buffer);
GGML_ASSERT(memcmp(data, check_buffer, size));
free(check_buffer);
#endif
ACL_CHECK(aclrtMemcpyAsync(
(char *)tensor->data + offset, size, transform_buffer, size,
ACL_MEMCPY_HOST_TO_DEVICE, cann_ctx->stream()));
@ -1691,7 +1686,7 @@ static enum ggml_status ggml_backend_cann_graph_compute(
* @return bool Returns true if the operation is supported by the backend,
* otherwise false.
*/
static bool ggml_backend_cann_supports_op(ggml_backend_t backend,
static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
const ggml_tensor* op) {
switch (op->op) {
case GGML_OP_UNARY:
@ -1782,7 +1777,7 @@ static bool ggml_backend_cann_supports_op(ggml_backend_t backend,
return false;
}
GGML_UNUSED(backend);
GGML_UNUSED(dev);
}
/**
@ -1800,31 +1795,6 @@ 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 Checks if the CANN backend supports a specific backend buffer type.
*
* This function determines whether the CANN backend supports the given backend
* buffer type by comparing the device context of the backend and buffer type.
* It returns true if the devices are same between the backend context and
* buffer type context.
*
* @param backend Pointer to the CANN backend.
* @param buft Pointer to the backend buffer type to check.
* @return bool Returns true if the CANN backend supports the buffer type,
* otherwise false.
*/
static bool ggml_backend_cann_supports_buft(
ggml_backend_t backend, ggml_backend_buffer_type_t buft) {
if (ggml_backend_buft_is_cann(buft)) {
ggml_backend_cann_context * cann_ctx =
(ggml_backend_cann_context *)backend->context;
ggml_backend_cann_buffer_type_context * buft_ctx =
(ggml_backend_cann_buffer_type_context *)buft->context;
return buft_ctx->device == cann_ctx->device;
}
return false;
}
/**
* @brief Determines if a tensor operation should be offloaded to the CANN
* backend.
@ -1839,54 +1809,14 @@ static bool ggml_backend_cann_supports_buft(
* @return bool Returns true if the operation should be offloaded, otherwise
* false.
*/
static bool ggml_backend_cann_offload_op(ggml_backend_t backend,
static bool ggml_backend_cann_offload_op(ggml_backend_dev_t dev,
const ggml_tensor* op) {
const int min_batch_size = 32;
GGML_UNUSED(backend);
GGML_UNUSED(dev);
return op->ne[1] >= min_batch_size && op->op != GGML_OP_GET_ROWS;
}
/**
* @brief Creates a new event for the CANN backend.
*
* This function initializes a new event for the CANN backend by setting the
* device and creating an ACL runtime event. The created event is then wrapped
* in a ggml_backend_event structure and returned.
*
* @param backend Pointer to the CANN backend.
* @return ggml_backend_event_t Returns a pointer to the new event structure.
*/
static ggml_backend_event_t ggml_backend_cann_event_new(
ggml_backend_t backend) {
ggml_backend_cann_context* cann_ctx =
(ggml_backend_cann_context*)backend->context;
ggml_cann_set_device(cann_ctx->device);
aclrtEvent event;
ACL_CHECK(aclrtCreateEvent(&event));
return new ggml_backend_event{
/* .backend = */ backend,
/* .context = */ event,
};
}
/**
* @brief Frees a CANN backend event.
*
* This function destroys the ACL runtime event associated with the given CANN
* backend event and then deletes the event structure itself.
*
* @param event Pointer to the event structure to be freed.
*/
static void ggml_backend_cann_event_free(ggml_backend_event_t event) {
ACL_CHECK(aclrtDestroyEvent((aclrtEvent)event->context));
delete event;
}
/**
* @brief Records an event on the CANN backend stream.
*
@ -1895,10 +1825,9 @@ static void ggml_backend_cann_event_free(ggml_backend_event_t event) {
*
* @param event Pointer to the event structure to be recorded.
*/
static void ggml_backend_cann_event_record(ggml_backend_event_t event) {
static void ggml_backend_cann_event_record(ggml_backend_t backend, ggml_backend_event_t event) {
ggml_backend_cann_context* cann_ctx =
(ggml_backend_cann_context*)event->backend->context;
(ggml_backend_cann_context*)backend->context;
ACL_CHECK(aclrtRecordEvent((aclrtEvent)event->context, cann_ctx->stream()));
}
@ -1916,8 +1845,7 @@ static void ggml_backend_cann_event_wait(ggml_backend_t backend,
ggml_backend_event_t event) {
ggml_backend_cann_context* cann_ctx =
(ggml_backend_cann_context*)backend->context;
if (ggml_backend_is_cann(event->backend)) {
if (ggml_backend_is_cann(backend)) {
ACL_CHECK(aclrtStreamWaitEvent(cann_ctx->stream(),
(aclrtEvent)event->context));
} else {
@ -1925,17 +1853,6 @@ static void ggml_backend_cann_event_wait(ggml_backend_t backend,
}
}
/**
* @brief Synchronizes the given event on the CANN backend.
*
* This function waits for the specified event to complete on the ACL runtime.
*
* @param event Pointer to the event structure to be synchronized.
*/
static void ggml_backend_cann_event_synchronize(ggml_backend_event_t event) {
ACL_CHECK(aclrtSynchronizeEvent((aclrtEvent)event->context));
}
/**
* @brief Structure defining the interface for the CANN backend.
*
@ -1943,7 +1860,7 @@ static void ggml_backend_cann_event_synchronize(ggml_backend_event_t event) {
* supported by the CANN backend, including name retrieval, memory
* management, tensor operations, synchronization, and event handling.
*/
static ggml_backend_i ggml_backend_cann_interface = {
static const ggml_backend_i ggml_backend_cann_interface = {
/* .get_name = */ ggml_backend_cann_name,
/* .free = */ ggml_backend_cann_free,
/* .get_default_buffer_type = */ ggml_backend_cann_get_default_buffer_type,
@ -1956,9 +1873,9 @@ static ggml_backend_i ggml_backend_cann_interface = {
/* .graph_plan_update = */ NULL,
/* .graph_plan_compute = */ NULL,
/* .graph_compute = */ ggml_backend_cann_graph_compute,
/* .supports_op = */ ggml_backend_cann_supports_op,
/* .supports_buft = */ ggml_backend_cann_supports_buft,
/* .offload_op = */ ggml_backend_cann_offload_op,
/* .supports_op = */ NULL, // moved to device
/* .supports_buft = */ NULL, // moved to device
/* .offload_op = */ NULL, // moved to device
/* .event_record = */ ggml_backend_cann_event_record,
/* .event_wait = */ ggml_backend_cann_event_wait,
};
@ -1977,6 +1894,234 @@ static ggml_guid_t ggml_backend_cann_guid() {
return &guid;
}
// backend device
struct ggml_backend_cann_device_context {
int device;
std::string name;
std::string description;
};
static const char * ggml_backend_cann_device_get_name(ggml_backend_dev_t dev) {
ggml_backend_cann_device_context * ctx = (ggml_backend_cann_device_context *)dev->context;
return ctx->name.c_str();
}
static const char* ggml_backend_cann_device_get_description(ggml_backend_dev_t dev) {
ggml_backend_cann_device_context * ctx = (ggml_backend_cann_device_context *)dev->context;
return ctx->description.c_str();
}
static void ggml_backend_cann_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
ggml_backend_cann_device_context * ctx = (ggml_backend_cann_device_context *)dev->context;
ggml_backend_cann_get_device_memory(ctx->device, free, total);
}
static enum ggml_backend_dev_type ggml_backend_cann_device_get_type(ggml_backend_dev_t dev) {
GGML_UNUSED(dev);
return GGML_BACKEND_DEVICE_TYPE_GPU_FULL;
}
static void ggml_backend_cann_device_get_props(ggml_backend_dev_t dev, ggml_backend_dev_props * props) {
props->name = ggml_backend_cann_device_get_name(dev);
props->description = ggml_backend_cann_device_get_description(dev);
props->type = ggml_backend_cann_device_get_type(dev);
ggml_backend_cann_device_get_memory(dev, &props->memory_free, &props->memory_total);
bool host_buffer = getenv("GGML_CANN_NO_PINNED") == nullptr;
props->caps = {
/* .async = */ false,
/* .host_buffer = */ host_buffer,
/* .buffer_from_host_ptr = */ false,
/* .events = */ true,
};
}
static ggml_backend_t ggml_backend_cann_device_init(ggml_backend_dev_t dev, const char * params) {
GGML_UNUSED(params);
ggml_backend_cann_device_context * ctx = (ggml_backend_cann_device_context *)dev->context;
return ggml_backend_cann_init(ctx->device);
}
/**
* @brief Checks if the CANN backend supports a specific backend buffer type.
*
* This function determines whether the CANN backend supports the given backend
* buffer type by comparing the device context of the backend and buffer type.
* It returns true if the devices are same between the backend context and
* buffer type context.
*
* @param backend Pointer to the CANN backend.
* @param buft Pointer to the backend buffer type to check.
* @return bool Returns true if the CANN backend supports the buffer type,
* otherwise false.
*/
static bool ggml_backend_cann_supports_buft(
ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
if (ggml_backend_buft_is_cann(buft)) {
ggml_backend_cann_device_context * dev_ctx = (ggml_backend_cann_device_context *)dev->context;
ggml_backend_cann_buffer_type_context * buft_ctx =
(ggml_backend_cann_buffer_type_context *)buft->context;
return buft_ctx->device == dev_ctx->device;
}
return false;
}
static ggml_backend_buffer_type_t ggml_backend_cann_device_get_buffer_type(ggml_backend_dev_t dev) {
ggml_backend_cann_device_context * ctx = (ggml_backend_cann_device_context *)dev->context;
return ggml_backend_cann_buffer_type(ctx->device);
}
static ggml_backend_buffer_type_t ggml_backend_cann_device_get_host_buffer_type(ggml_backend_dev_t dev) {
GGML_UNUSED(dev);
return ggml_backend_cann_host_buffer_type();
}
/**
* @brief Creates a new event for the CANN backend device.
*
* This function initializes a new event for the CANN backend by setting the
* device and creating an ACL runtime event. The created event is then wrapped
* in a ggml_backend_event structure and returned.
*
* @param backend Pointer to the CANN backend.
* @return ggml_backend_event_t Returns a pointer to the new event structure.
*/
static ggml_backend_event_t ggml_backend_cann_device_event_new(
ggml_backend_dev_t dev) {
ggml_backend_cann_device_context * dev_ctx = (ggml_backend_cann_device_context *)dev->context;
ggml_cann_set_device(dev_ctx->device);
aclrtEvent event;
ACL_CHECK(aclrtCreateEvent(&event));
return new ggml_backend_event{
/* .device = */ ggml_backend_reg_dev_get(ggml_backend_cann_reg(), dev_ctx->device),
/* .context = */ event,
};
}
/**
* @brief Frees a CANN backend event.
*
* This function destroys the ACL runtime event associated with the given CANN
* backend event and then deletes the event structure itself.
*
* @param event Pointer to the event structure to be freed.
*/
static void ggml_backend_cann_device_event_free(ggml_backend_dev_t dev, ggml_backend_event_t event) {
ACL_CHECK(aclrtDestroyEvent((aclrtEvent)event->context));
delete event;
GGML_UNUSED(dev);
}
/**
* @brief Synchronizes the given event on the CANN backend.
*
* This function waits for the specified event to complete on the ACL runtime.
*
* @param event Pointer to the event structure to be synchronized.
*/
static void ggml_backend_cann_device_event_synchronize(ggml_backend_dev_t dev, ggml_backend_event_t event) {
ACL_CHECK(aclrtSynchronizeEvent((aclrtEvent)event->context));
GGML_UNUSED(dev);
}
static const ggml_backend_device_i ggml_backend_cann_device_interface = {
/* .get_name = */ ggml_backend_cann_device_get_name,
/* .get_description = */ ggml_backend_cann_device_get_description,
/* .get_memory = */ ggml_backend_cann_device_get_memory,
/* .get_type = */ ggml_backend_cann_device_get_type,
/* .get_props = */ ggml_backend_cann_device_get_props,
/* .init_backend = */ ggml_backend_cann_device_init, // called for every card
/* .get_buffer_type = */ ggml_backend_cann_device_get_buffer_type,
/* .get_host_buffer_type = */ ggml_backend_cann_device_get_host_buffer_type,
/* .buffer_from_host_ptr = */ NULL, // not supported for CANN
/* .supports_op = */ ggml_backend_cann_supports_op,
/* .supports_buft = */ ggml_backend_cann_supports_buft,
/* .offload_op = */ ggml_backend_cann_offload_op,
/* .event_new = */ ggml_backend_cann_device_event_new,
/* .event_free = */ ggml_backend_cann_device_event_free,
/* .event_synchronize = */ ggml_backend_cann_device_event_synchronize,
};
// backend reg
struct ggml_backend_cann_reg_context {
std::vector<ggml_backend_dev_t> devices;
};
static const char * ggml_backend_cann_reg_get_name(ggml_backend_reg_t reg) {
GGML_UNUSED(reg);
return GGML_CANN_NAME;
}
static size_t ggml_backend_cann_reg_get_device_count(ggml_backend_reg_t reg) {
ggml_backend_cann_reg_context * ctx = (ggml_backend_cann_reg_context *)reg->context;
return ctx->devices.size();
}
static ggml_backend_dev_t ggml_backend_cann_reg_get_device(ggml_backend_reg_t reg, size_t index) {
ggml_backend_cann_reg_context * ctx = (ggml_backend_cann_reg_context *)reg->context;
GGML_ASSERT(index < ctx->devices.size());
return ctx->devices[index];
}
static void * ggml_backend_cann_reg_get_proc_address(ggml_backend_reg_t reg, const char * name) {
GGML_UNUSED(reg);
GGML_UNUSED(name);
// reserved for future use
return nullptr;
}
static const ggml_backend_reg_i ggml_backend_cann_reg_interface = {
/* .get_name = */ ggml_backend_cann_reg_get_name,
/* .get_device_count = */ ggml_backend_cann_reg_get_device_count,
/* .get_device_get = */ ggml_backend_cann_reg_get_device,
/* .get_proc_address = */ ggml_backend_cann_reg_get_proc_address,
};
// backend registry, called only once for cann backend
ggml_backend_reg_t ggml_backend_cann_reg() {
static ggml_backend_reg reg;
static bool initialized = false;
{
static std::mutex mutex;
std::lock_guard<std::mutex> lock(mutex);
if (!initialized) {
aclInit(nullptr);
ggml_backend_cann_reg_context * ctx = new ggml_backend_cann_reg_context;
for (int i = 0; i < ggml_cann_info().device_count; i++) {
ggml_backend_cann_device_context* dev_ctx = new ggml_backend_cann_device_context();
dev_ctx->description = aclrtGetSocName();
dev_ctx->device = i;
dev_ctx->name = GGML_CANN_NAME + std::to_string(i);
ggml_cann_set_device(i);
ggml_backend_dev_t dev = new ggml_backend_device {
/* .interface = */ ggml_backend_cann_device_interface,
/* .reg = */ &reg,
/* .context = */ dev_ctx
};
ctx->devices.push_back(dev);
}
reg = ggml_backend_reg {
/* .interface = */ ggml_backend_cann_reg_interface,
/* .context = */ ctx
};
}
initialized = true;
}
return &reg;
}
ggml_backend_t ggml_backend_cann_init(int32_t device) {
aclInit(nullptr);
if (device < 0 || device >= ggml_backend_cann_get_device_count()) {
@ -1993,7 +2138,7 @@ ggml_backend_t ggml_backend_cann_init(int32_t device) {
ggml_backend_t cann_backend =
new ggml_backend{/* .guid = */ ggml_backend_cann_guid(),
/* .interface = */ ggml_backend_cann_interface,
/* .device = */ nullptr,
/* .device = */ ggml_backend_reg_dev_get(ggml_backend_cann_reg(), device),
/* .context = */ ctx};
return cann_backend;

42
ggml/src/ggml-cuda.cu
View File

@ -291,7 +291,7 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool {
return;
}
}
GGML_LOG_WARN(GGML_CUDA_NAME " buffer pool full, increase MAX_CUDA_BUFFERS\n");
GGML_LOG_DEBUG(GGML_CUDA_NAME " buffer pool full, increase MAX_CUDA_BUFFERS\n");
ggml_cuda_set_device(device);
CUDA_CHECK(cudaFree(ptr));
pool_size -= size;
@ -980,7 +980,7 @@ static void * ggml_cuda_host_malloc(size_t size) {
if (err != cudaSuccess) {
// clear the error
cudaGetLastError();
GGML_LOG_WARN("%s: failed to allocate %.2f MiB of pinned memory: %s\n", __func__,
GGML_LOG_DEBUG("%s: failed to allocate %.2f MiB of pinned memory: %s\n", __func__,
size / 1024.0 / 1024.0, cudaGetErrorString(err));
return nullptr;
}
@ -1151,8 +1151,8 @@ static cudaError_t ggml_cuda_cpy_tensor_2d(
void * dst, const struct ggml_tensor * src, int64_t i3, int64_t i2, int64_t i1_low, int64_t i1_high, cudaStream_t stream) {
GGML_ASSERT(ggml_backend_buffer_is_cuda(src->buffer));
char * src_ptr = (char *) src->data;
char * dst_ptr = (char *) dst;
const char * src_ptr = (const char *) src->data;
char * dst_ptr = (char *) dst;
const int64_t ne0 = src->ne[0];
const int64_t nb0 = src->nb[0];
@ -1162,7 +1162,7 @@ static cudaError_t ggml_cuda_cpy_tensor_2d(
const enum ggml_type type = src->type;
const int64_t ts = ggml_type_size(type);
const int64_t bs = ggml_blck_size(type);
int64_t i1_diff = i1_high - i1_low;
const int64_t i1_diff = i1_high - i1_low;
const char * x = src_ptr + i1_low*nb1 + i2*nb2 + i3*nb3;
if (nb0 == ts && nb1 == ts*ne0/bs) {
@ -1479,13 +1479,18 @@ static void ggml_cuda_op_mul_mat(
if (src0_is_contiguous) {
dev[id].src0_dd = split ? (char *) src0_extra->data_device[id] : (char *) src0->data;
} else {
dev[id].src0_dd = dev[id].src0_dd_alloc.alloc(ctx.pool(id), ggml_nbytes(src0));
// If src0 is not contiguous it will be copied to a temporary buffer.
// This buffer needs to be cleared entirely because multiple regions will function as padding.
const size_t nbytes_data = ggml_nbytes(src0);
const size_t nbytes_padding = ggml_row_size(src0->type, MATRIX_ROW_PADDING - ne00 % MATRIX_ROW_PADDING);
dev[id].src0_dd = dev[id].src0_dd_alloc.alloc(ctx.pool(id), nbytes_data + nbytes_padding);
CUDA_CHECK(cudaMemsetAsync(dev[id].src0_dd, 0, nbytes_data + nbytes_padding, stream));
}
// If src0 is on a temporary compute buffers (partial offloading) there may be some padding that needs to be cleared:
// If src0 is on a temporary compute buffer (partial offloading) there may be some padding that needs to be cleared:
if (ne00 % MATRIX_ROW_PADDING != 0 && ggml_is_quantized(src0->type) && ggml_backend_buffer_get_usage(src0->buffer) == GGML_BACKEND_BUFFER_USAGE_COMPUTE && src0->view_src == nullptr) {
const int64_t nbytes_data = ggml_row_size(src0->type, (dev[id].row_high - dev[id].row_low)*ne00);
const int64_t nbytes_padding = ggml_row_size(src0->type, MATRIX_ROW_PADDING - ne00 % MATRIX_ROW_PADDING);
const size_t nbytes_data = ggml_row_size(src0->type, (dev[id].row_high - dev[id].row_low)*ne00);
const size_t nbytes_padding = ggml_row_size(src0->type, MATRIX_ROW_PADDING - ne00 % MATRIX_ROW_PADDING);
CUDA_CHECK(cudaMemsetAsync(dev[id].src0_dd + nbytes_data , 0, nbytes_padding, stream));
}
@ -2406,7 +2411,7 @@ static bool ggml_backend_cuda_cpy_tensor_async(ggml_backend_t backend_src, ggml_
if (cuda_ctx_src->device != buf_ctx_src->device || cuda_ctx_dst->device != buf_ctx_dst->device) {
#ifndef NDEBUG
GGML_LOG_WARN("%s: backend and buffer devices do not match\n", __func__);
GGML_LOG_DEBUG("%s: backend and buffer devices do not match\n", __func__);
#endif
return false;
}
@ -2524,7 +2529,7 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend,
if (ggml_cuda_info().devices[cuda_ctx->device].cc < CC_AMPERE) {
cuda_ctx->cuda_graph->disable_due_to_gpu_arch = true;
#ifndef NDEBUG
GGML_LOG_WARN("%s: disabling CUDA graphs due to GPU architecture\n", __func__);
GGML_LOG_DEBUG("%s: disabling CUDA graphs due to GPU architecture\n", __func__);
#endif
}
}
@ -2575,14 +2580,14 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend,
if (node->src[0] && node->src[0]->buffer && ggml_backend_buffer_is_cuda_split(node->src[0]->buffer)) {
use_cuda_graph = false; // Split buffers are not supported by CUDA graph capture
#ifndef NDEBUG
GGML_LOG_WARN("%s: disabling CUDA graphs due to split buffer\n", __func__);
GGML_LOG_DEBUG("%s: disabling CUDA graphs due to split buffer\n", __func__);
#endif
}
if (node->op == GGML_OP_MUL_MAT_ID) {
use_cuda_graph = false; // This node type is not supported by CUDA graph capture
#ifndef NDEBUG
GGML_LOG_WARN("%s: disabling CUDA graphs due to mul_mat_id\n", __func__);
GGML_LOG_DEBUG("%s: disabling CUDA graphs due to mul_mat_id\n", __func__);
#endif
}
@ -2591,7 +2596,7 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend,
// Changes in batch size or context size can cause changes to the grid size of some kernels.
use_cuda_graph = false;
#ifndef NDEBUG
GGML_LOG_WARN("%s: disabling CUDA graphs due to batch size > 1 [%s] [%ld %ld %ld %ld]\n", __func__, node->name, node->ne[0], node->ne[1], node->ne[2], node->ne[3]);
GGML_LOG_DEBUG("%s: disabling CUDA graphs due to batch size > 1 [%s] [%ld %ld %ld %ld]\n", __func__, node->name, node->ne[0], node->ne[1], node->ne[2], node->ne[3]);
#endif
}
@ -2603,7 +2608,7 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend,
if (!ptr) {
use_cuda_graph = false;
#ifndef NDEBUG
GGML_LOG_WARN("%s: disabling CUDA graphs due to unsupported copy op\n", __func__);
GGML_LOG_DEBUG("%s: disabling CUDA graphs due to unsupported copy op\n", __func__);
#endif
} else {
if (std::find(ggml_cuda_cpy_fn_ptrs.begin(), ggml_cuda_cpy_fn_ptrs.end(), ptr) == ggml_cuda_cpy_fn_ptrs.end()) {
@ -2627,7 +2632,7 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend,
if (cuda_ctx->cuda_graph->number_consecutive_updates >= 4) {
cuda_ctx->cuda_graph->disable_due_to_too_many_updates = true;
#ifndef NDEBUG
GGML_LOG_WARN("%s: disabling CUDA graphs due to too many consecutive updates\n", __func__);
GGML_LOG_DEBUG("%s: disabling CUDA graphs due to too many consecutive updates\n", __func__);
#endif
}
}
@ -2685,7 +2690,7 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend,
use_cuda_graph = false;
cuda_ctx->cuda_graph->disable_due_to_failed_graph_capture = true;
#ifndef NDEBUG
GGML_LOG_WARN("%s: disabling CUDA graphs due to failed graph capture\n", __func__);
GGML_LOG_DEBUG("%s: disabling CUDA graphs due to failed graph capture\n", __func__);
#endif
} else {
graph_evaluated_or_captured = true; // CUDA graph has been captured
@ -2854,7 +2859,7 @@ bool ggml_backend_cuda_register_host_buffer(void * buffer, size_t size) {
// clear the error
cudaGetLastError();
GGML_LOG_WARN("%s: failed to register %.2f MiB of pinned memory: %s\n", __func__,
GGML_LOG_DEBUG("%s: failed to register %.2f MiB of pinned memory: %s\n", __func__,
size / 1024.0 / 1024.0, cudaGetErrorString(err));
return false;
}
@ -3141,7 +3146,6 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
case GGML_OP_ROPE:
return ggml_is_contiguous(op->src[0]);
case GGML_OP_IM2COL:
return op->src[0]->type == GGML_TYPE_F16;
case GGML_OP_POOL_2D:
case GGML_OP_SUM:
case GGML_OP_SUM_ROWS:

2
ggml/src/ggml-cuda/cpy.cuh
View File

@ -1,6 +1,6 @@
#include "common.cuh"
#define CUDA_CPY_BLOCK_SIZE 32
#define CUDA_CPY_BLOCK_SIZE 64
void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, ggml_tensor * src1);

34
ggml/src/ggml-cuda/dmmv.cu
View File

@ -416,10 +416,11 @@ static __global__ void dequantize_mul_mat_vec_q6_k(const void * __restrict__ vx,
static __device__ void convert_f16(const void * vx, const int64_t ib, const int iqs, dfloat2 & v){
const half * x = (const half *) vx;
// load 2 halfs into register in a single instruction
const half2 x_reg = *((half2 *) &(x[ib + iqs]));
// automatic half -> float type cast if dfloat == float
v.x = x[ib + iqs + 0];
v.y = x[ib + iqs + 1];
v.x = __low2float(x_reg);
v.y = __high2float(x_reg);
}
static constexpr __device__ dequantize_kernel_t get_dequantize_kernel(ggml_type type) {
@ -476,13 +477,28 @@ static __global__ void dequantize_mul_mat_vec(const void * __restrict__ vx, cons
// matrix multiplication
// for qr = 2 the y index needs to increase by 1 per j iter because of y_offset = qk/2
#ifdef GGML_CUDA_F16
tmp += __hmul2(v, {
y[iybs + iqs + j/qr + 0],
y[iybs + iqs + j/qr + y_offset]
});
if ( y_offset == 1 ) {
// load 2 dfloats into register in a single instruction
const dfloat2 y_reg = *((dfloat2 *) &(y[iybs + iqs + j/qr]));
tmp += __hmul2(v, y_reg);
}
else {
tmp += __hmul2(v, {
y[iybs + iqs + j/qr + 0],
y[iybs + iqs + j/qr + y_offset]
});
}
#else
tmp += v.x * y[iybs + iqs + j/qr + 0];
tmp += v.y * y[iybs + iqs + j/qr + y_offset];
if ( y_offset == 1 ) {
// load 2 dfloats into register in a single instruction
const dfloat2 y_reg = *((dfloat2 *) &(y[iybs + iqs + j/qr]));
tmp += v.x * y_reg.x;
tmp += v.y * y_reg.y;
}
else {
tmp += v.x * y[iybs + iqs + j/qr + 0];
tmp += v.y * y[iybs + iqs + j/qr + y_offset];
}
#endif // GGML_CUDA_F16
}
}

6
ggml/src/ggml-cuda/im2col.cu
View File

@ -91,9 +91,9 @@ void ggml_cuda_op_im2col(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const int64_t OH = is_2D ? dst->ne[2] : 1;
const int64_t OW = dst->ne[1];
const size_t delta_offset = src1->nb[is_2D ? 2 : 1] / 4; // nb is byte offset, src is type float32
const int64_t batch = src1->ne[3];
const size_t batch_offset = src1->nb[3] / 4; // nb is byte offset, src is type float32
const size_t delta_offset = src1->nb[is_2D ? 2 : 1] / 4; // nb is byte offset, src is type float32
const int64_t batch = src1->ne[is_2D ? 3 : 2];
const size_t batch_offset = src1->nb[is_2D ? 3 : 2] / 4; // nb is byte offset, src is type float32
if(dst->type == GGML_TYPE_F16) {
im2col_cuda_f16(src1_d, (half *) dst_d, IW, IH, OW, OH, KW, KH, IC, batch, batch_offset, delta_offset, s0, s1, p0, p1, d0, d1, stream);

4
ggml/src/ggml-cuda/mmq.cu
View File

@ -8,8 +8,6 @@ void ggml_cuda_op_mul_mat_q(
const int64_t ne00 = src0->ne[0];
const int64_t nb01 = src0->nb[1];
const int64_t ne10 = src1->ne[0];
const int64_t ne11 = src1->ne[1];
GGML_ASSERT(ne10 % QK8_1 == 0);
@ -17,7 +15,7 @@ void ggml_cuda_op_mul_mat_q(
const int64_t ne0 = dst->ne[0];
const int64_t row_diff = row_high - row_low;
const int64_t stride00 = nb01 / ggml_type_size(src0->type);
const int64_t stride00 = ne00 / ggml_blck_size(src0->type);
int id = ggml_cuda_get_device();
const int compute_capability = ggml_cuda_info().devices[id].cc;

8
ggml/src/ggml-impl.h
View File

@ -19,6 +19,9 @@ extern "C" {
#define MIN(a, b) ((a) < (b) ? (a) : (b))
#define MAX(a, b) ((a) > (b) ? (a) : (b))
// required for mmap as gguf only guarantees 32-byte alignment
#define TENSOR_ALIGNMENT 32
// static_assert should be a #define, but if it's not,
// fall back to the _Static_assert C11 keyword.
// if C99 - static_assert is noop
@ -196,6 +199,11 @@ struct ggml_cgraph {
struct ggml_cgraph ggml_graph_view(struct ggml_cgraph * cgraph, int i0, int i1);
// Memory allocation
void * ggml_aligned_malloc(size_t size);
void ggml_aligned_free(void * ptr, size_t size);
#ifdef __cplusplus
}
#endif

203
ggml/src/ggml-metal.m
View File

@ -241,6 +241,8 @@ enum ggml_metal_kernel_type {
GGML_METAL_KERNEL_TYPE_ROPE_NEOX_F16,
GGML_METAL_KERNEL_TYPE_IM2COL_F16,
GGML_METAL_KERNEL_TYPE_IM2COL_F32,
GGML_METAL_KERNEL_TYPE_IM2COL_EXT_F16,
GGML_METAL_KERNEL_TYPE_IM2COL_EXT_F32,
GGML_METAL_KERNEL_TYPE_UPSCALE_F32,
GGML_METAL_KERNEL_TYPE_PAD_F32,
GGML_METAL_KERNEL_TYPE_ARANGE_F32,
@ -272,6 +274,8 @@ enum ggml_metal_kernel_type {
GGML_METAL_KERNEL_TYPE_SIN,
GGML_METAL_KERNEL_TYPE_COS,
GGML_METAL_KERNEL_TYPE_SUM_ROWS,
GGML_METAL_KERNEL_TYPE_POOL_2D_AVG_F32,
GGML_METAL_KERNEL_TYPE_POOL_2D_MAX_F32,
GGML_METAL_KERNEL_TYPE_COUNT
};
@ -685,6 +689,8 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_NEOX_F16, rope_neox_f16, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_IM2COL_F16, im2col_f16, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_IM2COL_F32, im2col_f32, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_IM2COL_EXT_F16, im2col_ext_f16, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_IM2COL_EXT_F32, im2col_ext_f32, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_UPSCALE_F32, upscale_f32, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_PAD_F32, pad_f32, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_TIMESTEP_EMBEDDING_F32, timestep_embedding_f32, true);
@ -716,6 +722,8 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SIN, sin, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_COS, cos, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SUM_ROWS, sum_rows, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_POOL_2D_AVG_F32, pool_2d_avg_f32, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_POOL_2D_MAX_F32, pool_2d_max_f32, true);
}
[metal_library release];
@ -844,8 +852,8 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
case GGML_OP_IM2COL:
return op->src[0]->type == GGML_TYPE_F16;
case GGML_OP_POOL_1D:
case GGML_OP_POOL_2D:
return false;
case GGML_OP_POOL_2D:
case GGML_OP_UPSCALE:
case GGML_OP_PAD:
case GGML_OP_ARANGE:
@ -1007,19 +1015,21 @@ static void ggml_metal_encode_node(
id<MTLBuffer> id_src2 = src2 ? ggml_metal_get_buffer(src2, &offs_src2) : nil;
id<MTLBuffer> id_dst = dst ? ggml_metal_get_buffer(dst, &offs_dst) : nil;
//GGML_LOG_INFO("%s: op - %s\n", __func__, ggml_op_name(dst->op));
//if (src0) {
// GGML_LOG_INFO("%s: src0 - %4s [%5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(src0t), ne00, ne01, ne02,
// ggml_is_contiguous(src0), src0->name);
//}
//if (src1) {
// GGML_LOG_INFO("%s: src1 - %4s [%5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(src1t), ne10, ne11, ne12,
// ggml_is_contiguous(src1), src1->name);
//}
//if (dst) {
// GGML_LOG_INFO("%s: dst - %4s [%5lld, %5lld, %5lld], 1, %s\n", __func__, ggml_type_name(dstt), ne0, ne1, ne2,
// dst->name);
//}
#if 0
GGML_LOG_INFO("%s: op - %s\n", __func__, ggml_op_name(dst->op));
if (src0) {
GGML_LOG_INFO("%s: src0 - %4s [%5lld, %5lld, %5lld, %5lld] [%5lld, %5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(src0t), ne00, ne01, ne02, ne03, nb00, nb01, nb02, nb03,
ggml_is_contiguous(src0), src0->name);
}
if (src1) {
GGML_LOG_INFO("%s: src1 - %4s [%5lld, %5lld, %5lld, %5lld] [%5lld, %5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(src1t), ne10, ne11, ne12, ne13, nb10, nb11, nb12, nb13,
ggml_is_contiguous(src1), src1->name);
}
if (dst) {
GGML_LOG_INFO("%s: dst - %4s [%5lld, %5lld, %5lld, %5lld] [%5lld, %5lld, %5lld, %5lld], 1, %s\n", __func__, ggml_type_name(dstt), ne0, ne1, ne2, ne3, nb0, nb1, nb2, nb3,
dst->name);
}
#endif
id<MTLDevice> device = ctx_dev->mtl_device;
@ -1802,14 +1812,16 @@ static void ggml_metal_encode_node(
[encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4];
[encoder setBytes:&nb01 length:sizeof(nb01) atIndex:5];
[encoder setBytes:&nb02 length:sizeof(nb02) atIndex:6];
[encoder setBytes:&ne12 length:sizeof(ne12) atIndex:7];
[encoder setBytes:&nb10 length:sizeof(nb10) atIndex:8];
[encoder setBytes:&nb11 length:sizeof(nb11) atIndex:9];
[encoder setBytes:&nb12 length:sizeof(nb12) atIndex:10];
[encoder setBytes:&ne0 length:sizeof(ne0) atIndex:11];
[encoder setBytes:&ne1 length:sizeof(ne1) atIndex:12];
[encoder setBytes:&r2 length:sizeof(r2) atIndex:13];
[encoder setBytes:&r3 length:sizeof(r3) atIndex:14];
[encoder setBytes:&nb03 length:sizeof(nb03) atIndex:7];
[encoder setBytes:&ne12 length:sizeof(ne12) atIndex:8];
[encoder setBytes:&nb10 length:sizeof(nb10) atIndex:9];
[encoder setBytes:&nb11 length:sizeof(nb11) atIndex:10];
[encoder setBytes:&nb12 length:sizeof(nb12) atIndex:11];
[encoder setBytes:&nb13 length:sizeof(nb13) atIndex:12];
[encoder setBytes:&ne0 length:sizeof(ne0) atIndex:13];
[encoder setBytes:&ne1 length:sizeof(ne1) atIndex:14];
[encoder setBytes:&r2 length:sizeof(r2) atIndex:15];
[encoder setBytes:&r3 length:sizeof(r3) atIndex:16];
[encoder setThreadgroupMemoryLength:8192 atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake( (ne11 + 31)/32, (ne01 + 63)/64, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)];
} else {
@ -1978,20 +1990,22 @@ static void ggml_metal_encode_node(
[encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6];
[encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7];
[encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8];
[encoder setBytes:&ne10 length:sizeof(ne10) atIndex:9];
[encoder setBytes:&ne11 length:sizeof(ne11) atIndex:10];
[encoder setBytes:&ne12 length:sizeof(ne12) atIndex:11];
[encoder setBytes:&nb10 length:sizeof(nb10) atIndex:12];
[encoder setBytes:&nb11 length:sizeof(nb11) atIndex:13];
[encoder setBytes:&nb12 length:sizeof(nb12) atIndex:14];
[encoder setBytes:&ne0 length:sizeof(ne0) atIndex:15];
[encoder setBytes:&ne1 length:sizeof(ne1) atIndex:16];
[encoder setBytes:&r2 length:sizeof(r2) atIndex:17];
[encoder setBytes:&r3 length:sizeof(r3) atIndex:18];
[encoder setBytes:&nb03 length:sizeof(nb03) atIndex:9];
[encoder setBytes:&ne10 length:sizeof(ne10) atIndex:10];
[encoder setBytes:&ne11 length:sizeof(ne11) atIndex:11];
[encoder setBytes:&ne12 length:sizeof(ne12) atIndex:12];
[encoder setBytes:&nb10 length:sizeof(nb10) atIndex:13];
[encoder setBytes:&nb11 length:sizeof(nb11) atIndex:14];
[encoder setBytes:&nb12 length:sizeof(nb12) atIndex:15];
[encoder setBytes:&nb13 length:sizeof(nb13) atIndex:16];
[encoder setBytes:&ne0 length:sizeof(ne0) atIndex:17];
[encoder setBytes:&ne1 length:sizeof(ne1) atIndex:18];
[encoder setBytes:&r2 length:sizeof(r2) atIndex:19];
[encoder setBytes:&r3 length:sizeof(r3) atIndex:20];
if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1 || src0t == GGML_TYPE_Q5_0 ||
src0t == GGML_TYPE_Q5_1 || src0t == GGML_TYPE_Q8_0 || src0t == GGML_TYPE_Q2_K ||
src0t == GGML_TYPE_IQ1_S || src0t == GGML_TYPE_IQ1_M || src0t == GGML_TYPE_IQ2_S) {
src0t == GGML_TYPE_Q5_1 || src0t == GGML_TYPE_Q8_0 || src0t == GGML_TYPE_Q2_K ||
src0t == GGML_TYPE_IQ1_S || src0t == GGML_TYPE_IQ1_M || src0t == GGML_TYPE_IQ2_S) {
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
}
else if (src0t == GGML_TYPE_IQ2_XXS || src0t == GGML_TYPE_IQ2_XS) {
@ -2040,6 +2054,9 @@ static void ggml_metal_encode_node(
GGML_ASSERT(src1t == GGML_TYPE_F32);
GGML_ASSERT(ne03 == 1);
GGML_ASSERT(ne13 == 1);
// find the break-even point where the matrix-matrix kernel becomes more efficient compared
// to the matrix-vector kernel
// ne20 = n_used_experts
@ -2545,6 +2562,8 @@ static void ggml_metal_encode_node(
} break;
case GGML_OP_IM2COL:
{
GGML_ASSERT(ggml_is_contiguous(src0));
GGML_ASSERT(ggml_is_contiguous(src1));
GGML_ASSERT(src0->type == GGML_TYPE_F16);
GGML_ASSERT(src1->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F16 || dst->type == GGML_TYPE_F32);
@ -2574,30 +2593,54 @@ static void ggml_metal_encode_node(
const int32_t ofs0 = src1->nb[is_2D ? 3 : 2] / 4;
const int32_t ofs1 = src1->nb[is_2D ? 2 : 1] / 4;
id<MTLComputePipelineState> pipeline = nil;
id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_IM2COL_F32].pipeline;
const bool is_gt_mttpt = ((size_t)(N * KH * KW)) > pipeline.maxTotalThreadsPerThreadgroup;
switch (dst->type) {
case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_IM2COL_F32].pipeline; break;
case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_IM2COL_F16].pipeline; break;
case GGML_TYPE_F32: {
pipeline = (is_gt_mttpt ?
ctx->kernels[GGML_METAL_KERNEL_TYPE_IM2COL_EXT_F32].pipeline
:
ctx->kernels[GGML_METAL_KERNEL_TYPE_IM2COL_F32].pipeline);
} break;
case GGML_TYPE_F16: {
pipeline = (is_gt_mttpt ?
ctx->kernels[GGML_METAL_KERNEL_TYPE_IM2COL_EXT_F16].pipeline
:
ctx->kernels[GGML_METAL_KERNEL_TYPE_IM2COL_F16].pipeline);
} break;
default: GGML_ABORT("fatal error");
};
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src1 offset:offs_src1 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
[encoder setBytes:&ofs0 length:sizeof( int32_t) atIndex:2];
[encoder setBytes:&ofs1 length:sizeof( int32_t) atIndex:3];
[encoder setBytes:&IW length:sizeof( int32_t) atIndex:4];
[encoder setBytes:&IH length:sizeof( int32_t) atIndex:5];
[encoder setBytes:&CHW length:sizeof( int32_t) atIndex:6];
[encoder setBytes:&s0 length:sizeof( int32_t) atIndex:7];
[encoder setBytes:&s1 length:sizeof( int32_t) atIndex:8];
[encoder setBytes:&p0 length:sizeof( int32_t) atIndex:9];
[encoder setBytes:&p1 length:sizeof( int32_t) atIndex:10];
[encoder setBytes:&d0 length:sizeof( int32_t) atIndex:11];
[encoder setBytes:&d1 length:sizeof( int32_t) atIndex:12];
[encoder setBuffer:id_src1 offset:offs_src1 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
[encoder setBytes:&ofs0 length:sizeof(int32_t) atIndex:2];
[encoder setBytes:&ofs1 length:sizeof(int32_t) atIndex:3];
[encoder setBytes:&IW length:sizeof(int32_t) atIndex:4];
[encoder setBytes:&IH length:sizeof(int32_t) atIndex:5];
[encoder setBytes:&CHW length:sizeof(int32_t) atIndex:6];
[encoder setBytes:&s0 length:sizeof(int32_t) atIndex:7];
[encoder setBytes:&s1 length:sizeof(int32_t) atIndex:8];
[encoder setBytes:&p0 length:sizeof(int32_t) atIndex:9];
[encoder setBytes:&p1 length:sizeof(int32_t) atIndex:10];
[encoder setBytes:&d0 length:sizeof(int32_t) atIndex:11];
[encoder setBytes:&d1 length:sizeof(int32_t) atIndex:12];
[encoder dispatchThreadgroups:MTLSizeMake(IC, OH, OW) threadsPerThreadgroup:MTLSizeMake(N, KH, KW)];
if (is_gt_mttpt) {
[encoder setBytes:&N length:sizeof(int32_t) atIndex:13];
[encoder setBytes:&KH length:sizeof(int32_t) atIndex:14];
[encoder setBytes:&KW length:sizeof(int32_t) atIndex:15];
const uint64_t n_threads = MIN(pipeline.maxTotalThreadsPerThreadgroup, (uint64_t)N);
const int64_t quotient = N / n_threads + (N % n_threads > 0 ? 1 : 0);
[encoder dispatchThreadgroups:MTLSizeMake(quotient * CHW, OH, OW) threadsPerThreadgroup:MTLSizeMake(n_threads, 1, 1)];
} else {
[encoder dispatchThreadgroups:MTLSizeMake(IC, OH, OW) threadsPerThreadgroup:MTLSizeMake(N, KH, KW)];
}
} break;
case GGML_OP_UPSCALE:
{
@ -3001,6 +3044,64 @@ static void ggml_metal_encode_node(
[encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
} break;
case GGML_OP_POOL_2D:
{
GGML_ASSERT(ggml_is_contiguous(src0));
GGML_ASSERT(src0t == GGML_TYPE_F32 && src0t == dstt);
const int32_t * opts = dst->op_params;
enum ggml_op_pool op = opts[0];
id<MTLComputePipelineState> pipeline = nil;
switch (src0t) {
case GGML_TYPE_F32: {
switch(op) {
case GGML_OP_POOL_AVG:
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_POOL_2D_AVG_F32].pipeline; break;
case GGML_OP_POOL_MAX:
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_POOL_2D_MAX_F32].pipeline; break;
default: GGML_ASSERT(false && "not implemented");
}
} break;
default: GGML_ASSERT(false && "not implemented");
}
const int32_t k0 = opts[1];
const int32_t k1 = opts[2];
const int32_t s0 = opts[3];
const int32_t s1 = opts[4];
const int32_t p0 = opts[5];
const int32_t p1 = opts[6];
const int64_t IH = src0->ne[1];
const int64_t IW = src0->ne[0];
const int64_t N = dst->ne[3];
const int64_t OC = dst->ne[2];
const int64_t OH = dst->ne[1];
const int64_t OW = dst->ne[0];
const int64_t parallel_elements = N * OC * OH * OW;
const int64_t n_threads = MIN((int64_t)[pipeline maxTotalThreadsPerThreadgroup], parallel_elements);
const int64_t n_tg = (parallel_elements + n_threads - 1) / n_threads;
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
[encoder setBytes:&k0 length:sizeof(int32_t) atIndex:2];
[encoder setBytes:&k1 length:sizeof(int32_t) atIndex:3];
[encoder setBytes:&s0 length:sizeof(int32_t) atIndex:4];
[encoder setBytes:&s1 length:sizeof(int32_t) atIndex:5];
[encoder setBytes:&p0 length:sizeof(int32_t) atIndex:6];
[encoder setBytes:&p1 length:sizeof(int32_t) atIndex:7];
[encoder setBytes:&IH length:sizeof(int64_t) atIndex:8];
[encoder setBytes:&IW length:sizeof(int64_t) atIndex:9];
[encoder setBytes:&OH length:sizeof(int64_t) atIndex:10];
[encoder setBytes:&OW length:sizeof(int64_t) atIndex:11];
[encoder setBytes:&parallel_elements length:sizeof(int64_t) atIndex:12];
[encoder dispatchThreadgroups:MTLSizeMake(n_tg, 1, 1) threadsPerThreadgroup:MTLSizeMake(n_threads, 1, 1)];
} break;
default:
{
GGML_LOG_ERROR("%s: error: node %3d, op = %8s not implemented\n", __func__, idx, ggml_op_name(dst->op));

756
ggml/src/ggml-metal.metal
View File

File diff suppressed because it is too large Load Diff

560
ggml/src/ggml-rpc.cpp
View File

@ -57,8 +57,9 @@ struct socket_t {
}
};
// ggml_tensor is serialized into rpc_tensor
// all RPC structures must be packed
#pragma pack(push, 1)
// ggml_tensor is serialized into rpc_tensor
struct rpc_tensor {
uint64_t id;
uint32_t type;
@ -76,7 +77,6 @@ struct rpc_tensor {
char padding[4];
};
#pragma pack(pop)
static_assert(sizeof(rpc_tensor) % 8 == 0, "rpc_tensor size must be multiple of 8");
@ -96,6 +96,65 @@ enum rpc_cmd {
RPC_CMD_COUNT,
};
struct rpc_msg_alloc_buffer_req {
uint64_t size;
};
struct rpc_msg_alloc_buffer_rsp {
uint64_t remote_ptr;
uint64_t remote_size;
};
struct rpc_msg_get_alignment_rsp {
uint64_t alignment;
};
struct rpc_msg_get_max_size_rsp {
uint64_t max_size;
};
struct rpc_msg_buffer_get_base_req {
uint64_t remote_ptr;
};
struct rpc_msg_buffer_get_base_rsp {
uint64_t base_ptr;
};
struct rpc_msg_free_buffer_req {
uint64_t remote_ptr;
};
struct rpc_msg_buffer_clear_req {
uint64_t remote_ptr;
uint8_t value;
};
struct rpc_msg_get_tensor_req {
rpc_tensor tensor;
uint64_t offset;
uint64_t size;
};
struct rpc_msg_copy_tensor_req {
rpc_tensor src;
rpc_tensor dst;
};
struct rpc_msg_copy_tensor_rsp {
uint8_t result;
};
struct rpc_msg_graph_compute_rsp {
uint8_t result;
};
struct rpc_msg_get_device_memory_rsp {
uint64_t free_mem;
uint64_t total_mem;
};
#pragma pack(pop)
// RPC data structures
static ggml_guid_t ggml_backend_rpc_guid() {
@ -240,6 +299,38 @@ static bool recv_data(sockfd_t sockfd, void * data, size_t size) {
return true;
}
static bool send_msg(sockfd_t sockfd, const void * msg, size_t msg_size) {
if (!send_data(sockfd, &msg_size, sizeof(msg_size))) {
return false;
}
return send_data(sockfd, msg, msg_size);
}
static bool recv_msg(sockfd_t sockfd, void * msg, size_t msg_size) {
uint64_t size;
if (!recv_data(sockfd, &size, sizeof(size))) {
return false;
}
if (size != msg_size) {
return false;
}
return recv_data(sockfd, msg, msg_size);
}
static bool recv_msg(sockfd_t sockfd, std::vector<uint8_t> & input) {
uint64_t size;
if (!recv_data(sockfd, &size, sizeof(size))) {
return false;
}
try {
input.resize(size);
} catch (const std::bad_alloc & e) {
fprintf(stderr, "Failed to allocate input buffer of size %" PRIu64 "\n", size);
return false;
}
return recv_data(sockfd, input.data(), size);
}
static bool parse_endpoint(const std::string & endpoint, std::string & host, int & port) {
size_t pos = endpoint.find(':');
if (pos == std::string::npos) {
@ -252,28 +343,27 @@ static bool parse_endpoint(const std::string & endpoint, std::string & host, int
// RPC request : | rpc_cmd (1 byte) | request_size (8 bytes) | request_data (request_size bytes) |
// RPC response: | response_size (8 bytes) | response_data (response_size bytes) |
static bool send_rpc_cmd(const std::shared_ptr<socket_t> & sock, enum rpc_cmd cmd, const std::vector<uint8_t> & input, std::vector<uint8_t> & output) {
static bool send_rpc_cmd(const std::shared_ptr<socket_t> & sock, enum rpc_cmd cmd, const void * input, size_t input_size, void * output, size_t output_size) {
uint8_t cmd_byte = cmd;
if (!send_data(sock->fd, &cmd_byte, sizeof(cmd_byte))) {
return false;
}
uint64_t input_size = input.size();
if (!send_data(sock->fd, &input_size, sizeof(input_size))) {
return false;
}
if (!send_data(sock->fd, input.data(), input.size())) {
if (!send_data(sock->fd, input, input_size)) {
return false;
}
uint64_t output_size;
if (!recv_data(sock->fd, &output_size, sizeof(output_size))) {
// TODO: currently the output_size is always known, do we need support for commands with variable output size?
// even if we do, we can skip sending output_size from the server for commands with known output size
uint64_t out_size;
if (!recv_data(sock->fd, &out_size, sizeof(out_size))) {
return false;
}
if (output_size == 0) {
output.clear();
return true;
if (out_size != output_size) {
return false;
}
output.resize(output_size);
if (!recv_data(sock->fd, output.data(), output_size)) {
if (!recv_data(sock->fd, output, output_size)) {
return false;
}
return true;
@ -326,14 +416,9 @@ static const char * ggml_backend_rpc_buffer_get_name(ggml_backend_buffer_t buffe
static void ggml_backend_rpc_buffer_free_buffer(ggml_backend_buffer_t buffer) {
ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context;
// input serialization format: | remote_ptr (8 bytes) |
std::vector<uint8_t> input(sizeof(uint64_t), 0);
uint64_t remote_ptr = ctx->remote_ptr;
memcpy(input.data(), &remote_ptr, sizeof(remote_ptr));
std::vector<uint8_t> output;
bool status = send_rpc_cmd(ctx->sock, RPC_CMD_FREE_BUFFER, input, output);
rpc_msg_free_buffer_req request = {ctx->remote_ptr};
bool status = send_rpc_cmd(ctx->sock, RPC_CMD_FREE_BUFFER, &request, sizeof(request), nullptr, 0);
GGML_ASSERT(status);
GGML_ASSERT(output.empty());
delete ctx;
}
@ -342,20 +427,13 @@ static void * ggml_backend_rpc_buffer_get_base(ggml_backend_buffer_t buffer) {
if (ctx->base_cache.find(buffer) != ctx->base_cache.end()) {
return ctx->base_cache[buffer];
}
// input serialization format: | remote_ptr (8 bytes) |
std::vector<uint8_t> input(sizeof(uint64_t), 0);
uint64_t remote_ptr = ctx->remote_ptr;
memcpy(input.data(), &remote_ptr, sizeof(remote_ptr));
std::vector<uint8_t> output;
bool status = send_rpc_cmd(ctx->sock, RPC_CMD_BUFFER_GET_BASE, input, output);
rpc_msg_buffer_get_base_req request = {ctx->remote_ptr};
rpc_msg_buffer_get_base_rsp response;
bool status = send_rpc_cmd(ctx->sock, RPC_CMD_BUFFER_GET_BASE, &request, sizeof(request), &response, sizeof(response));
GGML_ASSERT(status);
GGML_ASSERT(output.size() == sizeof(uint64_t));
// output serialization format: | base_ptr (8 bytes) |
uint64_t base_ptr;
memcpy(&base_ptr, output.data(), sizeof(base_ptr));
void * base = reinterpret_cast<void *>(base_ptr);
ctx->base_cache[buffer] = base;
return base;
void * base_ptr = reinterpret_cast<void *>(response.base_ptr);
ctx->base_cache[buffer] = base_ptr;
return base_ptr;
}
static rpc_tensor serialize_tensor(const ggml_tensor * tensor) {
@ -405,26 +483,18 @@ static void ggml_backend_rpc_buffer_set_tensor(ggml_backend_buffer_t buffer, ggm
memcpy(input.data(), &rpc_tensor, sizeof(rpc_tensor));
memcpy(input.data() + sizeof(rpc_tensor), &offset, sizeof(offset));
memcpy(input.data() + sizeof(rpc_tensor) + sizeof(offset), data, size);
std::vector<uint8_t> output;
bool status = send_rpc_cmd(ctx->sock, RPC_CMD_SET_TENSOR, input, output);
bool status = send_rpc_cmd(ctx->sock, RPC_CMD_SET_TENSOR, input.data(), input.size(), nullptr, 0);
GGML_ASSERT(status);
}
static void ggml_backend_rpc_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context;
// input serialization format: | rpc_tensor | offset (8 bytes) | size (8 bytes) |
int input_size = sizeof(rpc_tensor) + 2*sizeof(uint64_t);
std::vector<uint8_t> input(input_size, 0);
rpc_tensor rpc_tensor = serialize_tensor(tensor);
memcpy(input.data(), &rpc_tensor, sizeof(rpc_tensor));
memcpy(input.data() + sizeof(rpc_tensor), &offset, sizeof(offset));
memcpy(input.data() + sizeof(rpc_tensor) + sizeof(offset), &size, sizeof(size));
std::vector<uint8_t> output;
bool status = send_rpc_cmd(ctx->sock, RPC_CMD_GET_TENSOR, input, output);
rpc_msg_get_tensor_req request;
request.tensor = serialize_tensor(tensor);
request.offset = offset;
request.size = size;
bool status = send_rpc_cmd(ctx->sock, RPC_CMD_GET_TENSOR, &request, sizeof(request), data, size);
GGML_ASSERT(status);
GGML_ASSERT(output.size() == size);
// output serialization format: | data (size bytes) |
memcpy(data, output.data(), size);
}
static bool ggml_backend_rpc_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * src, ggml_tensor * dst) {
@ -437,30 +507,19 @@ static bool ggml_backend_rpc_buffer_cpy_tensor(ggml_backend_buffer_t buffer, con
return false;
}
ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context;
// input serialization format: | rpc_tensor src | rpc_tensor dst |
int input_size = 2*sizeof(rpc_tensor);
std::vector<uint8_t> input(input_size, 0);
rpc_tensor rpc_src = serialize_tensor(src);
rpc_tensor rpc_dst = serialize_tensor(dst);
memcpy(input.data(), &rpc_src, sizeof(rpc_src));
memcpy(input.data() + sizeof(rpc_src), &rpc_dst, sizeof(rpc_dst));
std::vector<uint8_t> output;
bool status = send_rpc_cmd(ctx->sock, RPC_CMD_COPY_TENSOR, input, output);
rpc_msg_copy_tensor_req request;
request.src = serialize_tensor(src);
request.dst = serialize_tensor(dst);
rpc_msg_copy_tensor_rsp response;
bool status = send_rpc_cmd(ctx->sock, RPC_CMD_COPY_TENSOR, &request, sizeof(request), &response, sizeof(response));
GGML_ASSERT(status);
// output serialization format: | result (1 byte) |
GGML_ASSERT(output.size() == 1);
return output[0];
return response.result;
}
static void ggml_backend_rpc_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context;
// serialization format: | bufptr (8 bytes) | value (1 byte) |
int input_size = sizeof(uint64_t) + sizeof(uint8_t);
std::vector<uint8_t> input(input_size, 0);
memcpy(input.data(), &ctx->remote_ptr, sizeof(ctx->remote_ptr));
memcpy(input.data() + sizeof(ctx->remote_ptr), &value, sizeof(value));
std::vector<uint8_t> output;
bool status = send_rpc_cmd(ctx->sock, RPC_CMD_BUFFER_CLEAR, input, output);
rpc_msg_buffer_clear_req request = {ctx->remote_ptr, value};
bool status = send_rpc_cmd(ctx->sock, RPC_CMD_BUFFER_CLEAR, &request, sizeof(request), nullptr, 0);
GGML_ASSERT(status);
}
@ -484,25 +543,16 @@ static const char * ggml_backend_rpc_buffer_type_name(ggml_backend_buffer_type_t
static ggml_backend_buffer_t ggml_backend_rpc_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
ggml_backend_rpc_buffer_type_context * buft_ctx = (ggml_backend_rpc_buffer_type_context *)buft->context;
// input serialization format: | size (8 bytes) |
int input_size = sizeof(uint64_t);
std::vector<uint8_t> input(input_size, 0);
memcpy(input.data(), &size, sizeof(size));
std::vector<uint8_t> output;
rpc_msg_alloc_buffer_req request = {size};
rpc_msg_alloc_buffer_rsp response;
auto sock = get_socket(buft_ctx->endpoint);
bool status = send_rpc_cmd(sock, RPC_CMD_ALLOC_BUFFER, input, output);
bool status = send_rpc_cmd(sock, RPC_CMD_ALLOC_BUFFER, &request, sizeof(request), &response, sizeof(response));
GGML_ASSERT(status);
GGML_ASSERT(output.size() == 2*sizeof(uint64_t));
// output serialization format: | remote_ptr (8 bytes) | remote_size (8 bytes) |
uint64_t remote_ptr;
memcpy(&remote_ptr, output.data(), sizeof(remote_ptr));
size_t remote_size;
memcpy(&remote_size, output.data() + sizeof(uint64_t), sizeof(remote_size));
if (remote_ptr != 0) {
if (response.remote_ptr != 0) {
ggml_backend_buffer_t buffer = ggml_backend_buffer_init(buft,
ggml_backend_rpc_buffer_interface,
new ggml_backend_rpc_buffer_context{sock, {}, remote_ptr, "RPC[" + std::string(buft_ctx->endpoint) + "]"},
remote_size);
new ggml_backend_rpc_buffer_context{sock, {}, response.remote_ptr, "RPC[" + std::string(buft_ctx->endpoint) + "]"},
response.remote_size);
return buffer;
} else {
return nullptr;
@ -510,16 +560,10 @@ static ggml_backend_buffer_t ggml_backend_rpc_buffer_type_alloc_buffer(ggml_back
}
static size_t get_alignment(const std::shared_ptr<socket_t> & sock) {
// input serialization format: | 0 bytes |
std::vector<uint8_t> input;
std::vector<uint8_t> output;
bool status = send_rpc_cmd(sock, RPC_CMD_GET_ALIGNMENT, input, output);
rpc_msg_get_alignment_rsp response;
bool status = send_rpc_cmd(sock, RPC_CMD_GET_ALIGNMENT, nullptr, 0, &response, sizeof(response));
GGML_ASSERT(status);
GGML_ASSERT(output.size() == sizeof(uint64_t));
// output serialization format: | alignment (8 bytes) |
uint64_t alignment;
memcpy(&alignment, output.data(), sizeof(alignment));
return alignment;
return response.alignment;
}
static size_t ggml_backend_rpc_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
@ -528,16 +572,10 @@ static size_t ggml_backend_rpc_buffer_type_get_alignment(ggml_backend_buffer_typ
}
static size_t get_max_size(const std::shared_ptr<socket_t> & sock) {
// input serialization format: | 0 bytes |
std::vector<uint8_t> input;
std::vector<uint8_t> output;
bool status = send_rpc_cmd(sock, RPC_CMD_GET_MAX_SIZE, input, output);
rpc_msg_get_max_size_rsp response;
bool status = send_rpc_cmd(sock, RPC_CMD_GET_MAX_SIZE, nullptr, 0, &response, sizeof(response));
GGML_ASSERT(status);
GGML_ASSERT(output.size() == sizeof(uint64_t));
// output serialization format: | max_size (8 bytes) |
uint64_t max_size;
memcpy(&max_size, output.data(), sizeof(max_size));
return max_size;
return response.max_size;
}
static size_t ggml_backend_rpc_get_max_size(ggml_backend_buffer_type_t buft) {
@ -622,12 +660,11 @@ static enum ggml_status ggml_backend_rpc_graph_compute(ggml_backend_t backend, g
ggml_backend_rpc_context * rpc_ctx = (ggml_backend_rpc_context *)backend->context;
std::vector<uint8_t> input;
serialize_graph(cgraph, input);
std::vector<uint8_t> output;
rpc_msg_graph_compute_rsp response;
auto sock = get_socket(rpc_ctx->endpoint);
bool status = send_rpc_cmd(sock, RPC_CMD_GRAPH_COMPUTE, input, output);
bool status = send_rpc_cmd(sock, RPC_CMD_GRAPH_COMPUTE, input.data(), input.size(), &response, sizeof(response));
GGML_ASSERT(status);
GGML_ASSERT(output.size() == 1);
return (enum ggml_status)output[0];
return (enum ggml_status)response.result;
}
static ggml_backend_i ggml_backend_rpc_interface = {
@ -702,19 +739,11 @@ GGML_API bool ggml_backend_is_rpc(ggml_backend_t backend) {
}
static void get_device_memory(const std::shared_ptr<socket_t> & sock, size_t * free, size_t * total) {
// input serialization format: | 0 bytes |
std::vector<uint8_t> input;
std::vector<uint8_t> output;
bool status = send_rpc_cmd(sock, RPC_CMD_GET_DEVICE_MEMORY, input, output);
rpc_msg_get_device_memory_rsp response;
bool status = send_rpc_cmd(sock, RPC_CMD_GET_DEVICE_MEMORY, nullptr, 0, &response, sizeof(response));
GGML_ASSERT(status);
GGML_ASSERT(output.size() == 2*sizeof(uint64_t));
// output serialization format: | free (8 bytes) | total (8 bytes) |
uint64_t free_mem;
memcpy(&free_mem, output.data(), sizeof(free_mem));
uint64_t total_mem;
memcpy(&total_mem, output.data() + sizeof(uint64_t), sizeof(total_mem));
*free = free_mem;
*total = total_mem;
*free = response.free_mem;
*total = response.total_mem;
}
GGML_API void ggml_backend_rpc_get_device_memory(const char * endpoint, size_t * free, size_t * total) {
@ -734,16 +763,16 @@ public:
rpc_server(ggml_backend_t backend) : backend(backend) {}
~rpc_server();
bool alloc_buffer(const std::vector<uint8_t> & input, std::vector<uint8_t> & output);
void get_alignment(std::vector<uint8_t> & output);
void get_max_size(std::vector<uint8_t> & output);
bool buffer_get_base(const std::vector<uint8_t> & input, std::vector<uint8_t> & output);
bool free_buffer(const std::vector<uint8_t> & input);
bool buffer_clear(const std::vector<uint8_t> & input);
void alloc_buffer(const rpc_msg_alloc_buffer_req & request, rpc_msg_alloc_buffer_rsp & response);
void get_alignment(rpc_msg_get_alignment_rsp & response);
void get_max_size(rpc_msg_get_max_size_rsp & response);
bool buffer_get_base(const rpc_msg_buffer_get_base_req & request, rpc_msg_buffer_get_base_rsp & response);
bool free_buffer(const rpc_msg_free_buffer_req & request);
bool buffer_clear(const rpc_msg_buffer_clear_req & request);
bool set_tensor(const std::vector<uint8_t> & input);
bool get_tensor(const std::vector<uint8_t> & input, std::vector<uint8_t> & output);
bool copy_tensor(const std::vector<uint8_t> & input, std::vector<uint8_t> & output);
bool graph_compute(const std::vector<uint8_t> & input, std::vector<uint8_t> & output);
bool get_tensor(const rpc_msg_get_tensor_req & request, std::vector<uint8_t> & response);
bool copy_tensor(const rpc_msg_copy_tensor_req & request, rpc_msg_copy_tensor_rsp & response);
bool graph_compute(const std::vector<uint8_t> & input, rpc_msg_graph_compute_rsp & response);
private:
ggml_tensor * deserialize_tensor(struct ggml_context * ctx, const rpc_tensor * tensor);
@ -757,80 +786,50 @@ private:
std::unordered_set<ggml_backend_buffer_t> buffers;
};
bool rpc_server::alloc_buffer(const std::vector<uint8_t> & input, std::vector<uint8_t> & output) {
// input serialization format: | size (8 bytes) |
if (input.size() != sizeof(uint64_t)) {
return false;
}
uint64_t size;
memcpy(&size, input.data(), sizeof(size));
void rpc_server::alloc_buffer(const rpc_msg_alloc_buffer_req & request, rpc_msg_alloc_buffer_rsp & response) {
ggml_backend_buffer_type_t buft = ggml_backend_get_default_buffer_type(backend);
ggml_backend_buffer_t buffer = ggml_backend_buft_alloc_buffer(buft, size);
uint64_t remote_ptr = 0;
uint64_t remote_size = 0;
ggml_backend_buffer_t buffer = ggml_backend_buft_alloc_buffer(buft, request.size);
response.remote_ptr = 0;
response.remote_size = 0;
if (buffer != nullptr) {
remote_ptr = reinterpret_cast<uint64_t>(buffer);
remote_size = buffer->size;
GGML_PRINT_DEBUG("[%s] size: %" PRIu64 " -> remote_ptr: %" PRIx64 ", remote_size: %" PRIu64 "\n", __func__, size, remote_ptr, remote_size);
response.remote_ptr = reinterpret_cast<uint64_t>(buffer);
response.remote_size = buffer->size;
GGML_PRINT_DEBUG("[%s] size: %" PRIu64 " -> remote_ptr: %" PRIx64 ", remote_size: %" PRIu64 "\n", __func__, request.size, response.remote_ptr, response.remote_size);
buffers.insert(buffer);
} else {
GGML_PRINT_DEBUG("[%s] size: %" PRIu64 " -> failed\n", __func__, size);
GGML_PRINT_DEBUG("[%s] size: %" PRIu64 " -> failed\n", __func__, request.size);
}
// output serialization format: | remote_ptr (8 bytes) | remote_size (8 bytes) |
output.resize(2*sizeof(uint64_t), 0);
memcpy(output.data(), &remote_ptr, sizeof(remote_ptr));
memcpy(output.data() + sizeof(uint64_t), &remote_size, sizeof(remote_size));
return true;
}
void rpc_server::get_alignment(std::vector<uint8_t> & output) {
void rpc_server::get_alignment(rpc_msg_get_alignment_rsp & response) {
ggml_backend_buffer_type_t buft = ggml_backend_get_default_buffer_type(backend);
size_t alignment = ggml_backend_buft_get_alignment(buft);
GGML_PRINT_DEBUG("[%s] alignment: %lu\n", __func__, alignment);
// output serialization format: | alignment (8 bytes) |
output.resize(sizeof(uint64_t), 0);
memcpy(output.data(), &alignment, sizeof(alignment));
response.alignment = alignment;
}
void rpc_server::get_max_size(std::vector<uint8_t> & output) {
void rpc_server::get_max_size(rpc_msg_get_max_size_rsp & response) {
ggml_backend_buffer_type_t buft = ggml_backend_get_default_buffer_type(backend);
size_t max_size = ggml_backend_buft_get_max_size(buft);
GGML_PRINT_DEBUG("[%s] max_size: %lu\n", __func__, max_size);
// output serialization format: | max_size (8 bytes) |
output.resize(sizeof(uint64_t), 0);
memcpy(output.data(), &max_size, sizeof(max_size));
response.max_size = max_size;
}
bool rpc_server::buffer_get_base(const std::vector<uint8_t> & input, std::vector<uint8_t> & output) {
// input serialization format: | remote_ptr (8 bytes) |
if (input.size() != sizeof(uint64_t)) {
return false;
}
uint64_t remote_ptr;
memcpy(&remote_ptr, input.data(), sizeof(remote_ptr));
GGML_PRINT_DEBUG("[%s] remote_ptr: %" PRIx64 "\n", __func__, remote_ptr);
ggml_backend_buffer_t buffer = reinterpret_cast<ggml_backend_buffer_t>(remote_ptr);
bool rpc_server::buffer_get_base(const rpc_msg_buffer_get_base_req & request, rpc_msg_buffer_get_base_rsp & response) {
GGML_PRINT_DEBUG("[%s] remote_ptr: %" PRIx64 "\n", __func__, request.remote_ptr);
ggml_backend_buffer_t buffer = reinterpret_cast<ggml_backend_buffer_t>(request.remote_ptr);
if (buffers.find(buffer) == buffers.end()) {
GGML_PRINT_DEBUG("[%s] buffer not found\n", __func__);
return false;
}
void * base = ggml_backend_buffer_get_base(buffer);
// output serialization format: | base_ptr (8 bytes) |
uint64_t base_ptr = reinterpret_cast<uint64_t>(base);
output.resize(sizeof(uint64_t), 0);
memcpy(output.data(), &base_ptr, sizeof(base_ptr));
response.base_ptr = reinterpret_cast<uint64_t>(base);
return true;
}
bool rpc_server::free_buffer(const std::vector<uint8_t> & input) {
// input serialization format: | remote_ptr (8 bytes) |
if (input.size() != sizeof(uint64_t)) {
return false;
}
uint64_t remote_ptr;
memcpy(&remote_ptr, input.data(), sizeof(remote_ptr));
GGML_PRINT_DEBUG("[%s] remote_ptr: %" PRIx64 "\n", __func__, remote_ptr);
ggml_backend_buffer_t buffer = reinterpret_cast<ggml_backend_buffer_t>(remote_ptr);
bool rpc_server::free_buffer(const rpc_msg_free_buffer_req & request) {
GGML_PRINT_DEBUG("[%s] remote_ptr: %" PRIx64 "\n", __func__, request.remote_ptr);
ggml_backend_buffer_t buffer = reinterpret_cast<ggml_backend_buffer_t>(request.remote_ptr);
if (buffers.find(buffer) == buffers.end()) {
GGML_PRINT_DEBUG("[%s] buffer not found\n", __func__);
return false;
@ -840,22 +839,14 @@ bool rpc_server::free_buffer(const std::vector<uint8_t> & input) {
return true;
}
bool rpc_server::buffer_clear(const std::vector<uint8_t> & input) {
// input serialization format: | remote_ptr (8 bytes) | value (1 byte) |
if (input.size() != sizeof(uint64_t) + sizeof(uint8_t)) {
return false;
}
uint64_t remote_ptr;
memcpy(&remote_ptr, input.data(), sizeof(remote_ptr));
uint8_t value;
memcpy(&value, input.data() + sizeof(uint64_t), sizeof(value));
GGML_PRINT_DEBUG("[%s] remote_ptr: %" PRIx64 ", value: %u\n", __func__, remote_ptr, value);
ggml_backend_buffer_t buffer = reinterpret_cast<ggml_backend_buffer_t>(remote_ptr);
bool rpc_server::buffer_clear(const rpc_msg_buffer_clear_req & request) {
GGML_PRINT_DEBUG("[%s] remote_ptr: %" PRIx64 ", value: %u\n", __func__, request.remote_ptr, request.value);
ggml_backend_buffer_t buffer = reinterpret_cast<ggml_backend_buffer_t>(request.remote_ptr);
if (buffers.find(buffer) == buffers.end()) {
GGML_PRINT_DEBUG("[%s] buffer not found\n", __func__);
return false;
}
ggml_backend_buffer_clear(buffer, value);
ggml_backend_buffer_clear(buffer, request.value);
return true;
}
@ -930,74 +921,55 @@ bool rpc_server::set_tensor(const std::vector<uint8_t> & input) {
return true;
}
bool rpc_server::get_tensor(const std::vector<uint8_t> & input, std::vector<uint8_t> & output) {
// serialization format: | rpc_tensor | offset (8 bytes) | size (8 bytes) |
if (input.size() != sizeof(rpc_tensor) + 2*sizeof(uint64_t)) {
return false;
}
const rpc_tensor * in_tensor = (const rpc_tensor *)input.data();
uint64_t offset;
memcpy(&offset, input.data() + sizeof(rpc_tensor), sizeof(offset));
uint64_t size;
memcpy(&size, input.data() + sizeof(rpc_tensor) + sizeof(offset), sizeof(size));
bool rpc_server::get_tensor(const rpc_msg_get_tensor_req & request, std::vector<uint8_t> & response) {
struct ggml_init_params params {
/*.mem_size =*/ ggml_tensor_overhead(),
/*.mem_buffer =*/ NULL,
/*.no_alloc =*/ true,
};
struct ggml_context * ctx = ggml_init(params);
ggml_tensor * tensor = deserialize_tensor(ctx, in_tensor);
ggml_tensor * tensor = deserialize_tensor(ctx, &request.tensor);
if (tensor == nullptr) {
GGML_PRINT_DEBUG("[%s] error deserializing tensor\n", __func__);
ggml_free(ctx);
return false;
}
GGML_PRINT_DEBUG("[%s] buffer: %p, data: %p, offset: %" PRIu64 ", size: %" PRIu64 "\n", __func__, (void*)tensor->buffer, tensor->data, offset, size);
GGML_PRINT_DEBUG("[%s] buffer: %p, data: %p, offset: %" PRIu64 ", size: %" PRIu64 "\n", __func__, (void*)tensor->buffer, tensor->data, request.offset, request.size);
// sanitize tensor->data
{
const size_t p0 = (size_t) ggml_backend_buffer_get_base(tensor->buffer);
const size_t p1 = p0 + ggml_backend_buffer_get_size(tensor->buffer);
if (in_tensor->data + offset < p0 || in_tensor->data + offset >= p1 || size > (p1 - in_tensor->data - offset)) {
GGML_ABORT("[%s] tensor->data out of bounds\n", __func__);
if (request.tensor.data + request.offset < p0 ||
request.tensor.data + request.offset >= p1 ||
request.size > (p1 - request.tensor.data - request.offset)) {
GGML_ABORT("[%s] tensor->data out of bounds\n", __func__);
}
}
// output serialization format: | data (size bytes) |
output.resize(size, 0);
ggml_backend_tensor_get(tensor, output.data(), offset, size);
response.resize(request.size, 0);
ggml_backend_tensor_get(tensor, response.data(), request.offset, request.size);
ggml_free(ctx);
return true;
}
bool rpc_server::copy_tensor(const std::vector<uint8_t> & input, std::vector<uint8_t> & output) {
// serialization format: | rpc_tensor src | rpc_tensor dst |
if (input.size() != 2*sizeof(rpc_tensor)) {
return false;
}
const rpc_tensor * rpc_src = (const rpc_tensor *)input.data();
const rpc_tensor * rpc_dst = (const rpc_tensor *)(input.data() + sizeof(rpc_src));
bool rpc_server::copy_tensor(const rpc_msg_copy_tensor_req & request, rpc_msg_copy_tensor_rsp & response) {
struct ggml_init_params params {
/*.mem_size =*/ 2*ggml_tensor_overhead(),
/*.mem_buffer =*/ NULL,
/*.no_alloc =*/ true,
};
struct ggml_context * ctx = ggml_init(params);
ggml_tensor * src = deserialize_tensor(ctx, rpc_src);
ggml_tensor * dst = deserialize_tensor(ctx, rpc_dst);
ggml_tensor * src = deserialize_tensor(ctx, &request.src);
ggml_tensor * dst = deserialize_tensor(ctx, &request.dst);
if (src == nullptr || dst == nullptr) {
GGML_PRINT_DEBUG("[%s] error deserializing tensors\n", __func__);
ggml_free(ctx);
return false;
}
GGML_PRINT_DEBUG("[%s] src->buffer: %p, dst->buffer: %p\n", __func__, (void*)src->buffer, (void*)dst->buffer);
bool result = ggml_backend_buffer_copy_tensor(src, dst);
// output serialization format: | result (1 byte) |
output.resize(1, 0);
output[0] = result;
response.result = ggml_backend_buffer_copy_tensor(src, dst);
ggml_free(ctx);
return true;
}
@ -1026,7 +998,7 @@ ggml_tensor * rpc_server::create_node(uint64_t id,
return result;
}
bool rpc_server::graph_compute(const std::vector<uint8_t> & input, std::vector<uint8_t> & output) {
bool rpc_server::graph_compute(const std::vector<uint8_t> & input, rpc_msg_graph_compute_rsp & response) {
// serialization format:
// | n_nodes (4 bytes) | nodes (n_nodes * sizeof(uint64_t) | n_tensors (4 bytes) | tensors (n_tensors * sizeof(rpc_tensor)) |
if (input.size() < sizeof(uint32_t)) {
@ -1066,9 +1038,7 @@ bool rpc_server::graph_compute(const std::vector<uint8_t> & input, std::vector<u
graph->nodes[i] = create_node(id, ctx, tensor_ptrs, tensor_map);
}
ggml_status status = ggml_backend_graph_compute(backend, graph);
// output serialization format: | status (1 byte) |
output.resize(1, 0);
output[0] = status;
response.result = status;
ggml_free(ctx);
return true;
}
@ -1091,85 +1061,153 @@ static void rpc_serve_client(ggml_backend_t backend, sockfd_t sockfd, size_t fre
fprintf(stderr, "Unknown command: %d\n", cmd);
break;
}
std::vector<uint8_t> input;
std::vector<uint8_t> output;
uint64_t input_size;
if (!recv_data(sockfd, &input_size, sizeof(input_size))) {
break;
}
try {
input.resize(input_size);
} catch (const std::bad_alloc & e) {
fprintf(stderr, "Failed to allocate input buffer of size %" PRIu64 "\n", input_size);
break;
}
if (!recv_data(sockfd, input.data(), input_size)) {
break;
}
bool ok = true;
switch (cmd) {
case RPC_CMD_ALLOC_BUFFER: {
ok = server.alloc_buffer(input, output);
rpc_msg_alloc_buffer_req request;
if (!recv_msg(sockfd, &request, sizeof(request))) {
return;
}
rpc_msg_alloc_buffer_rsp response;
server.alloc_buffer(request, response);
if (!send_msg(sockfd, &response, sizeof(response))) {
return;
}
break;
}
case RPC_CMD_GET_ALIGNMENT: {
server.get_alignment(output);
if (!recv_msg(sockfd, nullptr, 0)) {
return;
}
rpc_msg_get_alignment_rsp response;
server.get_alignment(response);
if (!send_msg(sockfd, &response, sizeof(response))) {
return;
}
break;
}
case RPC_CMD_GET_MAX_SIZE: {
server.get_max_size(output);
if (!recv_msg(sockfd, nullptr, 0)) {
return;
}
rpc_msg_get_max_size_rsp response;
server.get_max_size(response);
if (!send_msg(sockfd, &response, sizeof(response))) {
return;
}
break;
}
case RPC_CMD_BUFFER_GET_BASE: {
ok = server.buffer_get_base(input, output);
rpc_msg_buffer_get_base_req request;
if (!recv_msg(sockfd, &request, sizeof(request))) {
return;
}
rpc_msg_buffer_get_base_rsp response;
if (!server.buffer_get_base(request, response)) {
return;
}
if (!send_msg(sockfd, &response, sizeof(response))) {
return;
}
break;
}
case RPC_CMD_FREE_BUFFER: {
ok = server.free_buffer(input);
rpc_msg_free_buffer_req request;
if (!recv_msg(sockfd, &request, sizeof(request))) {
return;
}
if (!server.free_buffer(request)) {
return;
}
if (!send_msg(sockfd, nullptr, 0)) {
return;
}
break;
}
case RPC_CMD_BUFFER_CLEAR: {
ok = server.buffer_clear(input);
rpc_msg_buffer_clear_req request;
if (!recv_msg(sockfd, &request, sizeof(request))) {
return;
}
if (!server.buffer_clear(request)) {
return;
}
if (!send_msg(sockfd, nullptr, 0)) {
return;
}
break;
}
case RPC_CMD_SET_TENSOR: {
ok = server.set_tensor(input);
std::vector<uint8_t> input;
if (!recv_msg(sockfd, input)) {
return;
}
if (!server.set_tensor(input)) {
return;
}
if (!send_msg(sockfd, nullptr, 0)) {
return;
}
break;
}
case RPC_CMD_GET_TENSOR: {
ok = server.get_tensor(input, output);
rpc_msg_get_tensor_req request;
if (!recv_msg(sockfd, &request, sizeof(request))) {
return;
}
std::vector<uint8_t> response;
if (!server.get_tensor(request, response)) {
return;
}
if (!send_msg(sockfd, response.data(), response.size())) {
return;
}
break;
}
case RPC_CMD_COPY_TENSOR: {
ok = server.copy_tensor(input, output);
rpc_msg_copy_tensor_req request;
if (!recv_msg(sockfd, &request, sizeof(request))) {
return;
}
rpc_msg_copy_tensor_rsp response;
if (!server.copy_tensor(request, response)) {
return;
}
if (!send_msg(sockfd, &response, sizeof(response))) {
return;
}
break;
}
case RPC_CMD_GRAPH_COMPUTE: {
ok = server.graph_compute(input, output);
std::vector<uint8_t> input;
if (!recv_msg(sockfd, input)) {
return;
}
rpc_msg_graph_compute_rsp response;
if (!server.graph_compute(input, response)) {
return;
}
if (!send_msg(sockfd, &response, sizeof(response))) {
return;
}
break;
}
case RPC_CMD_GET_DEVICE_MEMORY: {
// output serialization format: | free (8 bytes) | total (8 bytes) |
output.resize(2*sizeof(uint64_t), 0);
memcpy(output.data(), &free_mem, sizeof(free_mem));
memcpy(output.data() + sizeof(uint64_t), &total_mem, sizeof(total_mem));
if (!recv_msg(sockfd, nullptr, 0)) {
return;
}
rpc_msg_get_device_memory_rsp response;
response.free_mem = free_mem;
response.total_mem = total_mem;
if (!send_msg(sockfd, &response, sizeof(response))) {
return;
}
break;
}
default: {
fprintf(stderr, "Unknown command: %d\n", cmd);
ok = false;
return;
}
}
if (!ok) {
break;
}
uint64_t output_size = output.size();
if (!send_data(sockfd, &output_size, sizeof(output_size))) {
break;
}
if (!send_data(sockfd, output.data(), output_size)) {
break;
}
}
}

2689
ggml/src/ggml-sycl.cpp
View File

File diff suppressed because it is too large Load Diff

136
ggml/src/ggml-sycl/mmvq.cpp
View File

@ -1,6 +1,6 @@
#include "mmvq.hpp"
#include "vecdotq.hpp"
#include <cassert>
template <int qk, int qi, typename block_q_t, int vdr, vec_dot_q_sycl_t vec_dot_q_sycl>
static void mul_mat_vec_q(const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, const int ncols, const int nrows,
@ -13,7 +13,8 @@ static void mul_mat_vec_q(const void * __restrict__ vx, const void * __restrict_
}
const int blocks_per_row = ncols / qk;
const int blocks_per_warp = vdr * WARP_SIZE / qi;
const int blocks_per_warp = vdr * QK_WARP_SIZE / qi;
assert(blocks_per_warp>0);
// partial sum for each thread
float tmp = 0.0f;
@ -37,7 +38,7 @@ static void mul_mat_vec_q(const void * __restrict__ vx, const void * __restrict_
// sum up partial sums and write back result
#pragma unroll
for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) {
for (int mask = QK_WARP_SIZE / 2; mask > 0; mask >>= 1) {
tmp +=
dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask);
}
@ -61,7 +62,8 @@ static void mul_mat_vec_q_iq2_xxs_q8_1(const void *__restrict__ vx,
}
const int blocks_per_row = ncols / qk;
const int blocks_per_warp = vdr * WARP_SIZE / qi;
const int blocks_per_warp = vdr * QK_WARP_SIZE / qi;
assert(blocks_per_warp>0);
// partial sum for each thread
float tmp = 0.0f;
@ -85,7 +87,7 @@ static void mul_mat_vec_q_iq2_xxs_q8_1(const void *__restrict__ vx,
// sum up partial sums and write back result
#pragma unroll
for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) {
for (int mask = QK_WARP_SIZE / 2; mask > 0; mask >>= 1) {
tmp +=
dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask);
}
@ -109,8 +111,8 @@ static void mul_mat_vec_q_iq2_xs_q8_1(const void *__restrict__ vx,
}
const int blocks_per_row = ncols / qk;
const int blocks_per_warp = vdr * WARP_SIZE / qi;
const int blocks_per_warp = vdr * QK_WARP_SIZE / qi;
assert(blocks_per_warp>0);
// partial sum for each thread
float tmp = 0.0f;
@ -133,7 +135,7 @@ static void mul_mat_vec_q_iq2_xs_q8_1(const void *__restrict__ vx,
// sum up partial sums and write back result
#pragma unroll
for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) {
for (int mask = QK_WARP_SIZE / 2; mask > 0; mask >>= 1) {
tmp +=
dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask);
}
@ -157,8 +159,8 @@ static void mul_mat_vec_q_iq2_s_q8_1(const void *__restrict__ vx,
}
const int blocks_per_row = ncols / qk;
const int blocks_per_warp = vdr * WARP_SIZE / qi;
const int blocks_per_warp = vdr * QK_WARP_SIZE / qi;
assert(blocks_per_warp>0);
// partial sum for each thread
float tmp = 0.0f;
@ -181,7 +183,7 @@ static void mul_mat_vec_q_iq2_s_q8_1(const void *__restrict__ vx,
// sum up partial sums and write back result
#pragma unroll
for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) {
for (int mask = QK_WARP_SIZE / 2; mask > 0; mask >>= 1) {
tmp +=
dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask);
}
@ -205,8 +207,8 @@ static void mul_mat_vec_q_iq3_xxs_q8_1(const void *__restrict__ vx,
}
const int blocks_per_row = ncols / qk;
const int blocks_per_warp = vdr * WARP_SIZE / qi;
const int blocks_per_warp = vdr * QK_WARP_SIZE / qi;
assert(blocks_per_warp>0);
// partial sum for each thread
float tmp = 0.0f;
@ -229,7 +231,7 @@ static void mul_mat_vec_q_iq3_xxs_q8_1(const void *__restrict__ vx,
// sum up partial sums and write back result
#pragma unroll
for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) {
for (int mask = QK_WARP_SIZE / 2; mask > 0; mask >>= 1) {
tmp +=
dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask);
}
@ -253,8 +255,8 @@ static void mul_mat_vec_q_iq3_s_q8_1(const void *__restrict__ vx,
}
const int blocks_per_row = ncols / qk;
const int blocks_per_warp = vdr * WARP_SIZE / qi;
const int blocks_per_warp = vdr * QK_WARP_SIZE / qi;
assert(blocks_per_warp>0);
// partial sum for each thread
float tmp = 0.0f;
@ -277,7 +279,7 @@ static void mul_mat_vec_q_iq3_s_q8_1(const void *__restrict__ vx,
// sum up partial sums and write back result
#pragma unroll
for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) {
for (int mask = QK_WARP_SIZE / 2; mask > 0; mask >>= 1) {
tmp +=
dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask);
}
@ -301,8 +303,8 @@ static void mul_mat_vec_q_iq1_s_q8_1(const void *__restrict__ vx,
}
const int blocks_per_row = ncols / qk;
const int blocks_per_warp = vdr * WARP_SIZE / qi;
const int blocks_per_warp = vdr * QK_WARP_SIZE / qi;
assert(blocks_per_warp>0);
// partial sum for each thread
float tmp = 0.0f;
@ -325,7 +327,7 @@ static void mul_mat_vec_q_iq1_s_q8_1(const void *__restrict__ vx,
// sum up partial sums and write back result
#pragma unroll
for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) {
for (int mask = QK_WARP_SIZE / 2; mask > 0; mask >>= 1) {
tmp +=
dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask);
}
@ -349,8 +351,8 @@ static void mul_mat_vec_q_iq1_m_q8_1(const void *__restrict__ vx,
}
const int blocks_per_row = ncols / qk;
const int blocks_per_warp = vdr * WARP_SIZE / qi;
const int blocks_per_warp = vdr * QK_WARP_SIZE / qi;
assert(blocks_per_warp>0);
// partial sum for each thread
float tmp = 0.0f;
@ -373,7 +375,7 @@ static void mul_mat_vec_q_iq1_m_q8_1(const void *__restrict__ vx,
// sum up partial sums and write back result
#pragma unroll
for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) {
for (int mask = QK_WARP_SIZE / 2; mask > 0; mask >>= 1) {
tmp +=
dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask);
}
@ -397,8 +399,8 @@ static void mul_mat_vec_q_iq4_nl_q8_1(const void *__restrict__ vx,
}
const int blocks_per_row = ncols / qk;
const int blocks_per_warp = vdr * WARP_SIZE / qi;
const int blocks_per_warp = vdr * QK_WARP_SIZE / qi;
assert(blocks_per_warp>0);
// partial sum for each thread
float tmp = 0.0f;
@ -421,7 +423,7 @@ static void mul_mat_vec_q_iq4_nl_q8_1(const void *__restrict__ vx,
// sum up partial sums and write back result
#pragma unroll
for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) {
for (int mask = QK_WARP_SIZE / 2; mask > 0; mask >>= 1) {
tmp +=
dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask);
}
@ -446,8 +448,8 @@ static void mul_mat_vec_q_iq4_xs_q8_1(const void *__restrict__ vx,
}
const int blocks_per_row = ncols / qk;
const int blocks_per_warp = vdr * WARP_SIZE / qi;
const int blocks_per_warp = vdr * QK_WARP_SIZE / qi;
assert(blocks_per_warp>0);
// partial sum for each thread
float tmp = 0.0f;
@ -470,7 +472,7 @@ static void mul_mat_vec_q_iq4_xs_q8_1(const void *__restrict__ vx,
// sum up partial sums and write back result
#pragma unroll
for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) {
for (int mask = QK_WARP_SIZE / 2; mask > 0; mask >>= 1) {
tmp +=
dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask);
}
@ -487,7 +489,7 @@ static void mul_mat_vec_q4_0_q8_1_sycl(const void *vx, const void *vy,
GGML_ASSERT(ncols % QK4_0 == 0);
const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
const sycl::range<3> block_nums(1, 1, block_num_y);
const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE);
{
stream->submit([&](sycl::handler &cgh) {
@ -495,7 +497,7 @@ static void mul_mat_vec_q4_0_q8_1_sycl(const void *vx, const void *vy,
cgh.parallel_for(
sycl::nd_range<3>(block_nums * block_dims, block_dims),
[=](sycl::nd_item<3> item_ct1)
[[intel::reqd_sub_group_size(WARP_SIZE)]] {
[[intel::reqd_sub_group_size(QK_WARP_SIZE)]] {
mul_mat_vec_q<QK4_0, QI4_0, block_q4_0,
VDR_Q4_0_Q8_1_MMVQ, vec_dot_q4_0_q8_1>(
vx, vy, dst, ncols, nrows, item_ct1);
@ -511,7 +513,7 @@ static void mul_mat_vec_q4_1_q8_1_sycl(const void *vx, const void *vy,
GGML_ASSERT(ncols % QK4_1 == 0);
const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
const sycl::range<3> block_nums(1, 1, block_num_y);
const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE);
{
stream->submit([&](sycl::handler &cgh) {
@ -519,7 +521,7 @@ static void mul_mat_vec_q4_1_q8_1_sycl(const void *vx, const void *vy,
cgh.parallel_for(
sycl::nd_range<3>(block_nums * block_dims, block_dims),
[=](sycl::nd_item<3> item_ct1)
[[intel::reqd_sub_group_size(WARP_SIZE)]] {
[[intel::reqd_sub_group_size(QK_WARP_SIZE)]] {
mul_mat_vec_q<QK4_0, QI4_1, block_q4_1,
VDR_Q4_1_Q8_1_MMVQ, vec_dot_q4_1_q8_1>(
vx, vy, dst, ncols, nrows, item_ct1);
@ -535,7 +537,7 @@ static void mul_mat_vec_q5_0_q8_1_sycl(const void *vx, const void *vy,
GGML_ASSERT(ncols % QK5_0 == 0);
const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
const sycl::range<3> block_nums(1, 1, block_num_y);
const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE);
{
stream->submit([&](sycl::handler &cgh) {
@ -543,7 +545,7 @@ static void mul_mat_vec_q5_0_q8_1_sycl(const void *vx, const void *vy,
cgh.parallel_for(
sycl::nd_range<3>(block_nums * block_dims, block_dims),
[=](sycl::nd_item<3> item_ct1)
[[intel::reqd_sub_group_size(WARP_SIZE)]] {
[[intel::reqd_sub_group_size(QK_WARP_SIZE)]] {
mul_mat_vec_q<QK5_0, QI5_0, block_q5_0,
VDR_Q5_0_Q8_1_MMVQ, vec_dot_q5_0_q8_1>(
vx, vy, dst, ncols, nrows, item_ct1);
@ -559,7 +561,7 @@ static void mul_mat_vec_q5_1_q8_1_sycl(const void *vx, const void *vy,
GGML_ASSERT(ncols % QK5_1 == 0);
const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
const sycl::range<3> block_nums(1, 1, block_num_y);
const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE);
{
stream->submit([&](sycl::handler &cgh) {
@ -567,7 +569,7 @@ static void mul_mat_vec_q5_1_q8_1_sycl(const void *vx, const void *vy,
cgh.parallel_for(
sycl::nd_range<3>(block_nums * block_dims, block_dims),
[=](sycl::nd_item<3> item_ct1)
[[intel::reqd_sub_group_size(WARP_SIZE)]] {
[[intel::reqd_sub_group_size(QK_WARP_SIZE)]] {
mul_mat_vec_q<QK5_1, QI5_1, block_q5_1,
VDR_Q5_1_Q8_1_MMVQ, vec_dot_q5_1_q8_1>(
vx, vy, dst, ncols, nrows, item_ct1);
@ -583,7 +585,7 @@ static void mul_mat_vec_q8_0_q8_1_sycl(const void *vx, const void *vy,
GGML_ASSERT(ncols % QK8_0 == 0);
const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
const sycl::range<3> block_nums(1, 1, block_num_y);
const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE);
{
stream->submit([&](sycl::handler &cgh) {
@ -591,7 +593,7 @@ static void mul_mat_vec_q8_0_q8_1_sycl(const void *vx, const void *vy,
cgh.parallel_for(
sycl::nd_range<3>(block_nums * block_dims, block_dims),
[=](sycl::nd_item<3> item_ct1)
[[intel::reqd_sub_group_size(WARP_SIZE)]] {
[[intel::reqd_sub_group_size(QK_WARP_SIZE)]] {
mul_mat_vec_q<QK8_0, QI8_0, block_q8_0,
VDR_Q8_0_Q8_1_MMVQ, vec_dot_q8_0_q8_1>(
vx, vy, dst, ncols, nrows, item_ct1);
@ -607,7 +609,7 @@ static void mul_mat_vec_q2_K_q8_1_sycl(const void *vx, const void *vy,
GGML_ASSERT(ncols % QK_K == 0);
const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
const sycl::range<3> block_nums(1, 1, block_num_y);
const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE);
{
stream->submit([&](sycl::handler &cgh) {
@ -615,7 +617,7 @@ static void mul_mat_vec_q2_K_q8_1_sycl(const void *vx, const void *vy,
cgh.parallel_for(
sycl::nd_range<3>(block_nums * block_dims, block_dims),
[=](sycl::nd_item<3> item_ct1)
[[intel::reqd_sub_group_size(WARP_SIZE)]] {
[[intel::reqd_sub_group_size(QK_WARP_SIZE)]] {
mul_mat_vec_q<QK_K, QI2_K, block_q2_K,
VDR_Q2_K_Q8_1_MMVQ, vec_dot_q2_K_q8_1>(
vx, vy, dst, ncols, nrows, item_ct1);
@ -631,7 +633,7 @@ static void mul_mat_vec_q3_K_q8_1_sycl(const void *vx, const void *vy,
GGML_ASSERT(ncols % QK_K == 0);
const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
const sycl::range<3> block_nums(1, 1, block_num_y);
const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE);
{
stream->submit([&](sycl::handler &cgh) {
@ -639,7 +641,7 @@ static void mul_mat_vec_q3_K_q8_1_sycl(const void *vx, const void *vy,
cgh.parallel_for(
sycl::nd_range<3>(block_nums * block_dims, block_dims),
[=](sycl::nd_item<3> item_ct1)
[[intel::reqd_sub_group_size(WARP_SIZE)]] {
[[intel::reqd_sub_group_size(QK_WARP_SIZE)]] {
mul_mat_vec_q<QK_K, QI3_K, block_q3_K,
VDR_Q3_K_Q8_1_MMVQ, vec_dot_q3_K_q8_1>(
vx, vy, dst, ncols, nrows, item_ct1);
@ -655,7 +657,7 @@ static void mul_mat_vec_q4_K_q8_1_sycl(const void *vx, const void *vy,
GGML_ASSERT(ncols % QK_K == 0);
const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
const sycl::range<3> block_nums(1, 1, block_num_y);
const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE);
{
stream->submit([&](sycl::handler &cgh) {
@ -663,7 +665,7 @@ static void mul_mat_vec_q4_K_q8_1_sycl(const void *vx, const void *vy,
cgh.parallel_for(
sycl::nd_range<3>(block_nums * block_dims, block_dims),
[=](sycl::nd_item<3> item_ct1)
[[intel::reqd_sub_group_size(WARP_SIZE)]] {
[[intel::reqd_sub_group_size(QK_WARP_SIZE)]] {
mul_mat_vec_q<QK_K, QI4_K, block_q4_K,
VDR_Q4_K_Q8_1_MMVQ, vec_dot_q4_K_q8_1>(
vx, vy, dst, ncols, nrows, item_ct1);
@ -679,7 +681,7 @@ static void mul_mat_vec_q5_K_q8_1_sycl(const void *vx, const void *vy,
GGML_ASSERT(ncols % QK_K == 0);
const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
const sycl::range<3> block_nums(1, 1, block_num_y);
const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE);
{
stream->submit([&](sycl::handler &cgh) {
@ -687,7 +689,7 @@ static void mul_mat_vec_q5_K_q8_1_sycl(const void *vx, const void *vy,
cgh.parallel_for(
sycl::nd_range<3>(block_nums * block_dims, block_dims),
[=](sycl::nd_item<3> item_ct1)
[[intel::reqd_sub_group_size(WARP_SIZE)]] {
[[intel::reqd_sub_group_size(QK_WARP_SIZE)]] {
mul_mat_vec_q<QK_K, QI5_K, block_q5_K,
VDR_Q5_K_Q8_1_MMVQ, vec_dot_q5_K_q8_1>(
vx, vy, dst, ncols, nrows, item_ct1);
@ -703,7 +705,7 @@ static void mul_mat_vec_q6_K_q8_1_sycl(const void *vx, const void *vy,
GGML_ASSERT(ncols % QK_K == 0);
const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
const sycl::range<3> block_nums(1, 1, block_num_y);
const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE);
{
stream->submit([&](sycl::handler &cgh) {
@ -711,7 +713,7 @@ static void mul_mat_vec_q6_K_q8_1_sycl(const void *vx, const void *vy,
cgh.parallel_for(
sycl::nd_range<3>(block_nums * block_dims, block_dims),
[=](sycl::nd_item<3> item_ct1)
[[intel::reqd_sub_group_size(WARP_SIZE)]] {
[[intel::reqd_sub_group_size(QK_WARP_SIZE)]] {
mul_mat_vec_q<QK_K, QI6_K, block_q6_K,
VDR_Q6_K_Q8_1_MMVQ, vec_dot_q6_K_q8_1>(
vx, vy, dst, ncols, nrows, item_ct1);
@ -728,13 +730,13 @@ static void mul_mat_vec_iq2_xxs_q8_1_sycl(const void *vx, const void *vy,
GGML_ASSERT(ncols % QK_K == 0);
const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
const sycl::range<3> block_nums(1, 1, block_num_y);
const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE);
{
stream->submit([&](sycl::handler &cgh) {
cgh.parallel_for(
sycl::nd_range<3>(block_nums * block_dims, block_dims),
[=](sycl::nd_item<3> item_ct1)
[[intel::reqd_sub_group_size(WARP_SIZE)]] {
[[intel::reqd_sub_group_size(QK_WARP_SIZE)]] {
mul_mat_vec_q_iq2_xxs_q8_1<QK_K, QI2_XXS/2, block_iq2_xxs, 1>(
vx, vy, dst, ncols, nrows, item_ct1);
});
@ -749,7 +751,7 @@ static void mul_mat_vec_iq2_xs_q8_1_sycl(const void *vx, const void *vy,
GGML_ASSERT(ncols % QK_K == 0);
const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
const sycl::range<3> block_nums(1, 1, block_num_y);
const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE);
{
stream->submit([&](sycl::handler &cgh) {
@ -759,7 +761,7 @@ static void mul_mat_vec_iq2_xs_q8_1_sycl(const void *vx, const void *vy,
cgh.parallel_for(
sycl::nd_range<3>(block_nums * block_dims, block_dims),
[=](sycl::nd_item<3> item_ct1)
[[intel::reqd_sub_group_size(WARP_SIZE)]] {
[[intel::reqd_sub_group_size(QK_WARP_SIZE)]] {
mul_mat_vec_q_iq2_xs_q8_1<QK_K, QI2_XS/2, block_iq2_xs, 1>(
vx, vy, dst, ncols, nrows, item_ct1);
});
@ -774,7 +776,7 @@ static void mul_mat_vec_iq2_s_q8_1_sycl(const void *vx, const void *vy,
GGML_ASSERT(ncols % QK_K == 0);
const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
const sycl::range<3> block_nums(1, 1, block_num_y);
const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE);
{
stream->submit([&](sycl::handler &cgh) {
@ -784,7 +786,7 @@ static void mul_mat_vec_iq2_s_q8_1_sycl(const void *vx, const void *vy,
cgh.parallel_for(
sycl::nd_range<3>(block_nums * block_dims, block_dims),
[=](sycl::nd_item<3> item_ct1)
[[intel::reqd_sub_group_size(WARP_SIZE)]] {
[[intel::reqd_sub_group_size(QK_WARP_SIZE)]] {
mul_mat_vec_q_iq2_s_q8_1<QK_K, QI2_S/2, block_iq2_s, 1>(
vx, vy, dst, ncols, nrows, item_ct1);
});
@ -799,7 +801,7 @@ static void mul_mat_vec_iq3_xxs_q8_1_sycl(const void *vx, const void *vy,
GGML_ASSERT(ncols % QK_K == 0);
const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
const sycl::range<3> block_nums(1, 1, block_num_y);
const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE);
{
stream->submit([&](sycl::handler &cgh) {
@ -809,7 +811,7 @@ static void mul_mat_vec_iq3_xxs_q8_1_sycl(const void *vx, const void *vy,
cgh.parallel_for(
sycl::nd_range<3>(block_nums * block_dims, block_dims),
[=](sycl::nd_item<3> item_ct1)
[[intel::reqd_sub_group_size(WARP_SIZE)]] {
[[intel::reqd_sub_group_size(QK_WARP_SIZE)]] {
mul_mat_vec_q_iq3_xxs_q8_1<QK_K, QI3_XXS/2, block_iq3_xxs, 1>(
vx, vy, dst, ncols, nrows, item_ct1);
});
@ -824,7 +826,7 @@ static void mul_mat_vec_iq3_s_q8_1_sycl(const void *vx, const void *vy,
GGML_ASSERT(ncols % QK_K == 0);
const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
const sycl::range<3> block_nums(1, 1, block_num_y);
const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE);
{
stream->submit([&](sycl::handler &cgh) {
@ -833,7 +835,7 @@ static void mul_mat_vec_iq3_s_q8_1_sycl(const void *vx, const void *vy,
cgh.parallel_for(
sycl::nd_range<3>(block_nums * block_dims, block_dims),
[=](sycl::nd_item<3> item_ct1)
[[intel::reqd_sub_group_size(WARP_SIZE)]] {
[[intel::reqd_sub_group_size(QK_WARP_SIZE)]] {
mul_mat_vec_q_iq3_s_q8_1<QK_K, QI3_S/2, block_iq3_s, 1>(
vx, vy, dst, ncols, nrows, item_ct1);
});
@ -848,7 +850,7 @@ static void mul_mat_vec_iq1_s_q8_1_sycl(const void *vx, const void *vy,
GGML_ASSERT(ncols % QK_K == 0);
const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
const sycl::range<3> block_nums(1, 1, block_num_y);
const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE);
{
stream->submit([&](sycl::handler &cgh) {
@ -858,7 +860,7 @@ static void mul_mat_vec_iq1_s_q8_1_sycl(const void *vx, const void *vy,
cgh.parallel_for(
sycl::nd_range<3>(block_nums * block_dims, block_dims),
[=](sycl::nd_item<3> item_ct1)
[[intel::reqd_sub_group_size(WARP_SIZE)]] {
[[intel::reqd_sub_group_size(QK_WARP_SIZE)]] {
mul_mat_vec_q_iq1_s_q8_1<QK_K, QI1_S, block_iq1_s, 1>(
vx, vy, dst, ncols, nrows, item_ct1);
});
@ -873,13 +875,13 @@ static void mul_mat_vec_iq1_m_q8_1_sycl(const void *vx, const void *vy,
GGML_ASSERT(ncols % QK_K == 0);
const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
const sycl::range<3> block_nums(1, 1, block_num_y);
const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE);
{
stream->submit([&](sycl::handler &cgh) {
cgh.parallel_for(
sycl::nd_range<3>(block_nums * block_dims, block_dims),
[=](sycl::nd_item<3> item_ct1)
[[intel::reqd_sub_group_size(WARP_SIZE)]] {
[[intel::reqd_sub_group_size(QK_WARP_SIZE)]] {
mul_mat_vec_q_iq1_m_q8_1<QK_K, QI1_S, block_iq1_m, 1>(
vx, vy, dst, ncols, nrows, item_ct1);
});
@ -894,14 +896,14 @@ static void mul_mat_vec_iq4_nl_q8_1_sycl(const void *vx, const void *vy,
GGML_ASSERT(ncols % QK4_NL == 0);
const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
const sycl::range<3> block_nums(1, 1, block_num_y);
const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE);
{
stream->submit([&](sycl::handler &cgh) {
cgh.parallel_for(
sycl::nd_range<3>(block_nums * block_dims, block_dims),
[=](sycl::nd_item<3> item_ct1)
[[intel::reqd_sub_group_size(WARP_SIZE)]] {
[[intel::reqd_sub_group_size(QK_WARP_SIZE)]] {
mul_mat_vec_q_iq4_nl_q8_1<QK4_NL, QI4_NL, block_iq4_nl, 2>(
vx, vy, dst, ncols, nrows, item_ct1);
});
@ -916,14 +918,14 @@ static void mul_mat_vec_iq4_xs_q8_1_sycl(const void *vx, const void *vy,
GGML_ASSERT(ncols % QK_K == 0);
const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
const sycl::range<3> block_nums(1, 1, block_num_y);
const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE);
{
stream->submit([&](sycl::handler &cgh) {
cgh.parallel_for(
sycl::nd_range<3>(block_nums * block_dims, block_dims),
[=](sycl::nd_item<3> item_ct1)
[[intel::reqd_sub_group_size(WARP_SIZE)]] {
[[intel::reqd_sub_group_size(QK_WARP_SIZE)]] {
mul_mat_vec_q_iq4_xs_q8_1<QK_K, QI4_XS/4, block_iq4_xs, 1>(
vx, vy, dst, ncols, nrows, item_ct1);
});

284
ggml/src/ggml-vulkan.cpp
View File

@ -1941,7 +1941,7 @@ static vk_device ggml_vk_get_device(size_t idx) {
if (device->fp16) {
device_extensions.push_back("VK_KHR_shader_float16_int8");
}
device->name = device->properties.deviceName.data();
device->name = GGML_VK_NAME + std::to_string(idx);
device_create_info = {
vk::DeviceCreateFlags(),
@ -1968,7 +1968,7 @@ static vk_device ggml_vk_get_device(size_t idx) {
device->buffer_type = {
/* .iface = */ ggml_backend_vk_buffer_type_interface,
/* .device = */ nullptr,
/* .device = */ ggml_backend_reg_dev_get(ggml_backend_vk_reg(), idx),
/* .context = */ new ggml_backend_vk_buffer_type_context{ device->name, device },
};
@ -6378,7 +6378,7 @@ ggml_backend_buffer_type_t ggml_backend_vk_host_buffer_type() {
/* .get_alloc_size = */ ggml_backend_cpu_buffer_type()->iface.get_alloc_size,
/* .is_host = */ ggml_backend_cpu_buffer_type()->iface.is_host,
},
/* .device = */ nullptr,
/* .device = */ ggml_backend_reg_dev_get(ggml_backend_vk_reg(), 0),
/* .context = */ nullptr,
};
@ -6581,9 +6581,135 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
UNUSED(backend);
}
static bool ggml_backend_vk_supports_op(ggml_backend_t backend, const ggml_tensor * op) {
// ggml_backend_vk_context * ctx = (ggml_backend_vk_context *) backend->context;
// TODO: enable async and synchronize
static ggml_backend_i ggml_backend_vk_interface = {
/* .get_name = */ ggml_backend_vk_name,
/* .free = */ ggml_backend_vk_free,
/* .get_default_buffer_type = */ ggml_backend_vk_get_default_buffer_type,
/* .set_tensor_async = */ NULL, // ggml_backend_vk_set_tensor_async,
/* .get_tensor_async = */ NULL, // ggml_backend_vk_get_tensor_async,
/* .cpy_tensor_async = */ NULL, // ggml_backend_vk_cpy_tensor_async,
/* .synchronize = */ NULL, // ggml_backend_vk_synchronize,
/* .graph_plan_create = */ NULL,
/* .graph_plan_free = */ NULL,
/* .graph_plan_update = */ NULL,
/* .graph_plan_compute = */ NULL,
/* .graph_compute = */ ggml_backend_vk_graph_compute,
/* .supports_op = */ NULL,
/* .supports_buft = */ NULL,
/* .offload_op = */ NULL,
/* .event_record = */ NULL,
/* .event_wait = */ NULL,
};
static ggml_guid_t ggml_backend_vk_guid() {
static ggml_guid guid = { 0xb8, 0xf7, 0x4f, 0x86, 0x40, 0x3c, 0xe1, 0x02, 0x91, 0xc8, 0xdd, 0xe9, 0x02, 0x3f, 0xc0, 0x2b };
return &guid;
}
ggml_backend_t ggml_backend_vk_init(size_t dev_num) {
VK_LOG_DEBUG("ggml_backend_vk_init(" << dev_num << ")");
ggml_backend_vk_context * ctx = new ggml_backend_vk_context;
ggml_vk_init(ctx, dev_num);
ggml_backend_t vk_backend = new ggml_backend {
/* .guid = */ ggml_backend_vk_guid(),
/* .interface = */ ggml_backend_vk_interface,
/* .device = */ ggml_backend_reg_dev_get(ggml_backend_vk_reg(), dev_num),
/* .context = */ ctx,
};
return vk_backend;
}
bool ggml_backend_is_vk(ggml_backend_t backend) {
return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_vk_guid());
}
int ggml_backend_vk_get_device_count() {
return ggml_vk_get_device_count();
}
void ggml_backend_vk_get_device_description(int device, char * description, size_t description_size) {
GGML_ASSERT(device < (int) vk_instance.device_indices.size());
int dev_idx = vk_instance.device_indices[device];
ggml_vk_get_device_description(dev_idx, description, description_size);
}
void ggml_backend_vk_get_device_memory(int device, size_t * free, size_t * total) {
GGML_ASSERT(device < (int) vk_instance.device_indices.size());
vk::PhysicalDevice vkdev = vk_instance.instance.enumeratePhysicalDevices()[vk_instance.device_indices[device]];
vk::PhysicalDeviceMemoryProperties memprops = vkdev.getMemoryProperties();
for (const vk::MemoryHeap& heap : memprops.memoryHeaps) {
if (heap.flags & vk::MemoryHeapFlagBits::eDeviceLocal) {
*total = heap.size;
*free = heap.size;
break;
}
}
}
//////////////////////////
struct ggml_backend_vk_device_context {
int device;
std::string name;
std::string description;
};
static const char * ggml_backend_vk_device_get_name(ggml_backend_dev_t dev) {
ggml_backend_vk_device_context * ctx = (ggml_backend_vk_device_context *)dev->context;
return ctx->name.c_str();
}
static const char * ggml_backend_vk_device_get_description(ggml_backend_dev_t dev) {
ggml_backend_vk_device_context * ctx = (ggml_backend_vk_device_context *)dev->context;
return ctx->description.c_str();
}
static void ggml_backend_vk_device_get_memory(ggml_backend_dev_t device, size_t * free, size_t * total) {
ggml_backend_vk_device_context * ctx = (ggml_backend_vk_device_context *)device->context;
ggml_backend_vk_get_device_memory(ctx->device, free, total);
}
static ggml_backend_buffer_type_t ggml_backend_vk_device_get_buffer_type(ggml_backend_dev_t dev) {
ggml_backend_vk_device_context * ctx = (ggml_backend_vk_device_context *)dev->context;
return ggml_backend_vk_buffer_type(ctx->device);
}
static ggml_backend_buffer_type_t ggml_backend_vk_device_get_host_buffer_type(ggml_backend_dev_t dev) {
UNUSED(dev);
return ggml_backend_vk_host_buffer_type();
}
static enum ggml_backend_dev_type ggml_backend_vk_device_get_type(ggml_backend_dev_t dev) {
UNUSED(dev);
return GGML_BACKEND_DEVICE_TYPE_GPU_FULL;
}
static void ggml_backend_vk_device_get_props(ggml_backend_dev_t dev, struct ggml_backend_dev_props * props) {
props->name = ggml_backend_vk_device_get_name(dev);
props->description = ggml_backend_vk_device_get_description(dev);
props->type = ggml_backend_vk_device_get_type(dev);
ggml_backend_vk_device_get_memory(dev, &props->memory_free, &props->memory_total);
props->caps = {
/* async */ false,
/* host_buffer */ true,
/* events */ false,
};
}
static ggml_backend_t ggml_backend_vk_device_init(ggml_backend_dev_t dev, const char * params) {
UNUSED(params);
ggml_backend_vk_device_context * ctx = (ggml_backend_vk_device_context *)dev->context;
return ggml_backend_vk_init(ctx->device);
}
static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggml_tensor * op) {
switch (op->op) {
case GGML_OP_UNARY:
switch (ggml_get_unary_op(op)) {
@ -6701,97 +6827,101 @@ static bool ggml_backend_vk_supports_op(ggml_backend_t backend, const ggml_tenso
return false;
}
UNUSED(backend);
UNUSED(dev);
}
static bool ggml_backend_vk_offload_op(ggml_backend_t backend, const ggml_tensor * op) {
static bool ggml_backend_vk_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
if (buft->iface.get_name != ggml_backend_vk_buffer_type_name) {
return false;
}
ggml_backend_vk_device_context * ctx = (ggml_backend_vk_device_context *)dev->context;
ggml_backend_vk_buffer_type_context * buft_ctx = (ggml_backend_vk_buffer_type_context *)buft->context;
return buft_ctx->device->idx == ctx->device;
}
static bool ggml_backend_vk_device_offload_op(ggml_backend_dev_t dev, const ggml_tensor * op) {
const int min_batch_size = 32;
return (op->ne[1] >= min_batch_size && op->op != GGML_OP_GET_ROWS) ||
(op->ne[2] >= min_batch_size && op->op == GGML_OP_MUL_MAT_ID);
UNUSED(backend);
UNUSED(dev);
}
static bool ggml_backend_vk_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft) {
if (buft->iface.get_name != ggml_backend_vk_buffer_type_name) {
return false;
}
ggml_backend_vk_buffer_type_context * buft_ctx = (ggml_backend_vk_buffer_type_context *)buft->context;
ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context;
return buft_ctx->device == ctx->device;
}
// TODO: enable async and synchronize
static ggml_backend_i ggml_backend_vk_interface = {
/* .get_name = */ ggml_backend_vk_name,
/* .free = */ ggml_backend_vk_free,
/* .get_default_buffer_type = */ ggml_backend_vk_get_default_buffer_type,
/* .set_tensor_async = */ NULL, // ggml_backend_vk_set_tensor_async,
/* .get_tensor_async = */ NULL, // ggml_backend_vk_get_tensor_async,
/* .cpy_tensor_async = */ NULL, // ggml_backend_vk_cpy_tensor_async,
/* .synchronize = */ NULL, // ggml_backend_vk_synchronize,
/* .graph_plan_create = */ NULL,
/* .graph_plan_free = */ NULL,
/* .graph_plan_update = */ NULL,
/* .graph_plan_compute = */ NULL,
/* .graph_compute = */ ggml_backend_vk_graph_compute,
/* .supports_op = */ ggml_backend_vk_supports_op,
/* .supports_buft = */ ggml_backend_vk_supports_buft,
/* .offload_op = */ ggml_backend_vk_offload_op,
/* .event_record = */ NULL,
/* .event_wait = */ NULL,
static const struct ggml_backend_device_i ggml_backend_vk_device_i = {
/* .get_name = */ ggml_backend_vk_device_get_name,
/* .get_description = */ ggml_backend_vk_device_get_description,
/* .get_memory = */ ggml_backend_vk_device_get_memory,
/* .get_type = */ ggml_backend_vk_device_get_type,
/* .get_props = */ ggml_backend_vk_device_get_props,
/* .init_backend = */ ggml_backend_vk_device_init,
/* .get_buffer_type = */ ggml_backend_vk_device_get_buffer_type,
/* .get_host_buffer_type = */ ggml_backend_vk_device_get_host_buffer_type,
/* .buffer_from_host_ptr = */ NULL,
/* .supports_op = */ ggml_backend_vk_device_supports_op,
/* .supports_buft = */ ggml_backend_vk_device_supports_buft,
/* .offload_op = */ ggml_backend_vk_device_offload_op,
/* .event_new = */ NULL,
/* .event_free = */ NULL,
/* .event_synchronize = */ NULL,
};
static ggml_guid_t ggml_backend_vk_guid() {
static ggml_guid guid = { 0xb8, 0xf7, 0x4f, 0x86, 0x40, 0x3c, 0xe1, 0x02, 0x91, 0xc8, 0xdd, 0xe9, 0x02, 0x3f, 0xc0, 0x2b };
return &guid;
static const char * ggml_backend_vk_reg_get_name(ggml_backend_reg_t reg) {
UNUSED(reg);
return GGML_VK_NAME;
}
ggml_backend_t ggml_backend_vk_init(size_t dev_num) {
VK_LOG_DEBUG("ggml_backend_vk_init(" << dev_num << ")");
ggml_backend_vk_context * ctx = new ggml_backend_vk_context;
ggml_vk_init(ctx, dev_num);
ggml_backend_t vk_backend = new ggml_backend {
/* .guid = */ ggml_backend_vk_guid(),
/* .interface = */ ggml_backend_vk_interface,
/* .device = */ nullptr,
/* .context = */ ctx,
};
return vk_backend;
static size_t ggml_backend_vk_reg_get_device_count(ggml_backend_reg_t reg) {
UNUSED(reg);
return ggml_backend_vk_get_device_count();
}
bool ggml_backend_is_vk(ggml_backend_t backend) {
return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_vk_guid());
}
static ggml_backend_dev_t ggml_backend_vk_reg_get_device(ggml_backend_reg_t reg, size_t device) {
static std::vector<ggml_backend_dev_t> devices;
int ggml_backend_vk_get_device_count() {
return ggml_vk_get_device_count();
}
static bool initialized = false;
void ggml_backend_vk_get_device_description(int device, char * description, size_t description_size) {
ggml_vk_get_device_description(device, description, description_size);
}
void ggml_backend_vk_get_device_memory(int device, size_t * free, size_t * total) {
GGML_ASSERT(device < (int) vk_instance.device_indices.size());
vk::PhysicalDevice vkdev = vk_instance.instance.enumeratePhysicalDevices()[vk_instance.device_indices[device]];
vk::PhysicalDeviceMemoryProperties memprops = vkdev.getMemoryProperties();
for (const vk::MemoryHeap& heap : memprops.memoryHeaps) {
if (heap.flags & vk::MemoryHeapFlagBits::eDeviceLocal) {
*total = heap.size;
*free = heap.size;
break;
{
static std::mutex mutex;
std::lock_guard<std::mutex> lock(mutex);
if (!initialized) {
for (size_t i = 0; i < ggml_backend_vk_get_device_count(); i++) {
ggml_backend_vk_device_context * ctx = new ggml_backend_vk_device_context;
char desc[256];
ggml_backend_vk_get_device_description(i, desc, sizeof(desc));
ctx->device = i;
ctx->name = GGML_VK_NAME + std::to_string(i);
ctx->description = desc;
devices.push_back(new ggml_backend_device {
/* .iface = */ ggml_backend_vk_device_i,
/* .reg = */ reg,
/* .context = */ ctx,
});
}
initialized = true;
}
}
GGML_ASSERT(device < devices.size());
return devices[device];
}
static const struct ggml_backend_reg_i ggml_backend_vk_reg_i = {
/* .get_name = */ ggml_backend_vk_reg_get_name,
/* .get_device_count = */ ggml_backend_vk_reg_get_device_count,
/* .get_device = */ ggml_backend_vk_reg_get_device,
/* .get_proc_address = */ NULL,
};
ggml_backend_reg_t ggml_backend_vk_reg() {
static ggml_backend_reg reg = {
/* .iface = */ ggml_backend_vk_reg_i,
/* .context = */ nullptr,
};
return &reg;
}
// Extension availability

94
ggml/src/ggml.c
View File

@ -35,10 +35,6 @@
#include <omp.h>
#endif
#ifdef GGML_USE_METAL
#include <unistd.h>
#endif
#if defined(__ARM_FEATURE_SVE) || defined(__ARM_FEATURE_MATMUL_INT8)
#undef GGML_USE_LLAMAFILE
#endif
@ -189,6 +185,8 @@ typedef pthread_t ggml_thread_t;
#endif
#if defined(__APPLE__)
#include <unistd.h>
#include <mach/mach.h>
#include <TargetConditionals.h>
#endif
@ -326,8 +324,9 @@ struct ggml_logger_state {
static struct ggml_logger_state g_logger_state = {ggml_log_callback_default, NULL};
static void ggml_log_internal_v(enum ggml_log_level level, const char * format, va_list args) {
if (format == NULL)
if (format == NULL) {
return;
}
va_list args_copy;
va_copy(args_copy, args);
char buffer[128];
@ -386,22 +385,40 @@ void ggml_log_callback_default(enum ggml_log_level level, const char * text, voi
//#define GGML_SOFT_MAX_ACCELERATE
#endif
void * ggml_aligned_malloc(size_t size) {
#if defined(_MSC_VER) || defined(__MINGW32__)
#define GGML_ALIGNED_MALLOC(size) _aligned_malloc(size, GGML_MEM_ALIGN)
#define GGML_ALIGNED_FREE(ptr) _aligned_free(ptr)
return _aligned_malloc(size, TENSOR_ALIGNMENT);
#else
inline static void * ggml_aligned_malloc(size_t size) {
if (size == 0) {
GGML_LOG_WARN("Behavior may be unexpected when allocating 0 bytes for ggml_aligned_malloc!\n");
return NULL;
}
void * aligned_memory = NULL;
#ifdef GGML_USE_CPU_HBM
int result = hbw_posix_memalign(&aligned_memory, 16, size);
int result = hbw_posix_memalign(&aligned_memory, TENSOR_ALIGNMENT, size);
#elif TARGET_OS_OSX
kern_return_t alloc_status = vm_allocate((vm_map_t) mach_task_self(), (vm_address_t *) &aligned_memory, size, VM_FLAGS_ANYWHERE);
int result = EFAULT;
switch (alloc_status) {
case KERN_SUCCESS:
result = 0;
break;
case KERN_INVALID_ADDRESS:
result = EINVAL;
break;
case KERN_NO_SPACE:
result = ENOMEM;
break;
default:
result = EFAULT;
break;
}
#elif GGML_USE_METAL
int result = posix_memalign(&aligned_memory, sysconf(_SC_PAGESIZE), size);
const long page_size = sysconf(_SC_PAGESIZE);
int result = posix_memalign(&aligned_memory, MAX(TENSOR_ALIGNMENT, page_size), size);
#else
int result = posix_memalign(&aligned_memory, GGML_MEM_ALIGN, size);
int result = posix_memalign(&aligned_memory, TENSOR_ALIGNMENT, size);
#endif
if (result != 0) {
// Handle allocation failure
@ -419,14 +436,26 @@ inline static void * ggml_aligned_malloc(size_t size) {
return NULL;
}
return aligned_memory;
#endif
}
#define GGML_ALIGNED_MALLOC(size) ggml_aligned_malloc(size)
#ifdef GGML_USE_CPU_HBM
#define GGML_ALIGNED_FREE(ptr) if(NULL != ptr) hbw_free(ptr)
void ggml_aligned_free(void * ptr, size_t size) {
GGML_UNUSED(size);
#if defined(_MSC_VER) || defined(__MINGW32__)
_aligned_free(ptr);
#elif GGML_USE_CPU_HBM
if (ptr != NULL) {
hbw_free(ptr);
}
#elif TARGET_OS_OSX
if (ptr != NULL) {
vm_deallocate((vm_map_t)mach_task_self(), (vm_address_t)ptr, size);
}
#else
#define GGML_ALIGNED_FREE(ptr) free(ptr)
#endif
free(ptr);
#endif
}
inline static void * ggml_malloc(size_t size) {
if (size == 0) {
@ -3435,7 +3464,7 @@ int64_t ggml_nrows(const struct ggml_tensor * tensor) {
size_t ggml_nbytes(const struct ggml_tensor * tensor) {
size_t nbytes;
size_t blck_size = ggml_blck_size(tensor->type);
const size_t blck_size = ggml_blck_size(tensor->type);
if (blck_size == 1) {
nbytes = ggml_type_size(tensor->type);
for (int i = 0; i < GGML_MAX_DIMS; ++i) {
@ -3823,10 +3852,6 @@ struct ggml_context * ggml_init(struct ggml_init_params params) {
},
};
for (int i = 0; i < GGML_MAX_CONTEXTS; ++i) {
g_state.contexts[i].used = false;
}
const uint64_t t_end = ggml_time_us(); UNUSED(t_end);
GGML_PRINT_DEBUG("%s: g_state initialized in %f ms\n", __func__, (t_end - t_start)/1000.0f);
@ -3869,7 +3894,7 @@ struct ggml_context * ggml_init(struct ggml_init_params params) {
*ctx = (struct ggml_context) {
/*.mem_size =*/ mem_size,
/*.mem_buffer =*/ params.mem_buffer ? params.mem_buffer : GGML_ALIGNED_MALLOC(mem_size),
/*.mem_buffer =*/ params.mem_buffer ? params.mem_buffer : ggml_aligned_malloc(mem_size),
/*.mem_buffer_owned =*/ params.mem_buffer ? false : true,
/*.no_alloc =*/ params.no_alloc,
/*.no_alloc_save =*/ params.no_alloc,
@ -3909,7 +3934,7 @@ void ggml_free(struct ggml_context * ctx) {
__func__, i, ggml_used_mem(ctx));
if (ctx->mem_buffer_owned) {
GGML_ALIGNED_FREE(ctx->mem_buffer);
ggml_aligned_free(ctx->mem_buffer, ctx->mem_size);
}
found = true;
@ -15695,6 +15720,9 @@ static void ggml_compute_forward_flash_attn_ext_f16(
ggml_vec_dot_t const kq_vec_dot = type_traits[k->type].vec_dot;
ggml_to_float_t const v_to_float = type_traits[v->type].to_float;
GGML_ASSERT(q_to_vec_dot && "fattn: unsupported K-type");
GGML_ASSERT(v_to_float && "fattn: unsupported V-type");
// loop over n_batch and n_head
for (int ir = ir0; ir < ir1; ++ir) {
// q indices
@ -19608,9 +19636,10 @@ static void ggml_thread_cpumask_next(const bool * global_mask, bool * local_mask
void ggml_threadpool_free(struct ggml_threadpool* threadpool) {
if (!threadpool) return;
const int n_threads = threadpool->n_threads_max;
#ifndef GGML_USE_OPENMP
struct ggml_compute_state* workers = threadpool->workers;
const int n_threads = threadpool->n_threads_max;
ggml_mutex_lock(&threadpool->mutex);
@ -19630,8 +19659,9 @@ void ggml_threadpool_free(struct ggml_threadpool* threadpool) {
ggml_cond_destroy(&threadpool->cond);
#endif // GGML_USE_OPENMP
GGML_ALIGNED_FREE(threadpool->workers);
GGML_ALIGNED_FREE(threadpool);
const size_t workers_size = sizeof(struct ggml_compute_state) * n_threads;
ggml_aligned_free(threadpool->workers, workers_size);
ggml_aligned_free(threadpool, sizeof(struct ggml_threadpool));
}
#ifndef GGML_USE_OPENMP
@ -20063,7 +20093,7 @@ static struct ggml_threadpool * ggml_threadpool_new_impl(
struct ggml_cplan * cplan) {
struct ggml_threadpool * threadpool =
GGML_ALIGNED_MALLOC(sizeof(struct ggml_threadpool));
ggml_aligned_malloc(sizeof(struct ggml_threadpool));
{
threadpool->cgraph = cgraph;
threadpool->cplan = cplan;
@ -20084,7 +20114,7 @@ static struct ggml_threadpool * ggml_threadpool_new_impl(
// Allocate and init workers state
const size_t workers_size = sizeof(struct ggml_compute_state) * tpp->n_threads;
struct ggml_compute_state * workers = GGML_ALIGNED_MALLOC(workers_size);
struct ggml_compute_state * workers = ggml_aligned_malloc(workers_size);
memset(workers, 0, workers_size);
for (int j = 0; j < tpp->n_threads; j++) {
@ -23222,6 +23252,14 @@ int ggml_cpu_has_avx512_bf16(void) {
#endif
}
int ggml_cpu_has_amx_int8(void) {
#if defined(__AMX_INT8__)
return 1;
#else
return 0;
#endif
}
int ggml_cpu_has_fma(void) {
#if defined(__FMA__)
return 1;

57
ggml/src/llamafile/sgemm.cpp
View File

@ -942,6 +942,36 @@ class tinyBLAS_Q0_AVX {
return _mm_sub_epi8(_mm_and_si128(_mm_set1_epi8(15), _mm_srli_epi16(x, 4)), _mm_set1_epi8(8));
}
inline __m256i load(const block_q5_0 *b) {
return _mm256_or_si256(denibble(b->qs), bittobyte(b->qh));
}
inline __m128i load0(const block_q5_0* b) {
const __m128i x = _mm_loadu_si128((const __m128i *)(b->qs));
uint32_t x32;
memcpy(&x32, b->qh, sizeof(uint32_t));
__m128i qxl = _mm_and_si128(_mm_set1_epi8(15), x);
__m128i bytesl = _mm_cmpeq_epi8(_mm_set1_epi64x(-1),
_mm_or_si128(_mm_set1_epi64x(0x7fbfdfeff7fbfdfe),
_mm_shuffle_epi8(_mm_set1_epi32(x32),
_mm_set_epi64x(0x0101010101010101, 0x0000000000000000))));
bytesl = _mm_andnot_si128(bytesl, _mm_set1_epi8((char)0xF0));
return _mm_or_si128(qxl, bytesl);
}
inline __m128i load1(const block_q5_0* b) {
const __m128i x = _mm_loadu_si128((const __m128i *)(b->qs));
uint32_t x32;
memcpy(&x32, b->qh, sizeof(uint32_t));
__m128i qxh = _mm_and_si128(_mm_set1_epi8(15), _mm_srli_epi16(x, 4));
__m128i bytesh = _mm_cmpeq_epi8(_mm_set1_epi64x(-1),
_mm_or_si128(_mm_set1_epi64x(0x7fbfdfeff7fbfdfe),
_mm_shuffle_epi8(_mm_set1_epi32(x32),
_mm_set_epi64x(0x0303030303030303, 0x0202020202020202))));
bytesh = _mm_andnot_si128(bytesh, _mm_set1_epi8((char)0xF0));
return _mm_or_si128(qxh, bytesh);
}
inline __m256i load(const block_iq4_nl *b) {
return MM256_SET_M128I(load1(b), load0(b));
}
@ -973,6 +1003,17 @@ class tinyBLAS_Q0_AVX {
_mm_srli_epi16(x, 4), 1));
}
static inline __m256i bittobyte(const uint8_t *p) {
uint32_t x32;
memcpy(&x32, p, sizeof(uint32_t));
__m256i bytes = _mm256_cmpeq_epi8(_mm256_set1_epi64x(-1),
_mm256_or_si256(_mm256_set1_epi64x(0x7fbfdfeff7fbfdfe),
_mm256_shuffle_epi8(_mm256_set1_epi32(x32),
_mm256_set_epi64x(0x0303030303030303, 0x0202020202020202,
0x0101010101010101, 0x0000000000000000))));
return _mm256_andnot_si256(bytes, _mm256_set1_epi8((char)0xF0));
}
const TA *const A;
const TB *const B;
TC *const C;
@ -1182,6 +1223,22 @@ bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda
#endif
}
case GGML_TYPE_Q5_0: {
if (Btype != GGML_TYPE_Q8_0)
return false;
#if defined(__AVX2__) || defined(__AVX512F__) || defined(__AVX__)
tinyBLAS_Q0_AVX<block_q5_0, block_q8_0, float> tb{
k, (const block_q5_0 *)A, lda,
(const block_q8_0 *)B, ldb,
(float *)C, ldc,
ith, nth};
tb.matmul(m, n);
return true;
#else
return false;
#endif
}
case GGML_TYPE_IQ4_NL: {
if (Btype != GGML_TYPE_Q8_0)
return false;

26
gguf-py/gguf/constants.py
View File

@ -152,6 +152,8 @@ class Keys:
MERGES = "tokenizer.ggml.merges"
BOS_ID = "tokenizer.ggml.bos_token_id"
EOS_ID = "tokenizer.ggml.eos_token_id"
EOT_ID = "tokenizer.ggml.eot_token_id"
EOM_ID = "tokenizer.ggml.eom_token_id"
UNK_ID = "tokenizer.ggml.unknown_token_id"
SEP_ID = "tokenizer.ggml.seperator_token_id"
PAD_ID = "tokenizer.ggml.padding_token_id"
@ -168,11 +170,16 @@ class Keys:
CHAT_TEMPLATE_N = "tokenizer.chat_template.{name}"
CHAT_TEMPLATES = "tokenizer.chat_templates"
# FIM/Infill special tokens constants
FIM_PRE_ID = "tokenizer.ggml.fim_pre_token_id"
FIM_SUF_ID = "tokenizer.ggml.fim_suf_token_id"
FIM_MID_ID = "tokenizer.ggml.fim_mid_token_id"
FIM_PAD_ID = "tokenizer.ggml.fim_pad_token_id"
FIM_REP_ID = "tokenizer.ggml.fim_rep_token_id"
FIM_SEP_ID = "tokenizer.ggml.fim_sep_token_id"
# deprecated:
PREFIX_ID = "tokenizer.ggml.prefix_token_id"
SUFFIX_ID = "tokenizer.ggml.suffix_token_id"
MIDDLE_ID = "tokenizer.ggml.middle_token_id"
EOT_ID = "tokenizer.ggml.eot_token_id"
EOM_ID = "tokenizer.ggml.eom_token_id"
class Adapter:
TYPE = "adapter.type"
@ -1579,6 +1586,8 @@ KEY_TOKENIZER_SCORES = Keys.Tokenizer.SCORES
KEY_TOKENIZER_MERGES = Keys.Tokenizer.MERGES
KEY_TOKENIZER_BOS_ID = Keys.Tokenizer.BOS_ID
KEY_TOKENIZER_EOS_ID = Keys.Tokenizer.EOS_ID
KEY_TOKENIZER_EOT_ID = Keys.Tokenizer.EOT_ID
KEY_TOKENIZER_EOM_ID = Keys.Tokenizer.EOM_ID
KEY_TOKENIZER_UNK_ID = Keys.Tokenizer.UNK_ID
KEY_TOKENIZER_SEP_ID = Keys.Tokenizer.SEP_ID
KEY_TOKENIZER_PAD_ID = Keys.Tokenizer.PAD_ID
@ -1586,8 +1595,15 @@ KEY_TOKENIZER_CLS_ID = Keys.Tokenizer.CLS_ID
KEY_TOKENIZER_MASK_ID = Keys.Tokenizer.MASK_ID
KEY_TOKENIZER_HF_JSON = Keys.Tokenizer.HF_JSON
KEY_TOKENIZER_RWKV = Keys.Tokenizer.RWKV
KEY_TOKENIZER_PRIFIX_ID = Keys.Tokenizer.PREFIX_ID
KEY_TOKENIZER_FIM_PRE_ID = Keys.Tokenizer.FIM_PRE_ID
KEY_TOKENIZER_FIM_SUF_ID = Keys.Tokenizer.FIM_SUF_ID
KEY_TOKENIZER_FIM_MID_ID = Keys.Tokenizer.FIM_MID_ID
KEY_TOKENIZER_FIM_PAD_ID = Keys.Tokenizer.FIM_PAD_ID
KEY_TOKENIZER_FIM_REP_ID = Keys.Tokenizer.FIM_REP_ID
KEY_TOKENIZER_FIM_SEP_ID = Keys.Tokenizer.FIM_SEP_ID
# deprecated
KEY_TOKENIZER_PREFIX_ID = Keys.Tokenizer.PREFIX_ID
KEY_TOKENIZER_SUFFIX_ID = Keys.Tokenizer.SUFFIX_ID
KEY_TOKENIZER_MIDDLE_ID = Keys.Tokenizer.MIDDLE_ID
KEY_TOKENIZER_EOT_ID = Keys.Tokenizer.EOT_ID
KEY_TOKENIZER_EOM_ID = Keys.Tokenizer.EOM_ID

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

@ -843,15 +843,6 @@ class GGUFWriter:
self.add_string(Keys.Tokenizer.CHAT_TEMPLATE, value)
def add_prefix_token_id(self, id: int) -> None:
self.add_uint32(Keys.Tokenizer.PREFIX_ID, id)
def add_suffix_token_id(self, id: int) -> None:
self.add_uint32(Keys.Tokenizer.SUFFIX_ID, id)
def add_middle_token_id(self, id: int) -> None:
self.add_uint32(Keys.Tokenizer.MIDDLE_ID, id)
def add_eot_token_id(self, id: int) -> None:
self.add_uint32(Keys.Tokenizer.EOT_ID, id)

82
include/llama.h
View File

@ -217,6 +217,7 @@ extern "C" {
typedef struct llama_token_data_array {
// TODO: consider SoA
// NOTE: this pointer can be modified by the samplers
llama_token_data * data;
size_t size;
int64_t selected; // this is the index in the data array (i.e. not the token id)
@ -232,8 +233,11 @@ extern "C" {
// - token : the token ids of the input (used when embd is NULL)
// - embd : token embeddings (i.e. float vector of size n_embd) (used when token is NULL)
// - pos : the positions of the respective token in the sequence
// (if set to NULL, the token position will be tracked automatically by llama_decode)
// - seq_id : the sequence to which the respective token belongs
// (if set to NULL, the sequence ID will be assumed to be 0)
// - logits : if zero, the logits (and/or the embeddings) for the respective token will not be output
// (if set to NULL, only the logits for last token will be returned)
//
typedef struct llama_batch {
int32_t n_tokens;
@ -244,15 +248,6 @@ extern "C" {
int32_t * n_seq_id;
llama_seq_id ** seq_id;
int8_t * logits; // TODO: rename this to "output"
// NOTE: helpers for smooth API transition - can be deprecated in the future
// for future-proof code, use the above fields instead and ignore everything below
//
// pos[i] = all_pos_0 + i*all_pos_1
//
llama_pos all_pos_0; // used if pos == NULL
llama_pos all_pos_1; // used if pos == NULL
llama_seq_id all_seq_id; // used if seq_id == NULL
} llama_batch;
enum llama_model_kv_override_type {
@ -776,15 +771,15 @@ extern "C" {
// Decoding
//
// Return batch for single sequence of tokens starting at pos_0
// Return batch for single sequence of tokens
// The sequence ID will be fixed to 0
// The position of the tokens will be tracked automatically by llama_decode
//
// NOTE: this is a helper function to facilitate transition to the new batch API - avoid using it
//
LLAMA_API struct llama_batch llama_batch_get_one(
llama_token * tokens,
int32_t n_tokens,
llama_pos pos_0,
llama_seq_id seq_id);
int32_t n_tokens);
// Allocates a batch of tokens on the heap that can hold a maximum of n_tokens
// Each token can be assigned up to n_seq_max sequence ids
@ -897,6 +892,7 @@ extern "C" {
// Special tokens
LLAMA_API llama_token llama_token_bos(const struct llama_model * model); // beginning-of-sentence
LLAMA_API llama_token llama_token_eos(const struct llama_model * model); // end-of-sentence
LLAMA_API llama_token llama_token_eot(const struct llama_model * model); // end-of-turn
LLAMA_API llama_token llama_token_cls(const struct llama_model * model); // classification
LLAMA_API llama_token llama_token_sep(const struct llama_model * model); // sentence separator
LLAMA_API llama_token llama_token_nl (const struct llama_model * model); // next-line
@ -905,11 +901,17 @@ extern "C" {
LLAMA_API bool llama_add_bos_token(const struct llama_model * model);
LLAMA_API bool llama_add_eos_token(const struct llama_model * model);
// Codellama infill tokens
LLAMA_API llama_token llama_token_prefix(const struct llama_model * model); // Beginning of infill prefix
LLAMA_API llama_token llama_token_middle(const struct llama_model * model); // Beginning of infill middle
LLAMA_API llama_token llama_token_suffix(const struct llama_model * model); // Beginning of infill suffix
LLAMA_API llama_token llama_token_eot (const struct llama_model * model); // End of infill middle
// infill tokens
DEPRECATED(LLAMA_API llama_token llama_token_prefix(const struct llama_model * model), "use llama_token_fim_pre instead");
DEPRECATED(LLAMA_API llama_token llama_token_middle(const struct llama_model * model), "use llama_token_fim_mid instead");
DEPRECATED(LLAMA_API llama_token llama_token_suffix(const struct llama_model * model), "use llama_token_fim_suf instead");
LLAMA_API llama_token llama_token_fim_pre(const struct llama_model * model);
LLAMA_API llama_token llama_token_fim_suf(const struct llama_model * model);
LLAMA_API llama_token llama_token_fim_mid(const struct llama_model * model);
LLAMA_API llama_token llama_token_fim_pad(const struct llama_model * model);
LLAMA_API llama_token llama_token_fim_rep(const struct llama_model * model);
LLAMA_API llama_token llama_token_fim_sep(const struct llama_model * model);
//
// Tokenization
@ -1068,12 +1070,13 @@ extern "C" {
// available samplers:
LLAMA_API struct llama_sampler * llama_sampler_init_greedy (void);
LLAMA_API struct llama_sampler * llama_sampler_init_dist (uint32_t seed);
LLAMA_API struct llama_sampler * llama_sampler_init_greedy(void);
LLAMA_API struct llama_sampler * llama_sampler_init_dist (uint32_t seed);
/// @details Sorts candidate tokens by their logits in descending order and calculate probabilities based on logits.
/// NOTE: Avoid using on the full vocabulary as the sorting can become slow. For example, apply top-k or top-p sampling first.
LLAMA_API struct llama_sampler * llama_sampler_init_softmax (void);
DEPRECATED(LLAMA_API struct llama_sampler * llama_sampler_init_softmax (void),
"will be removed in the future (see https://github.com/ggerganov/llama.cpp/pull/9896#discussion_r1800920915)");
/// @details Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
LLAMA_API struct llama_sampler * llama_sampler_init_top_k (int32_t k);
@ -1089,11 +1092,16 @@ extern "C" {
/// @details Locally Typical Sampling implementation described in the paper https://arxiv.org/abs/2202.00666.
LLAMA_API struct llama_sampler * llama_sampler_init_typical (float p, size_t min_keep);
/// #details Updates the logits l_i` = l_i/t. When t <= 0.0f, the maximum logit is kept at it's original value, the rest are set to -inf
LLAMA_API struct llama_sampler * llama_sampler_init_temp (float t);
/// @details Dynamic temperature implementation (a.k.a. entropy) described in the paper https://arxiv.org/abs/2309.02772.
LLAMA_API struct llama_sampler * llama_sampler_init_temp_ext (float t, float delta, float exponent);
/// @details XTC sampler as described in https://github.com/oobabooga/text-generation-webui/pull/6335
LLAMA_API struct llama_sampler * llama_sampler_init_xtc (float p, float t, size_t min_keep, uint32_t seed);
/// @details Mirostat 1.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words.
/// @param candidates A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text.
/// @param tau The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text.
@ -1133,11 +1141,43 @@ extern "C" {
bool penalize_nl, // consider newlines as a repeatable token
bool ignore_eos); // ignore the end-of-sequence token
/// @details DRY sampler, designed by p-e-w, as described in: https://github.com/oobabooga/text-generation-webui/pull/5677, porting Koboldcpp implementation authored by pi6am: https://github.com/LostRuins/koboldcpp/pull/982
LLAMA_API struct llama_sampler * llama_sampler_init_dry(
const struct llama_model * model,
float dry_multiplier,
float dry_base,
int32_t dry_allowed_length,
int32_t dry_penalty_last_n,
const char ** seq_breakers,
size_t num_breakers);
LLAMA_API struct llama_sampler * llama_sampler_init_logit_bias(
int32_t n_vocab,
int32_t n_logit_bias,
const llama_logit_bias * logit_bias);
// this sampler is meant to be used for fill-in-the-middle infilling
// it's supposed to be used after top_k + top_p sampling
//
// 1. if the sum of the EOG probs times the number of candidates is higher than the sum of the other probs -> pick EOG
// 2. combine probs of tokens that have the same prefix
//
// example:
//
// - before:
// "hel": 0.5
// "hell": 0.2
// "hello": 0.1
// "dummy": 0.1
//
// - after:
// "hel": 0.8
// "dummy": 0.1
//
// 3. discard non-EOG tokens with low prob
// 4. if no tokens are left -> pick EOT
//
LLAMA_API struct llama_sampler * llama_sampler_init_infill(const struct llama_model * model);
// Returns the seed used by the sampler if applicable, LLAMA_DEFAULT_SEED otherwise
LLAMA_API uint32_t llama_sampler_get_seed(const struct llama_sampler * smpl);

6
scripts/run-with-preset.py
View File

@ -15,7 +15,7 @@ CLI_ARGS_LLAMA_CLI_PERPLEXITY = [
"export", "file", "frequency-penalty", "grammar", "grammar-file", "hellaswag",
"hellaswag-tasks", "ignore-eos", "in-prefix", "in-prefix-bos", "in-suffix",
"interactive", "interactive-first", "keep", "logdir", "logit-bias", "lora", "lora-base",
"low-vram", "main-gpu", "memory-f32", "mirostat", "mirostat-ent", "mirostat-lr", "mlock",
"low-vram", "main-gpu", "mirostat", "mirostat-ent", "mirostat-lr", "mlock",
"model", "multiline-input", "n-gpu-layers", "n-predict", "no-mmap", "no-mul-mat-q",
"np-penalize-nl", "numa", "ppl-output-type", "ppl-stride", "presence-penalty", "prompt",
"prompt-cache", "prompt-cache-all", "prompt-cache-ro", "repeat-last-n",
@ -25,12 +25,12 @@ CLI_ARGS_LLAMA_CLI_PERPLEXITY = [
]
CLI_ARGS_LLAMA_BENCH = [
"batch-size", "memory-f32", "low-vram", "model", "mul-mat-q", "n-gen", "n-gpu-layers",
"batch-size", "low-vram", "model", "mul-mat-q", "n-gen", "n-gpu-layers",
"n-prompt", "output", "repetitions", "tensor-split", "threads", "verbose"
]
CLI_ARGS_LLAMA_SERVER = [
"alias", "batch-size", "ctx-size", "embedding", "host", "memory-f32", "lora", "lora-base",
"alias", "batch-size", "ctx-size", "embedding", "host", "lora", "lora-base",
"low-vram", "main-gpu", "mlock", "model", "n-gpu-layers", "n-probs", "no-mmap", "no-mul-mat-q",
"numa", "path", "port", "rope-freq-base", "timeout", "rope-freq-scale", "tensor-split",
"threads", "verbose"

7
scripts/sync-ggml-am.sh
View File

@ -76,6 +76,7 @@ while read c; do
src/ggml*.m \
src/ggml*.metal \
src/ggml*.cu \
src/ggml-amx/* \
src/ggml-cann/* \
src/ggml-cuda/* \
src/ggml-sycl/* \
@ -121,6 +122,8 @@ if [ -f $SRC_LLAMA/ggml-src.patch ]; then
# src/ggml-aarch64.c -> ggml/src/ggml-aarch64.c
# src/ggml-aarch64.h -> ggml/src/ggml-aarch64.h
# src/ggml-alloc.c -> ggml/src/ggml-alloc.c
# src/ggml-amx/* -> ggml/src/ggml-amx/
# src/ggml-amx.cpp -> ggml/src/ggml-amx.cpp
# src/ggml-backend-impl.h -> ggml/src/ggml-backend-impl.h
# src/ggml-backend.cpp -> ggml/src/ggml-backend.cpp
# src/ggml-cann/* -> ggml/src/ggml-cann/
@ -141,6 +144,7 @@ if [ -f $SRC_LLAMA/ggml-src.patch ]; then
#
# include/ggml.h -> ggml/include/ggml.h
# include/ggml-alloc.h -> ggml/include/ggml-alloc.h
# include/ggml-amx.h -> ggml/include/ggml-amx.h
# include/ggml-backend.h -> ggml/include/ggml-backend.h
# include/ggml-blas.h -> ggml/include/ggml-blas.h
# include/ggml-cann.h -> ggml/include/ggml-cann.h
@ -168,6 +172,8 @@ if [ -f $SRC_LLAMA/ggml-src.patch ]; then
-e 's/([[:space:]]|[ab]\/)src\/ggml-aarch64\.c/\1ggml\/src\/ggml-aarch64.c/g' \
-e 's/([[:space:]]|[ab]\/)src\/ggml-aarch64\.h/\1ggml\/src\/ggml-aarch64.h/g' \
-e 's/([[:space:]]|[ab]\/)src\/ggml-alloc\.c/\1ggml\/src\/ggml-alloc.c/g' \
-e 's/([[:space:]]|[ab]\/)src\/ggml-amx\//\1ggml\/src\/ggml-amx\//g' \
-e 's/([[:space:]]|[ab]\/)src\/ggml-amx\.cpp/\1ggml\/src\/ggml-amx.cpp/g' \
-e 's/([[:space:]]|[ab]\/)src\/ggml-backend-impl\.h/\1ggml\/src\/ggml-backend-impl.h/g' \
-e 's/([[:space:]]|[ab]\/)src\/ggml-backend\.cpp/\1ggml\/src\/ggml-backend.cpp/g' \
-e 's/([[:space:]]|[ab]\/)src\/ggml-cann\//\1ggml\/src\/ggml-cann\//g' \
@ -187,6 +193,7 @@ if [ -f $SRC_LLAMA/ggml-src.patch ]; then
-e 's/([[:space:]]|[ab]\/)src\/vulkan-shaders\//\1ggml\/src\/vulkan-shaders\//g' \
-e 's/([[:space:]]|[ab]\/)include\/ggml\.h/\1ggml\/include\/ggml.h/g' \
-e 's/([[:space:]]|[ab]\/)include\/ggml-alloc\.h/\1ggml\/include\/ggml-alloc.h/g' \
-e 's/([[:space:]]|[ab]\/)include\/ggml-amx\.h/\1ggml\/include\/ggml-amx.h/g' \
-e 's/([[:space:]]|[ab]\/)include\/ggml-backend\.h/\1ggml\/include\/ggml-backend.h/g' \
-e 's/([[:space:]]|[ab]\/)include\/ggml-blas\.h/\1ggml\/include\/ggml-blas.h/g' \
-e 's/([[:space:]]|[ab]\/)include\/ggml-cann\.h/\1ggml\/include\/ggml-cann.h/g' \

2
scripts/sync-ggml.last
View File

@ -1 +1 @@
564f42082f858f9674b2a2e06e9e779d9ed2c754
162e232411ee98ceb0cccfa84886118d917d2123

3
scripts/sync-ggml.sh
View File

@ -8,6 +8,8 @@ cp -rpv ../ggml/src/ggml.c ./ggml/src/ggml.c
cp -rpv ../ggml/src/ggml-aarch64.c ./ggml/src/ggml-aarch64.c
cp -rpv ../ggml/src/ggml-aarch64.h ./ggml/src/ggml-aarch64.h
cp -rpv ../ggml/src/ggml-alloc.c ./ggml/src/ggml-alloc.c
cp -rpv ../ggml/src/ggml-amx/* ./ggml/src/ggml-amx/
cp -rpv ../ggml/src/ggml-amx.cpp ./ggml/src/ggml-amx.cpp
cp -rpv ../ggml/src/ggml-backend-impl.h ./ggml/src/ggml-backend-impl.h
cp -rpv ../ggml/src/ggml-backend.cpp ./ggml/src/ggml-backend.cpp
cp -rpv ../ggml/src/ggml-cann/* ./ggml/src/ggml-cann/
@ -29,6 +31,7 @@ cp -rpv ../ggml/src/vulkan-shaders/* ./ggml/src/vulkan-shaders/
cp -rpv ../ggml/include/ggml.h ./ggml/include/ggml.h
cp -rpv ../ggml/include/ggml-alloc.h ./ggml/include/ggml-alloc.h
cp -rpv ../ggml/include/ggml-amx.h ./ggml/include/ggml-amx.h
cp -rpv ../ggml/include/ggml-backend.h ./ggml/include/ggml-backend.h
cp -rpv ../ggml/include/ggml-blas.h ./ggml/include/ggml-blas.h
cp -rpv ../ggml/include/ggml-cann.h ./ggml/include/ggml-cann.h

750
src/llama-sampling.cpp
View File

@ -63,6 +63,30 @@ static void llama_log_softmax(float * array, size_t size) {
}
*/
static void llama_sampler_temp_impl(llama_token_data_array * cur_p, float temp) {
if (temp <= 0.0f) {
// find the token with the highest logit and set the rest to -inf
size_t max_i = 0;
float max_l = cur_p->data[0].logit;
for (size_t i = 1; i < cur_p->size; ++i) {
if (cur_p->data[i ].logit > max_l) {
cur_p->data[max_i].logit = -INFINITY;
max_i = i;
max_l = cur_p->data[i].logit;
} else {
cur_p->data[i].logit = -INFINITY;
}
}
return;
}
for (size_t i = 0; i < cur_p->size; ++i) {
cur_p->data[i].logit /= temp;
}
}
static void llama_sampler_softmax_impl(llama_token_data_array * cur_p) {
GGML_ASSERT(cur_p->size > 0);
@ -427,6 +451,9 @@ static const char * llama_sampler_dist_name(const struct llama_sampler * /*smpl*
static void llama_sampler_dist_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
auto * ctx = (llama_sampler_dist *) smpl->ctx;
llama_sampler_softmax_impl(cur_p);
cur_p->selected = llama_sample_dist(cur_p, ctx->rng);
}
@ -912,9 +939,8 @@ static const char * llama_sampler_temp_name(const struct llama_sampler * /*smpl*
static void llama_sampler_temp_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
const auto * ctx = (llama_sampler_temp *) smpl->ctx;
for (size_t i = 0; i < cur_p->size; ++i) {
cur_p->data[i].logit /= ctx->temp;
}
llama_sampler_temp_impl(cur_p, ctx->temp);
}
static struct llama_sampler * llama_sampler_temp_clone(const struct llama_sampler * smpl) {
@ -961,6 +987,7 @@ static void llama_sampler_temp_ext_apply(struct llama_sampler * smpl, llama_toke
if (ctx->delta > 0) {
const float min_temp = std::max(0.0f, ctx->temp - ctx->delta);
const float max_temp = ctx->temp + ctx->delta;
float exponent_val = ctx->exponent;
// no need to do anything if there is only one (or zero) candidates
@ -998,9 +1025,7 @@ static void llama_sampler_temp_ext_apply(struct llama_sampler * smpl, llama_toke
#endif
// Apply the dynamically calculated temperature scaling
for (size_t i = 0; i < cur_p->size; ++i) {
cur_p->data[i].logit /= dyn_temp;
}
llama_sampler_temp_impl(cur_p, dyn_temp);
// Re-compute softmax probabilities after scaling logits with dynamic temperature
const double max_l_double = cur_p->data[0].logit;
@ -1024,9 +1049,7 @@ static void llama_sampler_temp_ext_apply(struct llama_sampler * smpl, llama_toke
}
#endif
} else {
for (size_t i = 0; i < cur_p->size; ++i) {
cur_p->data[i].logit /= ctx->temp;
}
llama_sampler_temp_impl(cur_p, ctx->temp);
}
}
@ -1059,6 +1082,101 @@ struct llama_sampler * llama_sampler_init_temp_ext(float temp, float delta, floa
};
}
// xtc
struct llama_sampler_xtc {
const float probability;
const float threshold;
const size_t min_keep;
const uint32_t seed;
uint32_t seed_cur;
std::mt19937 rng;
};
static const char * llama_sampler_xtc_name(const struct llama_sampler * /*smpl*/) {
return "xtc";
}
static void llama_sample_xtc_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
auto * ctx = (llama_sampler_xtc *) smpl->ctx;
if (ctx->probability <= 0.0f
|| ctx->threshold > 0.5f
|| cur_p->size < 2) {
return;
}
std::uniform_real_distribution<float> distribution(0.0f, 1.0f);
float chance = distribution(ctx->rng);
if (chance > ctx->probability) return;
// in case it's not sorted/recalculated yet
llama_sampler_softmax_impl(cur_p);
int pos_last = 0;
for (size_t i = 0; i < cur_p->size; ++i) {
if (cur_p->data[i].p >= ctx->threshold) {
pos_last = i;
} else break;
}
if (cur_p->size - pos_last >= ctx->min_keep && pos_last > 0) {
cur_p->data += pos_last;
cur_p->size -= pos_last;
}
}
static struct llama_sampler * llama_sampler_xtc_clone(const struct llama_sampler * smpl) {
const auto * ctx = (const llama_sampler_xtc *) smpl->ctx;
auto * result = llama_sampler_init_xtc(ctx->probability, ctx->threshold, ctx->min_keep, ctx->seed);
// copy the state
{
auto * result_ctx = (llama_sampler_xtc *) result->ctx;
result_ctx->rng = ctx->rng;
}
return result;
}
static void llama_sampler_xtc_free(struct llama_sampler * smpl) {
delete (llama_sampler_xtc *) smpl->ctx;
}
static void llama_sampler_xtc_reset(struct llama_sampler * smpl) {
auto * ctx = (llama_sampler_xtc *) smpl->ctx;
ctx->seed_cur = get_rng_seed(ctx->seed);
ctx->rng.seed(ctx->seed_cur);
}
static struct llama_sampler_i llama_sampler_xtc_i = {
/* .name = */ llama_sampler_xtc_name,
/* .accept = */ nullptr,
/* .apply = */ llama_sample_xtc_apply,
/* .reset = */ llama_sampler_xtc_reset,
/* .clone = */ llama_sampler_xtc_clone,
/* .free = */ llama_sampler_xtc_free,
};
struct llama_sampler * llama_sampler_init_xtc(float p, float t, size_t min_keep, uint32_t seed) {
auto seed_cur = get_rng_seed(seed);
return new llama_sampler {
/* .iface = */ &llama_sampler_xtc_i,
/* .ctx = */ new llama_sampler_xtc {
/* .probability = */ p,
/* .threshold = */ t,
/* .min_keep = */ min_keep,
/* .seed = */ seed,
/* .seed_cur = */ seed_cur,
/* .rng = */ std::mt19937(seed_cur),
},
};
}
// mirostat
struct llama_sampler_mirostat {
@ -1565,6 +1683,397 @@ struct llama_sampler * llama_sampler_init_penalties(
};
}
// DRY
struct llama_sampler_dry {
int32_t total_context_size;
const float dry_multiplier;
const float dry_base;
const int32_t dry_allowed_length;
const int32_t dry_penalty_last_n;
std::unordered_multimap<llama_token, std::vector<llama_token>> dry_processed_breakers;
std::vector<int> dry_repeat_count;
std::unordered_map<llama_token, int> dry_max_token_repeat;
ring_buffer<llama_token> last_tokens;
};
// Ported from Koboldcpp, original PR: https://github.com/LostRuins/koboldcpp/pull/982 (Original author: pi6am)
static void get_overlapping_token_sequences(const llama_vocab & vocab, const std::string& str, std::unordered_multimap<llama_token, std::vector<llama_token>>& token_sequences, int max_tail_len = -1) {
for (llama_token token_id = 0; token_id < (llama_token)vocab.n_vocab; token_id++) {
std::string word = llama_detokenize(vocab, {token_id}, true);
if (word.find(str) != std::string::npos) {
token_sequences.emplace(token_id, std::vector<llama_token>());
} else {
size_t word_len = word.size(), str_len = str.size();
size_t pos = -1;
while ((pos = word.find(str[0], pos + 1)) != std::string::npos) {
bool match = true;
size_t i;
for (i = 1; i < str_len && i + pos < word_len; ++i) {
if (word[pos + i] != str[i]) {
match = false;
break;
}
}
if (match) {
std::vector<llama_token> tokenization = llama_tokenize_internal(vocab, str.substr(i), false, false);
if (max_tail_len >= 0 && tokenization.size() > (size_t)max_tail_len) {
tokenization.resize(max_tail_len);
}
// Ensure we don't already have a duplicate matching tokenization
auto its = token_sequences.equal_range(token_id);
bool found = false;
for (auto it = its.first; it != its.second; ++it) {
if (tokenization == it->second) {
found = true;
break;
}
}
if (!found) {
token_sequences.emplace(token_id, tokenization);
}
}
}
}
}
}
static const char * llama_sampler_dry_name(const struct llama_sampler * /*smpl*/) {
return "dry";
}
static void llama_sampler_dry_accept(struct llama_sampler * smpl, llama_token token) {
auto * ctx = (llama_sampler_dry *) smpl->ctx;
if (ctx->dry_multiplier == 0.0f || ctx->dry_base < 1.0f || ctx->dry_penalty_last_n == 0) {
return;
}
ctx->last_tokens.push_back(token);
}
// Ported from Koboldcpp, original PR: https://github.com/LostRuins/koboldcpp/pull/982 (Original author: pi6am)
static void llama_sampler_dry_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
auto * ctx = (llama_sampler_dry *) smpl->ctx;
if (ctx->dry_multiplier == 0.0f || ctx->dry_base < 1.0f || ctx->dry_penalty_last_n == 0) {
return;
}
int32_t effective_dry_penalty_last_n = (ctx->dry_penalty_last_n == -1) ? ctx->total_context_size : std::max(ctx->dry_penalty_last_n, 0);
int last_n_repeat = std::min(std::min((int)ctx->last_tokens.size(), effective_dry_penalty_last_n), ctx->total_context_size);
if (last_n_repeat <= ctx->dry_allowed_length) {
return;
}
ctx->dry_repeat_count.assign(last_n_repeat, 0);
ctx->dry_max_token_repeat.clear();
// Step 1: Look for restart sequences to limit the maximum repetition length.
// Work backwards through the context looking for any token that begins a restart sequence.
//
// The collection `restart_sequences` is a mapping from a "head" token to all "tail"
// sequences that together comprise a restart sequence. This allows us to quickly check
// whether each token is the head of a complete sequence. Most restart sequences are actually
// a single token, and for these the "tail" is an empty vector.
//
// If the token is a "head", test all restart sequences that begin with this token
// (there will often only be one sequence for each token, but if sequences like 'aaaq1' and
// 'aaa1' are used as restart strings, both could start with 'aaa' when tokenized). The
// longest matching sequence (if any) is used to limit the maximum repetition length.
//
// Note that in the case case of a short sequence contained in a longer one, this might fail to
// find the smallest value for `rep_limit`. For example, if 'amniotic' and 'ni' are both used as
// restart sequences, 'ni' will be found first, and since it's shorter it will fail to suppress
// 'otic'. This is a minor issue since fully contained restart sequences are likely to be rare.
//
// This is theoretically worst-case O(N^2) for arbitrary restart sequences, which is why we
// have already clamped the maximum tail sequence length when generating `restart_sequences`.
// With clamping, this scan is O(N) in the context length.
int rep_limit = last_n_repeat;
for (int i = 0; i < last_n_repeat; ++i) {
llama_token token = ctx->last_tokens.rat(i);
auto its = ctx->dry_processed_breakers.equal_range(token);
if (its.first == ctx->dry_processed_breakers.end()) {
continue;
}
int longest_match = -1;
for (auto it = its.first; it != its.second; ++it) {
// Note that (*it) does not contain the head character, so seq_len will be
// the restart sequence length minus 1.
// In the common case of a single-token restart sequence, (*it) will be empty
// and we will trivially match.
int seq_len = (int)it->second.size();
if (seq_len > longest_match && seq_len <= (int)i) {
bool match = true;
for (int offset = 0; offset < seq_len; ++offset) {
// The -1 when indexing `last_tokens` is because we already matched the head.
if (it->second[offset] != ctx->last_tokens.rat(i - offset - 1)) {
match = false;
break;
}
}
if (match) {
longest_match = seq_len;
}
}
}
if (longest_match >= 0) {
// We found a restart sequence starting `i` tokens from the end and continuing for
// `longest_match` tokens.
rep_limit = i - longest_match;
break;
}
}
if (rep_limit < ctx->dry_allowed_length) {
return;
}
// Step 2: Iterate in reverse over the last N tokens of the context, using the "Z-algorithm" (in
// the reverse direction) to efficiently compute the positions and lengths of suffixes appearing
// elsewhere in the context. We limit the suffix length to `rep_limit` to respect restart sequences.
//
// This algorithm is not currently documented on Wikipedia, but there is a clear description here:
// https://ivanyu.me/blog/2014/10/15/z-algorithm/
//
// The code below is adapted from the public domain implementation by the same author here:
// https://github.com/ivanyu/string-algorithms/blob/master/z_algorithm.py
//
// Example:
// Last N tokens: a b c c b c y a b c
// Repeat counts: 0 0 3 1 0 2 0 0 0 0
// ^
// This `3` means that the last three tokens of the context (a b c) also appear here.
//
// This step is worst case O(N) since the Z-algorithm is linear, despite the appearance of nested
// for/while loops. This can be seen by observing that the `lt` and `rt` bounds are set after each
// repeated suffix is detected (i.e. after each while loop when n > 0). These bound variables
// ensure that the inner while loops only examine each token in the context once as the outer
// for loop iterates over the context.
{
const int last = last_n_repeat - 1;
int rt = 0, lt = 0;
for (int k = 1; k < last_n_repeat; ++k) {
if (k > rt) {
// If k is outside the current Z-box, do naive computation.
int n = 0;
while (n + k < last_n_repeat && ctx->last_tokens.rat(n) == ctx->last_tokens.rat(n+k)) {
++n;
}
ctx->dry_repeat_count[last - k] = std::min(n, rep_limit);
if (n > 0) {
lt = k;
rt = k+n-1;
}
} else {
// If k is inside the current Z-box, consider two cases.
int p = k - lt; // Pair index.
int right_part_len = rt - k + 1;
if (ctx->dry_repeat_count[last - p] < right_part_len) {
int n = std::min(ctx->dry_repeat_count[last - p], rep_limit);
ctx->dry_repeat_count[last - k] = n;
} else {
int i = rt + 1;
while (i < last_n_repeat && ctx->last_tokens.rat(i) == ctx->last_tokens.rat(i - k)) {
i += 1;
}
int n = std::min(i - k, rep_limit);
ctx->dry_repeat_count[last - k] = n;
lt = k;
rt = i - 1;
}
}
}
}
// Step 3: Iterate over dry_repeat_count and last_tokens, examining the maximum repeat length
// that would be generated by emitting each new token that would extend a sequence.
//
// Following the same example as above:
// Last N tokens: a b c c b c y a b c
// Repeat counts: 0 0 3 1 0 2 0 0 0 0
//
// For each non-zero, look ahead one token. This token, if emitted, would extend the repetition.
// c: 3 -> 4 (from `a b c` to `a b c c`)
// b: 1 -> 2 (from `c` to `c b`)
// y: 2 -> 3 (from `b c` to `b c y`)
for (int i = 0; i < last_n_repeat - 1; ++i) {
int repeat_len = ctx->dry_repeat_count[i];
if (repeat_len >= ctx->dry_allowed_length) {
// This token ends a repeat, so the next token would continue one.
// By convention, the value of `repeat_len` only includes the tokens currently
// in the context, not the new token that would be added.
llama_token token = ctx->last_tokens.rat(last_n_repeat - 2 - i);
// Track the maximum sequence ending in this token.
const auto& it = ctx->dry_max_token_repeat.find(token);
if (it == ctx->dry_max_token_repeat.end() || it->second < repeat_len) {
ctx->dry_max_token_repeat[token] = repeat_len;
}
}
}
// Step 4: Apply logit penalties based on the maximum repeat length for relevant tokens.
// Prevent floating point overflow in `pow(penalty_base, exponent)` by clamping to `max_exponent`.
// Compute it from `penalty_base` and the approximate log of `std::numeric_limits<float>::max()`
const float FLOAT_MAX_LOG = 88.7228391f;
int max_exponent = 0;
if (ctx->dry_base > 1.000001f) {
max_exponent = FLOAT_MAX_LOG / std::log(ctx->dry_base);
}
for (size_t i = 0; i < cur_p->size; ++i) {
const auto& af_kvp = ctx->dry_max_token_repeat.find(cur_p->data[i].id);
if (af_kvp != ctx->dry_max_token_repeat.end()) {
// Check all sequence breakers starting with this token
auto range = ctx->dry_processed_breakers.equal_range(cur_p->data[i].id);
bool is_single_token_breaker = false;
for (auto it = range.first; it != range.second; ++it) {
if (it->second.empty()) {
is_single_token_breaker = true;
break;
}
}
// Apply penalty only if it's not a single-token sequence breaker
if (!is_single_token_breaker) {
int repeat_exp = af_kvp->second - ctx->dry_allowed_length;
if (max_exponent > 0 && repeat_exp > max_exponent) {
repeat_exp = max_exponent;
}
float penalty = ctx->dry_multiplier * std::pow(ctx->dry_base, repeat_exp);
cur_p->data[i].logit -= penalty;
}
}
}
cur_p->sorted = false;
}
static void llama_sampler_dry_reset(struct llama_sampler * smpl) {
auto * ctx = (llama_sampler_dry *) smpl->ctx;
ctx->last_tokens.clear();
ctx->dry_repeat_count.clear();
ctx->dry_max_token_repeat.clear();
}
static struct llama_sampler * llama_sampler_dry_clone(const struct llama_sampler * smpl) {
const auto * ctx = (llama_sampler_dry *) smpl->ctx;
// nullptr is passed as vocab because it is only needed for raw sequence breaker processing, which we have already done and will be copying
auto * result = llama_sampler_init_dry(nullptr, ctx->dry_multiplier, ctx->dry_base, ctx->dry_allowed_length, ctx->dry_penalty_last_n, NULL, 0);
// Copy the state, including the processed breakers
{
auto * result_ctx = (llama_sampler_dry *) result->ctx;
result_ctx->dry_processed_breakers = ctx->dry_processed_breakers;
result_ctx->dry_repeat_count = ctx->dry_repeat_count;
result_ctx->dry_max_token_repeat = ctx->dry_max_token_repeat;
result_ctx->last_tokens = ctx->last_tokens;
}
return result;
}
static void llama_sampler_dry_free(struct llama_sampler * smpl) {
delete (llama_sampler_dry *) smpl->ctx;
}
static struct llama_sampler_i llama_sampler_dry_i = {
/* .name = */ llama_sampler_dry_name,
/* .accept = */ llama_sampler_dry_accept,
/* .apply = */ llama_sampler_dry_apply,
/* .reset = */ llama_sampler_dry_reset,
/* .clone = */ llama_sampler_dry_clone,
/* .free = */ llama_sampler_dry_free,
};
struct llama_sampler * llama_sampler_init_dry_impl(const struct llama_vocab & vocab, int32_t context_size, float dry_multiplier, float dry_base, int32_t dry_allowed_length, int32_t dry_penalty_last_n, const char** seq_breakers, size_t num_breakers) {
int32_t effective_dry_penalty_last_n = (dry_penalty_last_n == -1) ? context_size : std::max(dry_penalty_last_n, 0);
std::unordered_multimap<llama_token, std::vector<llama_token>> processed_breakers;
const int MAX_CHAR_LEN = 40;
const int MAX_SEQ_LEN = 20;
const bool dry_enabled = (dry_multiplier != 0.0f && dry_base >= 1.0f && dry_penalty_last_n != 0);
if (dry_enabled && seq_breakers != nullptr && num_breakers > 0) {
// Process sequence breakers
for (size_t i = 0; i < num_breakers; ++i) {
if (seq_breakers[i] == nullptr || std::strlen(seq_breakers[i]) == 0) {
LLAMA_LOG_WARN("skipping null or empty DRY sequence breaker at index %zu\n", i);
continue;
}
std::string sequence_break(seq_breakers[i]);
if (sequence_break.empty()) {
LLAMA_LOG_WARN("skipping empty DRY sequence breaker\n");
continue;
}
if (sequence_break.size() > MAX_CHAR_LEN) {
LLAMA_LOG_WARN("truncating DRY sequence breaker to %d characters\n", MAX_CHAR_LEN);
sequence_break.resize(MAX_CHAR_LEN);
}
get_overlapping_token_sequences(vocab, sequence_break, processed_breakers, MAX_SEQ_LEN);
}
}
return new llama_sampler {
/* .iface = */ &llama_sampler_dry_i,
/* .ctx = */ new llama_sampler_dry {
/* .total_context_size = */ context_size,
/* .dry_multiplier = */ dry_multiplier,
/* .dry_base = */ dry_base,
/* .dry_allowed_length = */ dry_allowed_length,
/* .dry_penalty_last_n = */ dry_penalty_last_n,
/* .dry_processed_breakers = */ std::move(processed_breakers),
/* .dry_repeat_count = */ dry_enabled ? std::vector<int>(effective_dry_penalty_last_n, 0) : std::vector<int>{},
/* .dry_max_token_repeat = */ {},
/* .last_tokens = */ dry_enabled ? ring_buffer<llama_token>(effective_dry_penalty_last_n) : ring_buffer<llama_token>(0),
},
};
}
// wrapper for test-sampling.cpp
struct llama_sampler * llama_sampler_init_dry_testing(int32_t context_size, float dry_multiplier, float dry_base, int32_t dry_allowed_length, int32_t dry_penalty_last_n, const std::vector<std::vector<llama_token>>& seq_breakers) {
llama_vocab dummy_vocab;
auto * result = llama_sampler_init_dry_impl(dummy_vocab, context_size, dry_multiplier, dry_base, dry_allowed_length, dry_penalty_last_n, NULL, 0);
auto * ctx = (llama_sampler_dry *) result->ctx;
// Process the token-based sequence breakers
ctx->dry_processed_breakers.clear();
if (seq_breakers.empty()) {
LLAMA_LOG_WARN("empty DRY sequence breakers list in llama_sampler_init_dry_testing\n");
} else {
for (const auto& breaker : seq_breakers) {
if (breaker.empty()) {
LLAMA_LOG_WARN("skipping DRY empty sequence breaker\n");
continue;
}
llama_token head_token = breaker[0];
std::vector<llama_token> tail_tokens(breaker.begin() + 1, breaker.end());
ctx->dry_processed_breakers.emplace(head_token, std::move(tail_tokens));
}
if (ctx->dry_processed_breakers.empty()) {
LLAMA_LOG_WARN("no valid DRY sequence breakers processed in llama_sampler_init_dry_testing\n");
}
}
return result;
}
// logit-bias
struct llama_sampler_logit_bias {
@ -1644,6 +2153,229 @@ struct llama_sampler * llama_sampler_init_logit_bias(
};
}
// infill
//#define GGML_DEBUG_SAMPLER_INFILL
struct llama_sampler_infill {
const struct llama_vocab * vocab;
std::vector<char> buf0;
std::vector<char> buf1;
};
static const char * llama_sampler_infill_name(const struct llama_sampler * /*smpl*/) {
return "infill";
}
static void llama_sampler_infill_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
auto * ctx = (llama_sampler_infill *) smpl->ctx;
llama_sampler_softmax_impl(cur_p);
#if defined(GGML_DEBUG_SAMPLER_INFILL)
#define LOG_DBG_CUR LLAMA_LOG_DEBUG
#else
#define LOG_DBG_CUR(...)
#endif
for (size_t i = 0; i < cur_p->size; ++i) {
LOG_DBG_CUR("%s: cur_p[%3zu] = { id: %6d, p: %.6f, logit: %6.3f }\n", __func__, i, cur_p->data[i].id, cur_p->data[i].p, cur_p->data[i].logit);
}
float p_txt_sum = 0.0f;
float p_eog_sum = 0.0f;
for (size_t i = 0; i < cur_p->size; ++i) {
if (llama_token_is_eog_impl(*ctx->vocab, cur_p->data[i].id)) {
p_eog_sum += cur_p->data[i].p;
} else {
p_txt_sum += cur_p->data[i].p;
}
}
const float rat = p_eog_sum == 0.0 ? INFINITY : p_txt_sum / p_eog_sum; GGML_UNUSED(rat);
LOG_DBG_CUR("%s: p_txt_sum = %.2f, p_eog_sum = %.2f, rat = %.2f, n = %zu\n", __func__, p_txt_sum, p_eog_sum, rat, cur_p->size);
if (3*p_eog_sum*cur_p->size > p_txt_sum) {
LOG_DBG_CUR("%s: the ratio p_txt/p_eog = %.2f is too low -> sampling EOG\n", __func__, p_txt_sum/p_eog_sum);
// keep just the EOG tokens
const auto size_org = cur_p->size;
cur_p->size = 0;
float p_sum = 0.0f;
for (size_t i = 0; i < size_org; ++i) {
if (llama_token_is_eog_impl(*ctx->vocab, cur_p->data[i].id)) {
p_sum += cur_p->data[i].p;
cur_p->data[cur_p->size++] = cur_p->data[i];
}
}
// normalize probs
for (size_t i = 0; i < cur_p->size; ++i) {
cur_p->data[i].p /= p_sum;
}
return;
}
size_t n_combined = 0; GGML_UNUSED(n_combined);
// combine tokens with common prefix
for (size_t i0 = 0; i0 < cur_p->size; ++i0) {
for (size_t i1 = 0; i1 < cur_p->size; ++i1) {
if (cur_p->data[i0].logit == -INFINITY) {
break;
}
if (i0 == i1 || cur_p->data[i1].logit == -INFINITY) {
continue;
}
int len0 = llama_token_to_piece_impl(*ctx->vocab, cur_p->data[i0].id, ctx->buf0.data(), ctx->buf0.size(), 0, false);
if (len0 < 0) {
ctx->buf0.resize(len0);
len0 = llama_token_to_piece_impl(*ctx->vocab, cur_p->data[i0].id, ctx->buf0.data(), ctx->buf0.size(), 0, false);
assert(len0 > 0);
}
int len1 = llama_token_to_piece_impl(*ctx->vocab, cur_p->data[i1].id, ctx->buf1.data(), ctx->buf1.size(), 0, false);
if (len1 < 0) {
ctx->buf1.resize(len1);
len1 = llama_token_to_piece_impl(*ctx->vocab, cur_p->data[i1].id, ctx->buf1.data(), ctx->buf1.size(), 0, false);
assert(len1 > 0);
}
// token i0 is a prefix of token i1
if (len0 > 0 && len0 <= len1 && memcmp(ctx->buf0.data(), ctx->buf1.data(), len0) == 0) {
int dst = i0;
int src = i1;
// merge into the token with higher probability
if (cur_p->data[i1].p > cur_p->data[i0].p) {
std::swap(dst, src);
}
cur_p->data[dst].p += cur_p->data[src].p;
cur_p->data[src].logit = -INFINITY;
cur_p->data[src].p = 0.0f;
n_combined++;
}
}
}
size_t n_non_eog = 0;
size_t size_org = cur_p->size;
float p_sum = 0.0f;
float thold = 0.2f;
cur_p->size = 0;
LOG_DBG_CUR("%s: n_combined = %zu, applying thold = %.3f\n", __func__, n_combined, thold);
for (size_t i = 0; i < size_org; ++i) {
const bool is_eog = llama_token_is_eog_impl(*ctx->vocab, cur_p->data[i].id);
if (cur_p->data[i].p < thold && !is_eog) {
continue;
}
if (!is_eog) {
++n_non_eog;
}
p_sum += cur_p->data[i].p;
// keep this token
cur_p->data[cur_p->size++] = cur_p->data[i];
}
LOG_DBG_CUR("%s: n_non_eog = %zu\n", __func__, n_non_eog);
// if no non-EOG tokens are left -> reduce cur_p to single EOT token
if (n_non_eog == 0) {
cur_p->size = 1;
cur_p->data[0].id = llama_token_eot_impl(*ctx->vocab);
cur_p->data[0].logit = 1.0f;
return;
}
// normalize probs
for (size_t i = 0; i < cur_p->size; ++i) {
cur_p->data[i].p /= p_sum;
LOG_DBG_CUR("%s: cur_p[%3zu] = { id: %6d, p: %.6f, logit: %6.3f }\n", __func__, i, cur_p->data[i].id, cur_p->data[i].p, cur_p->data[i].logit);
}
size_org = cur_p->size;
p_sum = 0.0f;
thold = 1.0/(n_non_eog + 1);
cur_p->size = 0;
LOG_DBG_CUR("%s: applying thold = %.3f\n", __func__, thold);
for (size_t i = 0; i < size_org; ++i) {
const bool is_eog = llama_token_is_eog_impl(*ctx->vocab, cur_p->data[i].id);
if (cur_p->data[i].p < thold && !is_eog) {
continue;
}
p_sum += cur_p->data[i].p;
cur_p->data[cur_p->size++] = cur_p->data[i];
}
// normalize probs
for (size_t i = 0; i < cur_p->size; ++i) {
cur_p->data[i].p /= p_sum;
LOG_DBG_CUR("%s: cur_p[%3zu] = { id: %6d, p: %.6f, logit: %6.3f }\n", __func__, i, cur_p->data[i].id, cur_p->data[i].p, cur_p->data[i].logit);
}
#undef LOG_DBG_CUR
}
static struct llama_sampler * llama_sampler_infill_clone(const struct llama_sampler * smpl) {
const auto * ctx = (const llama_sampler_infill *) smpl->ctx;
return llama_sampler_init_infill_impl(*ctx->vocab);
}
static void llama_sampler_infill_free(struct llama_sampler * smpl) {
delete (llama_sampler_infill *) smpl->ctx;
}
static struct llama_sampler_i llama_sampler_infill_i = {
/* .name = */ llama_sampler_infill_name,
/* .accept = */ nullptr,
/* .apply = */ llama_sampler_infill_apply,
/* .reset = */ nullptr,
/* .clone = */ llama_sampler_infill_clone,
/* .free = */ llama_sampler_infill_free,
};
struct llama_sampler * llama_sampler_init_infill_impl(
const struct llama_vocab & vocab) {
return new llama_sampler {
/* .iface = */ &llama_sampler_infill_i,
/* .ctx = */ new llama_sampler_infill {
/* .vocab = */ &vocab,
/* .buf0 = */ std::vector<char>(512),
/* .buf1 = */ std::vector<char>(512),
},
};
}
// utils
uint32_t llama_sampler_get_seed(const struct llama_sampler * smpl) {

23
src/llama-sampling.h
View File

@ -4,8 +4,6 @@
#include "llama-grammar.h"
#include <unordered_map>
struct llama_vocab;
struct llama_grammar;
@ -27,3 +25,24 @@ struct llama_sampler * llama_sampler_init_grammar_impl(
const struct llama_vocab & vocab,
const char * grammar_str,
const char * grammar_root);
struct llama_sampler * llama_sampler_init_infill_impl(
const struct llama_vocab & vocab);
struct llama_sampler * llama_sampler_init_dry_impl(
const struct llama_vocab & vocab,
int32_t context_size,
float dry_multiplier,
float dry_base,
int32_t dry_allowed_length,
int32_t dry_penalty_last_n,
const char ** seq_breakers,
size_t num_breakers);
struct llama_sampler * llama_sampler_init_dry_testing(
int32_t context_size,
float dry_multiplier,
float dry_base,
int32_t dry_allowed_length,
int32_t dry_penalty_last_n,
const std::vector<std::vector<llama_token>>& seq_breakers);

58
src/llama-vocab.cpp
View File

@ -221,7 +221,7 @@ struct llm_tokenizer_spm_session {
}
// seed the work queue with all possible 2-character tokens.
for (size_t i = 1; i < symbols.size(); ++i) {
for (int i = 1; i < (int) symbols.size(); ++i) {
try_add_bigram(i - 1, i);
}
@ -563,7 +563,7 @@ struct llm_tokenizer_bpe_session {
index++;
symbols.emplace_back(sym);
}
for (size_t i = 1; i < symbols.size(); ++i) {
for (int i = 1; i < (int) symbols.size(); ++i) {
add_new_bigram(i - 1, i);
}
@ -1663,6 +1663,14 @@ llama_token llama_token_eos_impl(const struct llama_vocab & vocab) {
return vocab.special_eos_id;
}
llama_token llama_token_eot_impl(const struct llama_vocab & vocab) {
return vocab.special_eot_id;
}
llama_token llama_token_eom_impl(const struct llama_vocab & vocab) {
return vocab.special_eom_id;
}
llama_token llama_token_cls_impl(const struct llama_vocab & vocab) {
return vocab.special_cls_id;
}
@ -1688,23 +1696,39 @@ bool llama_add_eos_token_impl(const struct llama_vocab & vocab) {
}
llama_token llama_token_prefix_impl(const struct llama_vocab & vocab) {
return vocab.special_prefix_id;
return vocab.special_fim_pre_id;
}
llama_token llama_token_middle_impl(const struct llama_vocab & vocab) {
return vocab.special_middle_id;
return vocab.special_fim_mid_id;
}
llama_token llama_token_suffix_impl(const struct llama_vocab & vocab) {
return vocab.special_suffix_id;
return vocab.special_fim_suf_id;
}
llama_token llama_token_eot_impl(const struct llama_vocab & vocab) {
return vocab.special_eot_id;
llama_token llama_token_fim_pre_impl(const struct llama_vocab & vocab) {
return vocab.special_fim_pre_id;
}
llama_token llama_token_eom_impl(const struct llama_vocab & vocab) {
return vocab.special_eom_id;
llama_token llama_token_fim_suf_impl(const struct llama_vocab & vocab) {
return vocab.special_fim_suf_id;
}
llama_token llama_token_fim_mid_impl(const struct llama_vocab & vocab) {
return vocab.special_fim_mid_id;
}
llama_token llama_token_fim_pad_impl(const struct llama_vocab & vocab) {
return vocab.special_fim_pad_id;
}
llama_token llama_token_fim_rep_impl(const struct llama_vocab & vocab) {
return vocab.special_fim_rep_id;
}
llama_token llama_token_fim_sep_impl(const struct llama_vocab & vocab) {
return vocab.special_fim_sep_id;
}
int32_t llama_tokenize_impl(
@ -1942,3 +1966,19 @@ int32_t llama_detokenize_impl(
return total <= text_len_max ? total : -total;
}
std::string llama_detokenize(const struct llama_vocab & vocab, const std::vector<llama_token> & tokens, bool special) {
std::string text;
text.resize(std::max(text.capacity(), tokens.size()));
int32_t n_chars = llama_detokenize_impl(vocab, tokens.data(), (int32_t)tokens.size(), &text[0], (int32_t)text.size(), false, special);
if (n_chars < 0) {
text.resize(-n_chars);
n_chars = llama_detokenize_impl(vocab, tokens.data(), (int32_t)tokens.size(), &text[0], (int32_t)text.size(), false, special);
GGML_ASSERT(n_chars <= (int32_t)text.size()); // whitespace trimming is performed after per-token detokenization
}
text.resize(n_chars);
// NOTE: the original tokenizer decodes bytes after collecting the pieces.
return text;
}

46
src/llama-vocab.h
View File

@ -37,20 +37,26 @@ struct llama_vocab {
std::map<std::pair<std::string, std::string>, int> bpe_ranks;
// default LLaMA special tokens
// TODO: should we set all of these to LLAMA_TOKEN_NULL?
id special_bos_id = 1;
id special_eos_id = 2;
id special_eot_id = LLAMA_TOKEN_NULL;
id special_eom_id = LLAMA_TOKEN_NULL;
id special_unk_id = 0;
id special_sep_id = LLAMA_TOKEN_NULL;
id special_pad_id = LLAMA_TOKEN_NULL;
id special_cls_id = LLAMA_TOKEN_NULL;
id special_mask_id = LLAMA_TOKEN_NULL;
id linefeed_id = 13;
id special_prefix_id = LLAMA_TOKEN_NULL;
id special_suffix_id = LLAMA_TOKEN_NULL;
id special_middle_id = LLAMA_TOKEN_NULL;
id special_eot_id = LLAMA_TOKEN_NULL; // TODO: move above after "eos_id", and here add "file separator" token
id special_eom_id = LLAMA_TOKEN_NULL;
id linefeed_id = 13;
// fim tokens
id special_fim_pre_id = LLAMA_TOKEN_NULL;
id special_fim_suf_id = LLAMA_TOKEN_NULL;
id special_fim_mid_id = LLAMA_TOKEN_NULL;
id special_fim_pad_id = LLAMA_TOKEN_NULL;
id special_fim_rep_id = LLAMA_TOKEN_NULL; // repo
id special_fim_sep_id = LLAMA_TOKEN_NULL; // file separator
// set of all tokens that cause "end of generation"
std::set<id> special_eog_ids;
@ -104,19 +110,26 @@ bool llama_token_is_control_impl(const struct llama_vocab & vocab, llama_token t
llama_token llama_token_bos_impl(const struct llama_vocab & vocab);
llama_token llama_token_eos_impl(const struct llama_vocab & vocab);
llama_token llama_token_eot_impl(const struct llama_vocab & vocab);
llama_token llama_token_eom_impl(const struct llama_vocab & vocab);
llama_token llama_token_cls_impl(const struct llama_vocab & vocab);
llama_token llama_token_sep_impl(const struct llama_vocab & vocab);
llama_token llama_token_nl_impl (const struct llama_vocab & vocab);
llama_token llama_token_pad_impl(const struct llama_vocab & vocab);
bool llama_add_bos_token_impl(const struct llama_vocab & vocab);
bool llama_add_eos_token_impl(const struct llama_vocab & vocab);
llama_token llama_token_prefix_impl(const struct llama_vocab & vocab);
llama_token llama_token_middle_impl(const struct llama_vocab & vocab);
llama_token llama_token_suffix_impl(const struct llama_vocab & vocab);
llama_token llama_token_eot_impl (const struct llama_vocab & vocab);
llama_token llama_token_eom_impl (const struct llama_vocab & vocab);
llama_token llama_token_fim_pre_impl(const struct llama_vocab & vocab);
llama_token llama_token_fim_suf_impl(const struct llama_vocab & vocab);
llama_token llama_token_fim_mid_impl(const struct llama_vocab & vocab);
llama_token llama_token_fim_pad_impl(const struct llama_vocab & vocab);
llama_token llama_token_fim_rep_impl(const struct llama_vocab & vocab);
llama_token llama_token_fim_sep_impl(const struct llama_vocab & vocab);
bool llama_add_bos_token_impl(const struct llama_vocab & vocab);
bool llama_add_eos_token_impl(const struct llama_vocab & vocab);
int32_t llama_tokenize_impl(
const struct llama_vocab & vocab,
@ -136,6 +149,12 @@ int32_t llama_token_to_piece_impl(
int32_t lstrip,
bool special);
// check if token0 is contained as a prefix in token1
bool llama_token_is_prefix_impl(
const struct llama_vocab & vocab,
llama_token token0,
llama_token token1);
int32_t llama_detokenize_impl(
const struct llama_vocab & vocab,
const llama_token * tokens,
@ -144,3 +163,8 @@ int32_t llama_detokenize_impl(
int32_t text_len_max,
bool remove_special,
bool unparse_special);
std::string llama_detokenize(
const struct llama_vocab & vocab,
const std::vector<llama_token> & tokens,
bool special);

1069
src/llama.cpp
View File

File diff suppressed because it is too large Load Diff

122
tests/test-backend-ops.cpp
View File

@ -1650,11 +1650,12 @@ struct test_mul_mat : public test_case {
const int64_t m;
const int64_t n;
const int64_t k;
const std::array<int64_t, 2> bs; // dims 3 and 4
const std::array<int64_t, 2> nr; // repeat in dims 3 and 4
const std::array<int64_t, 2> bs; // dims 3 and 4
const std::array<int64_t, 2> nr; // repeat in dims 3 and 4
const std::array<int64_t, 4> per; // permutation of dimensions
std::string vars() override {
return VARS_TO_STR7(type_a, type_b, m, n, k, bs, nr);
return VARS_TO_STR8(type_a, type_b, m, n, k, bs, nr, per);
}
double max_nmse_err() override {
@ -1669,17 +1670,44 @@ struct test_mul_mat : public test_case {
test_mul_mat(ggml_type type_a = GGML_TYPE_F32, ggml_type type_b = GGML_TYPE_F32,
int64_t m = 32, int64_t n = 32, int64_t k = 32,
std::array<int64_t, 2> bs = {10, 10},
std::array<int64_t, 2> nr = {2, 2})
: type_a(type_a), type_b(type_b), m(m), n(n), k(k), bs(bs), nr(nr) {}
std::array<int64_t, 2> nr = {2, 2},
std::array<int64_t, 4> per = {0, 1, 2, 3})
: type_a(type_a), type_b(type_b), m(m), n(n), k(k), bs(bs), nr(nr), per(per) {}
ggml_tensor * build_graph(ggml_context * ctx) override {
// C^T = A * B^T: (k, m) * (k, n) => (m, n)
ggml_tensor * a = ggml_new_tensor_4d(ctx, type_a, k, m, bs[0] , bs[1]);
ggml_tensor * b = ggml_new_tensor_4d(ctx, type_b, k, n, bs[0]*nr[0], bs[1]*nr[1]);
ggml_set_param(ctx, a);
ggml_set_param(ctx, b);
ggml_set_name(a, "a");
ggml_set_name(b, "b");
ggml_tensor * a;
ggml_tensor * b;
const int npermuted = (per[0] != 0) + (per[1] != 1) + (per[2] != 2) + (per[3] != 3);
if (npermuted > 0) {
GGML_ASSERT(npermuted == 2);
GGML_ASSERT(!ggml_is_quantized(type_a) || per[0] == 0);
GGML_ASSERT(!ggml_is_quantized(type_b) || per[0] == 0);
// Create tensors with the permuted dimensions, then permute them back to the dimensions given by m,n,k.
const int64_t ne_a[4] = {k, m, bs[0], bs[1]};
const int64_t ne_b[4] = {k, n, bs[0]*nr[0], bs[1]*nr[1]};
a = ggml_new_tensor_4d(ctx, type_a, ne_a[per[0]], ne_a[per[1]], ne_a[per[2]], ne_a[per[3]]);
b = ggml_new_tensor_4d(ctx, type_b, ne_b[per[0]], ne_b[per[1]], ne_b[per[2]], ne_b[per[3]]);
ggml_set_param(ctx, a);
ggml_set_param(ctx, b);
ggml_set_name(a, "a");
ggml_set_name(b, "b");
a = ggml_permute(ctx, a, per[0], per[1], per[2], per[3]);
b = ggml_permute(ctx, b, per[0], per[1], per[2], per[3]);
ggml_set_name(a, "a_permuted");
ggml_set_name(b, "b_permuted");
} else {
a = ggml_new_tensor_4d(ctx, type_a, k, m, bs[0], bs[1]);
b = ggml_new_tensor_4d(ctx, type_b, k, n, bs[0]*nr[0], bs[1]*nr[1]);
ggml_set_param(ctx, a);
ggml_set_param(ctx, b);
ggml_set_name(a, "a");
ggml_set_name(b, "b");
}
ggml_tensor * out = ggml_mul_mat(ctx, a, b);
ggml_set_name(out, "out");
@ -3308,13 +3336,49 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
}
}
test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F32, GGML_TYPE_F32));
test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F32));
test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16));
// test cases for 1D im2col
// im2col 1D
test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F32, GGML_TYPE_F32, {3000, 128, 1, 1}, {3, 128, 1280, 1}, 1, 0, 1, 0, 1, 0, false));
test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F32, {3000, 128, 1, 1}, {3, 128, 1280, 1}, 1, 0, 1, 0, 1, 0, false));
test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16, {3000, 128, 1, 1}, {3, 128, 1280, 1}, 1, 0, 1, 0, 1, 0, false));
for (int s0 : {1, 3}) {
for (int p0 : {0, 3}) {
for (int d0 : {1, 3}) {
test_cases.emplace_back(new test_im2col(
GGML_TYPE_F32, GGML_TYPE_F32, GGML_TYPE_F32, {20, 2, 2, 1}, {3, 2, 2, 1},
s0, 0, p0, 0, d0, 0, false));
}
}
}
// im2col 2D
test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F32, GGML_TYPE_F32));
test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F32));
test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16));
for (int s0 : {1, 3}) {
for (int s1 : {1, 3}) {
for (int p0 : {0, 3}) {
for (int p1 : {0, 3}) {
for (int d0 : {1, 3}) {
for (int d1 : {1, 3}) {
test_cases.emplace_back(new test_im2col(
GGML_TYPE_F32, GGML_TYPE_F32, GGML_TYPE_F32, {20, 20, 2, 2}, {3, 3, 2, 2},
s0, s1, p0, p1, d0, d1, true));
}
}
}
}
}
}
// extra tests for im2col 2D
test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16, {12, 12, 1, 32}, {3, 3, 1, 32}, 1, 1, 1, 1, 1, 1, true));
test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16, {12, 12, 2, 32}, {3, 3, 2, 32}, 1, 1, 1, 1, 1, 1, true));
test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16, {12, 12, 1, 1024}, {3, 3, 1, 1024}, 1, 1, 1, 1, 1, 1, true));
test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16, {12, 12, 2, 1024}, {3, 3, 2, 1024}, 1, 1, 1, 1, 1, 1, true));
test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16, {12, 12, 1, 2048}, {3, 3, 1, 2048}, 1, 1, 1, 1, 1, 1, true));
test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16, {12, 12, 2, 2048}, {3, 3, 2, 2048}, 1, 1, 1, 1, 1, 1, true));
test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16, {12, 12, 1, 2560}, {3, 3, 1, 2560}, 1, 1, 1, 1, 1, 1, true));
test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16, {12, 12, 2, 2560}, {3, 3, 2, 2560}, 1, 1, 1, 1, 1, 1, true));
// sycl backend will limit task global_range < MAX_INT
// test cases for 2D im2col with large input W and H (occurs in stable-diffusion)
@ -3442,13 +3506,14 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
#if 1
for (ggml_type type_a : base_types) {
for (ggml_type type_b : {GGML_TYPE_F32, GGML_TYPE_F16}) {
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, { 1, 1}, {1, 1}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, {10, 1}, {1, 1}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, {10, 1}, {2, 1}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, {10, 10}, {1, 1}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, {10, 10}, {2, 1}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, {10, 10}, {1, 2}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, {10, 10}, {2, 2}));
// test cases without permutation
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, { 1, 1}, {1, 1}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, {10, 1}, {1, 1}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, {10, 1}, {2, 1}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, {10, 10}, {1, 1}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, {10, 10}, {2, 1}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, {10, 10}, {1, 2}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, {10, 10}, {2, 2}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, { 1, 1}, {1, 1}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {10, 1}, {1, 1}));
@ -3457,6 +3522,19 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {10, 10}, {2, 1}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {10, 10}, {1, 2}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {10, 10}, {2, 2}));
// test cases with permutation
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, {2, 3}, {1, 1}, {0, 2, 1, 3}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, {2, 3}, {1, 1}, {0, 1, 3, 2}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, {2, 3}, {1, 1}, {0, 3, 2, 1}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 8, 256, {2, 3}, {1, 1}, {0, 2, 1, 3}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 8, 256, {2, 3}, {1, 1}, {0, 1, 3, 2}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 8, 256, {2, 3}, {1, 1}, {0, 3, 2, 1}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {2, 3}, {1, 1}, {0, 2, 1, 3}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {2, 3}, {1, 1}, {0, 1, 3, 2}));
test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {2, 3}, {1, 1}, {0, 3, 2, 1}));
}
}
for (ggml_type type_a : other_types) {

21
tests/test-json-schema-to-grammar.cpp
View File

@ -696,7 +696,7 @@ static void test_all(const std::string & lang, std::function<void(const TestCase
"pattern": "^abc?d*efg+(hij)?kl$"
})""",
R"""(
root ::= "\"" "ab" "c"? "d"* "ef" "g"+ ("hij")? "kl" "\"" space
root ::= "\"" ("ab" "c"? "d"* "ef" "g"+ ("hij")? "kl") "\"" space
space ::= | " " | "\n" [ \t]{0,20}
)"""
});
@ -709,7 +709,7 @@ static void test_all(const std::string & lang, std::function<void(const TestCase
"pattern": "^\\[\\]\\{\\}\\(\\)\\|\\+\\*\\?$"
})""",
R"""(
root ::= "\"" "[]{}()|+*?" "\"" space
root ::= "\"" ("[]{}()|+*?") "\"" space
space ::= | " " | "\n" [ \t]{0,20}
)"""
});
@ -722,7 +722,20 @@ static void test_all(const std::string & lang, std::function<void(const TestCase
"pattern": "^\"$"
})""",
R"""(
root ::= "\"" "\"" "\"" space
root ::= "\"" ("\"") "\"" space
space ::= | " " | "\n" [ \t]{0,20}
)"""
});
test({
SUCCESS,
"regexp with top-level alternation",
R"""({
"type": "string",
"pattern": "^A|B|C|D$"
})""",
R"""(
root ::= "\"" ("A" | "B" | "C" | "D") "\"" space
space ::= | " " | "\n" [ \t]{0,20}
)"""
});
@ -736,7 +749,7 @@ static void test_all(const std::string & lang, std::function<void(const TestCase
})""",
R"""(
dot ::= [^\x0A\x0D]
root ::= "\"" ("(" root-1{1,3} ")")? root-1{3,3} "-" root-1{4,4} " " "a"{3,5} "nd" dot dot dot "\"" space
root ::= "\"" (("(" root-1{1,3} ")")? root-1{3,3} "-" root-1{4,4} " " "a"{3,5} "nd" dot dot dot) "\"" space
root-1 ::= [0-9]
space ::= | " " | "\n" [ \t]{0,20}
)"""

311
tests/test-sampling.cpp
View File

@ -10,6 +10,8 @@
#include <string>
#include <vector>
extern struct llama_sampler * llama_sampler_init_dry_testing(int32_t context_size, float dry_multiplier, float dry_base, int32_t dry_allowed_length, int32_t dry_penalty_last_n, const std::vector<std::vector<llama_token>>& seq_breakers);
static void dump(const llama_token_data_array * cur_p) {
for (size_t i = 0; i < cur_p->size; i++) {
printf("%d: %f (%f)\n", cur_p->data[i].id, cur_p->data[i].p, cur_p->data[i].logit);
@ -18,181 +20,199 @@ static void dump(const llama_token_data_array * cur_p) {
#define DUMP(__cur_p) do { printf("%s:%d (%s)\n", __FILE__, __LINE__, __func__); dump((__cur_p)); printf("-\n"); } while(0)
#define APPLY(__cnstr, __cur_p) do { \
auto * cnstr = (__cnstr); \
llama_sampler_apply(cnstr, (__cur_p)); \
llama_sampler_free(cnstr); \
} while(0)
struct sampler_tester {
sampler_tester(size_t n_vocab) {
cur.reserve(n_vocab);
for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) {
const float logit = logf(token_id);
cur.emplace_back(llama_token_data{token_id, logit, 0.0f});
}
static void test_top_k(const std::vector<float> & probs, const std::vector<float> & expected_probs, int k) {
const size_t n_vocab = probs.size();
cur_p = llama_token_data_array { cur.data(), cur.size(), -1, false };
}
sampler_tester(const std::vector<float> & probs, const std::vector<float> & probs_expected) : probs_expected(probs_expected) {
cur.reserve(probs.size());
for (llama_token token_id = 0; token_id < (llama_token)probs.size(); token_id++) {
const float logit = logf(probs[token_id]);
cur.emplace_back(llama_token_data{token_id, logit, probs[token_id]});
}
cur_p = llama_token_data_array { cur.data(), cur.size(), -1, false };
}
void apply(llama_sampler * sampler) {
llama_sampler_apply(sampler, &cur_p);
llama_sampler_free(sampler);
}
void check() {
GGML_ASSERT(cur_p.size == probs_expected.size());
for (size_t i = 0; i < cur_p.size; i++) {
GGML_ASSERT(fabs(cur_p.data[i].p - probs_expected[i]) < 1e-5);
}
}
llama_token_data_array cur_p;
private:
const std::vector<float> probs_expected;
std::vector<llama_token_data> cur;
cur.reserve(n_vocab);
for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) {
const float logit = logf(probs[token_id]);
cur.emplace_back(llama_token_data{token_id, logit, 0.0f});
}
};
llama_token_data_array cur_p = { cur.data(), cur.size(), -1, false };
APPLY(llama_sampler_init_softmax(), &cur_p);
DUMP(&cur_p);
APPLY(llama_sampler_init_top_k(k), &cur_p);
DUMP(&cur_p);
static void test_temp(const std::vector<float> & probs, const std::vector<float> & probs_expected, float temp) {
sampler_tester tester(probs, probs_expected);
GGML_ASSERT(cur_p.size == expected_probs.size());
for (size_t i = 0; i < cur_p.size; i++) {
GGML_ASSERT(fabs(cur_p.data[i].p - expected_probs[i]) < 1e-5);
}
DUMP(&tester.cur_p);
tester.apply(llama_sampler_init_temp(temp));
tester.apply(llama_sampler_init_dist(0));
DUMP(&tester.cur_p);
tester.check();
}
static void test_top_p(const std::vector<float> & probs, const std::vector<float> & expected_probs, float p) {
const size_t n_vocab = probs.size();
static void test_temp_ext(const std::vector<float> & probs, const std::vector<float> & probs_expected, float temp, float delta, float exponent) {
sampler_tester tester(probs, probs_expected);
std::vector<llama_token_data> cur;
cur.reserve(n_vocab);
for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) {
const float logit = logf(probs[token_id]);
cur.emplace_back(llama_token_data{token_id, logit, 0.0f});
}
DUMP(&tester.cur_p);
tester.apply(llama_sampler_init_temp_ext(temp, delta, exponent));
tester.apply(llama_sampler_init_dist (0));
DUMP(&tester.cur_p);
llama_token_data_array cur_p = { cur.data(), cur.size(), -1, false };
APPLY(llama_sampler_init_softmax(), &cur_p);
DUMP(&cur_p);
APPLY(llama_sampler_init_top_p(p, 1), &cur_p);
DUMP(&cur_p);
GGML_ASSERT(cur_p.size == expected_probs.size());
for (size_t i = 0; i < cur_p.size; i++) {
GGML_ASSERT(fabs(cur_p.data[i].p - expected_probs[i]) < 1e-3);
}
tester.check();
}
static void test_tfs(const std::vector<float> & probs, const std::vector<float> & expected_probs, float z) {
const size_t n_vocab = probs.size();
static void test_top_k(const std::vector<float> & probs, const std::vector<float> & probs_expected, int k) {
sampler_tester tester(probs, probs_expected);
std::vector<llama_token_data> cur;
cur.reserve(n_vocab);
for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) {
const float logit = logf(probs[token_id]);
cur.emplace_back(llama_token_data{token_id, logit, 0.0f});
}
DUMP(&tester.cur_p);
tester.apply(llama_sampler_init_top_k(k));
tester.apply(llama_sampler_init_dist (0));
DUMP(&tester.cur_p);
llama_token_data_array cur_p = { cur.data(), cur.size(), -1, false };
DUMP(&cur_p);
APPLY(llama_sampler_init_tail_free(z, 1), &cur_p);
DUMP(&cur_p);
GGML_ASSERT(cur_p.size == expected_probs.size());
for (size_t i = 0; i < cur_p.size; i++) {
GGML_ASSERT(fabs(cur_p.data[i].p - expected_probs[i]) < 1e-3);
}
tester.check();
}
static void test_min_p(const std::vector<float> & probs, const std::vector<float> & expected_probs, float p) {
const size_t n_vocab = probs.size();
static void test_top_p(const std::vector<float> & probs, const std::vector<float> & probs_expected, float p) {
sampler_tester tester(probs, probs_expected);
std::vector<llama_token_data> cur;
cur.reserve(n_vocab);
for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) {
const float logit = logf(probs[token_id]);
cur.emplace_back(llama_token_data{token_id, logit, 0.0f});
}
DUMP(&tester.cur_p);
tester.apply(llama_sampler_init_top_p(p, 1));
tester.apply(llama_sampler_init_dist (0));
DUMP(&tester.cur_p);
llama_token_data_array cur_p = { cur.data(), cur.size(), -1, false };
DUMP(&cur_p);
APPLY(llama_sampler_init_min_p(p, 1), &cur_p);
DUMP(&cur_p);
APPLY(llama_sampler_init_softmax(), &cur_p);
GGML_ASSERT(cur_p.size == expected_probs.size());
for (size_t i = 0; i < cur_p.size; i++) {
GGML_ASSERT(fabs(cur_p.data[i].p - expected_probs[i]) < 1e-3);
}
tester.check();
}
static void test_typical(const std::vector<float> & probs, const std::vector<float> & expected_probs, float p) {
const size_t n_vocab = probs.size();
static void test_tfs(const std::vector<float> & probs, const std::vector<float> & probs_expected, float z) {
sampler_tester tester(probs, probs_expected);
std::vector<llama_token_data> cur;
cur.reserve(n_vocab);
for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) {
const float logit = logf(probs[token_id]);
cur.emplace_back(llama_token_data{token_id, logit, 0.0f});
}
DUMP(&tester.cur_p);
tester.apply(llama_sampler_init_tail_free(z, 1));
DUMP(&tester.cur_p);
llama_token_data_array cur_p = { cur.data(), cur.size(), -1, false };
DUMP(&cur_p);
APPLY(llama_sampler_init_typical(p, 1), &cur_p);
DUMP(&cur_p);
tester.check();
}
GGML_ASSERT(cur_p.size == expected_probs.size());
for (size_t i = 0; i < cur_p.size; i++) {
GGML_ASSERT(fabs(cur_p.data[i].p - expected_probs[i]) < 1e-3);
}
static void test_min_p(const std::vector<float> & probs, const std::vector<float> & probs_expected, float p) {
sampler_tester tester(probs, probs_expected);
DUMP(&tester.cur_p);
tester.apply(llama_sampler_init_min_p(p, 1));
tester.apply(llama_sampler_init_dist (0));
DUMP(&tester.cur_p);
tester.check();
}
static void test_xtc(const std::vector<float> & probs, const std::vector<float> & probs_expected, float p, float t) {
sampler_tester tester(probs, probs_expected);
DUMP(&tester.cur_p);
tester.apply(llama_sampler_init_xtc(p, t, 0, 0));
DUMP(&tester.cur_p);
tester.check();
}
static void test_typical(const std::vector<float> & probs, const std::vector<float> & probs_expected, float p) {
sampler_tester tester(probs, probs_expected);
DUMP(&tester.cur_p);
tester.apply(llama_sampler_init_typical(p, 1));
DUMP(&tester.cur_p);
tester.check();
}
static void test_penalties(
const std::vector<float> & probs, const std::vector<llama_token> & last_tokens,
const std::vector<float> & expected_probs, float repeat_penalty, float alpha_frequency, float alpha_presence
const std::vector<float> & probs_expected, float repeat_penalty, float alpha_frequency, float alpha_presence
) {
GGML_ASSERT(probs.size() == expected_probs.size());
GGML_ASSERT(probs.size() == probs_expected.size());
sampler_tester tester(probs, probs_expected);
const size_t n_vocab = probs.size();
std::vector<llama_token_data> cur;
cur.reserve(n_vocab);
for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) {
const float logit = logf(probs[token_id]);
cur.emplace_back(llama_token_data{token_id, logit, 0.0f});
}
llama_token_data_array cur_p = { cur.data(), cur.size(), -1, false };
auto * sampler = llama_sampler_init_penalties(n_vocab, LLAMA_TOKEN_NULL, LLAMA_TOKEN_NULL, last_tokens.size(), repeat_penalty, alpha_frequency, alpha_presence, false, false);
for (size_t i = 0; i < last_tokens.size(); i++) {
llama_sampler_accept(sampler, last_tokens[i]);
}
APPLY(llama_sampler_init_softmax(), &cur_p);
DUMP(&cur_p);
APPLY(sampler, &cur_p);
APPLY(llama_sampler_init_softmax(), &cur_p);
DUMP(&cur_p);
DUMP(&tester.cur_p);
tester.apply(sampler);
tester.apply(llama_sampler_init_dist(0));
DUMP(&tester.cur_p);
GGML_ASSERT(cur_p.size == expected_probs.size());
for (size_t i = 0; i < cur_p.size; i++) {
GGML_ASSERT(fabs(cur_p.data[i].p - expected_probs[i]) < 1e-3);
tester.check();
}
static void test_dry(
const std::vector<float> & probs, const std::vector<llama_token> & last_tokens,
const std::vector<float> & expected_probs, float dry_multiplier, float dry_base,
int dry_allowed_length, int dry_penalty_last_n,
const std::vector<std::vector<llama_token>> & seq_breakers
) {
GGML_ASSERT(probs.size() == expected_probs.size());
sampler_tester tester(probs, expected_probs);
auto * sampler = llama_sampler_init_dry_testing(1024, dry_multiplier, dry_base, dry_allowed_length, dry_penalty_last_n, seq_breakers);
for (size_t i = 0; i < last_tokens.size(); i++) {
llama_sampler_accept(sampler, last_tokens[i]);
}
DUMP(&tester.cur_p);
tester.apply(sampler);
tester.apply(llama_sampler_init_dist(0));
DUMP(&tester.cur_p);
tester.check();
}
static void test_sampler_queue(const size_t n_vocab, const std::string & samplers_sequence, const int top_k, const float top_p, const float min_p
) {
std::vector<llama_token_data> cur;
cur.reserve(n_vocab);
for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) {
const float logit = logf(token_id);
cur.emplace_back(llama_token_data{token_id, logit, 0.0f});
}
llama_token_data_array cur_p = { cur.data(), cur.size(), -1, false };
sampler_tester tester(n_vocab);
llama_token min_token_id = 0;
const llama_token max_token_id = n_vocab-1;
for (auto s : samplers_sequence) {
switch (s){
case 'k': APPLY(llama_sampler_init_top_k(top_k), &cur_p); break;
case 'k': tester.apply(llama_sampler_init_top_k(top_k)); break;
case 'f': GGML_ABORT("tail_free test not implemented");
case 'y': GGML_ABORT("typical test not implemented");
case 'p': APPLY(llama_sampler_init_top_p(top_p, 1), &cur_p); break;
case 'm': APPLY(llama_sampler_init_min_p(min_p, 1), &cur_p); break;
case 'p': tester.apply(llama_sampler_init_top_p(top_p, 1)); break;
case 'm': tester.apply(llama_sampler_init_min_p(min_p, 1)); break;
case 't': GGML_ABORT("temperature test not implemented");
default : GGML_ABORT("Unknown sampler");
}
APPLY(llama_sampler_init_softmax(), &cur_p); // make sure tokens are sorted for tests
tester.apply(llama_sampler_init_dist(0));
auto & cur_p = tester.cur_p;
const int size = cur_p.size;
@ -263,7 +283,7 @@ static void bench(llama_sampler * cnstr, const char * cnstr_name, const std::vec
}
const int64_t t_end = ggml_time_us();
llama_sampler_free(cnstr);
printf("%-42s: %8.3f us/iter\n", cnstr_name, (t_end - t_start) / (float)n_iter);
printf("%-43s: %8.3f us/iter\n", cnstr_name, (t_end - t_start) / (float)n_iter);
}
#define BENCH(__cnstr, __data, __n_iter) bench((__cnstr), #__cnstr, (__data), (__n_iter))
@ -279,26 +299,32 @@ static void test_perf() {
data.emplace_back(llama_token_data{i, logit, 0.0f});
}
BENCH(llama_sampler_init_top_k (40), data, 32);
BENCH(llama_sampler_init_top_p (0.8f, 1), data, 32);
BENCH(llama_sampler_init_min_p (0.2f, 1), data, 32);
BENCH(llama_sampler_init_tail_free(0.5f, 1), data, 32);
BENCH(llama_sampler_init_typical (0.5f, 1), data, 32);
BENCH(llama_sampler_init_softmax (), data, 32);
BENCH(llama_sampler_init_top_k (40), data, 32);
BENCH(llama_sampler_init_top_p (0.8f, 1), data, 32);
BENCH(llama_sampler_init_min_p (0.2f, 1), data, 32);
BENCH(llama_sampler_init_tail_free(0.5f, 1), data, 32);
BENCH(llama_sampler_init_typical (0.5f, 1), data, 32);
BENCH(llama_sampler_init_xtc (1.0f, 0.1f, 1, 1), data, 32);
}
int main(void) {
ggml_time_init();
test_top_k({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f}, 1);
test_top_k({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f, 0.2f}, 3);
test_temp({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f, 0.2f, 0.1f}, 1.0f);
test_temp({0.1f, 0.2f, 0.3f, 0.4f}, {1.0f, 0.0f, 0.0f, 0.0f}, 0.0f);
test_temp_ext({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f, 0.2f, 0.1f}, 1.0f, 0.0f, 1.0f);
test_temp_ext({0.1f, 0.2f, 0.3f, 0.4f}, {1.0f, 0.0f, 0.0f, 0.0f}, 0.0f, 0.0f, 1.0f);
test_top_k({0.1f, 0.2f, 0.3f, 0.4f}, {1.0f}, 1);
test_top_k({0.1f, 0.2f, 0.3f, 0.4f}, {0.44444f, 0.33333f, 0.22222f}, 3);
test_top_k({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f, 0.2f, 0.1f}, 4);
test_top_k({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f, 0.2f, 0.1f}, 0);
test_top_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f}, 0);
test_top_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f}, 0.7f);
test_top_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f, 0.2f}, 0.8f);
test_top_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f, 0.2f, 0.1f}, 1);
test_top_p({0.1f, 0.2f, 0.3f, 0.4f}, {1.0f}, 0);
test_top_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.571429f, 0.428571f}, 0.7f);
test_top_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.44444f, 0.33333f, 0.22222f}, 0.8f);
test_top_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f, 0.2f, 0.1f}, 1.0f);
test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f/1.0f, 0.3f/1.0f, 0.2f/1.0f, 0.1f/1.0f}, 0.00f);
test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f/1.0f, 0.3f/1.0f, 0.2f/1.0f, 0.1f/1.0f}, 0.24f);
@ -309,6 +335,14 @@ int main(void) {
test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f/0.4f}, 0.76f);
test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f/0.4f}, 1.00f);
printf("XTC should:\n");
test_xtc({0.4f, 0.3f, 0.2f, 0.1f}, {0.1f}, 0.99f, 0.09f);
test_xtc({0.4f, 0.3f, 0.2f, 0.1f}, {0.2f, 0.1f}, 0.99f, 0.19f);
test_xtc({0.4f, 0.3f, 0.2f, 0.1f}, {0.3f, 0.2f, 0.1f}, 0.99f, 0.29f);
printf("XTC should not:\n");
test_xtc({0.4f, 0.3f, 0.2f, 0.1f}, {0.4f, 0.3f, 0.2f, 0.1f}, 0.99f, 0.39f);
test_tfs({0.1f, 0.15f, 0.2f, 0.25f, 0.3f}, {0.3f}, 0.25f);
test_tfs({0.1f, 0.15f, 0.2f, 0.25f, 0.3f}, {0.3f, 0.25f}, 0.75f);
test_tfs({0.1f, 0.15f, 0.2f, 0.25f, 0.3f}, {0.3f, 0.25f}, 0.99f);
@ -324,6 +358,13 @@ int main(void) {
test_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2}, {0.499966f, 0.499966f, 0.000023f, 0.000023f, 0.000023f}, 1.0f, 5.0f, 5.0f);
test_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2, 0, 0}, {0.499977f, 0.499977f, 0.000023f, 0.000023f, 0.000000f}, 1.0f, 5.0f, 5.0f);
test_dry({0.25f, 0.25f, 0.25f, 0.25f}, {0, 1}, {0.25f, 0.25f, 0.25f, 0.25f}, 1.0f, 1.1f, 2, 4, {});
test_dry({0.25f, 0.25f, 0.25f, 0.25f}, {0, 1, 2, 0, 1}, {0.296923f, 0.296923f, 0.296923f, 0.109232f}, 1.0f, 1.1f, 2, 5, {});
test_dry({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 3, 4, 0, 1}, {0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, 1.0f, 1.1f, 2, 6, {{3}});
test_dry({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2, 0, 1}, {0.241818f, 0.241818f, 0.241818f, 0.241818f, 0.032727f}, 2.0f, 1.1f, 2, 5, {});
test_dry({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2, 3, 4, 0, 1}, {0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, 1.0f, 1.1f, 4, 7, {});
test_sampler_queue(10000, "k", 10000, 1.0f, 1.0f);
test_sampler_queue(10000, "k", 1, 1.0f, 1.0f);
test_sampler_queue(10000, "p", 10000, 1.0f, 1.0f);