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

No public description 

PiperOrigin-RevId: 617315030
This commit is contained in:
Eric Ye 2024-03-19 23:35:58 +01:00 committed by Jan Wassenberg
parent 5e0cafbdc2
commit fdc3812446
3 changed files with 16 additions and 10 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
gemma.cc
View File

@ -25,8 +25,6 @@
#include "compression/compress-inl.h" #include "compression/compress-inl.h"
// copybara:import_next_line:gemma_cpp // copybara:import_next_line:gemma_cpp
#include "ops.h" #include "ops.h"
// copybara:import_next_line:gemma_cpp
#include "util/args.h" // Path
#include "hwy/contrib/matvec/matvec-inl.h" #include "hwy/contrib/matvec/matvec-inl.h"
#include "hwy/highway.h" #include "hwy/highway.h"
#include "hwy/profiler.h" #include "hwy/profiler.h"
@ -815,9 +813,8 @@ void GemmaImpl<ConfigGemma7B>::Generate(
} }
Gemma::Gemma(const Path& tokenizer_path, const Path& compressed_weights_path, Gemma::Gemma(const Path& tokenizer_path, const Path& compressed_weights_path,
const Path& weights_path, Model model_type, ModelTraining training, const Path& weights_path, Model model_type,
hwy::ThreadPool& pool) hwy::ThreadPool& pool) {
: model_training(training) {
std::unique_ptr<sentencepiece::SentencePieceProcessor> tokenizer; std::unique_ptr<sentencepiece::SentencePieceProcessor> tokenizer;
{ {
PROFILER_ZONE("Startup.tokenizer"); PROFILER_ZONE("Startup.tokenizer");
@ -842,6 +839,11 @@ Gemma::Gemma(const Path& tokenizer_path, const Path& compressed_weights_path,
} }
} }
Gemma::Gemma(const Path& tokenizer_path, const Path& compressed_weights_path,
Model model_type, hwy::ThreadPool& pool)
: Gemma(tokenizer_path, compressed_weights_path, Path{""}, model_type,
pool) {}
Gemma::~Gemma() = default; // after GemmaInterface is defined Gemma::~Gemma() = default; // after GemmaInterface is defined
const sentencepiece::SentencePieceProcessor* Gemma::Tokenizer() const { const sentencepiece::SentencePieceProcessor* Gemma::Tokenizer() const {

10
gemma.h
View File

@ -16,9 +16,12 @@
#ifndef THIRD_PARTY_GEMMA_CPP_GEMMA_H_ #ifndef THIRD_PARTY_GEMMA_CPP_GEMMA_H_
#define THIRD_PARTY_GEMMA_CPP_GEMMA_H_ #define THIRD_PARTY_GEMMA_CPP_GEMMA_H_
#include <algorithm>
#include <cctype>
#include <functional> #include <functional>
#include <memory> #include <memory>
#include <random> #include <random>
#include <string>
#include <vector> #include <vector>
// copybara:import_next_line:gemma_cpp // copybara:import_next_line:gemma_cpp
@ -28,7 +31,7 @@
#include "configs.h" // kSeqLen #include "configs.h" // kSeqLen
// copybara:end // copybara:end
// copybara:import_next_line:gemma_cpp // copybara:import_next_line:gemma_cpp
#include "util/args.h" // Path #include "util/args.h" // ArgsBase
// copybara:end // copybara:end
#include "hwy/aligned_allocator.h" #include "hwy/aligned_allocator.h"
#include "hwy/base.h" // hwy::bfloat16_t #include "hwy/base.h" // hwy::bfloat16_t
@ -72,8 +75,9 @@ struct GemmaInterface;
struct Gemma { struct Gemma {
Gemma(const Path& tokenizer_path, const Path& compressed_weights_path, Gemma(const Path& tokenizer_path, const Path& compressed_weights_path,
const Path& weights_path, Model model_type, ModelTraining training, const Path& weights_path, Model model_type, hwy::ThreadPool& pool);
hwy::ThreadPool& pool); Gemma(const Path& tokenizer_path, const Path& compressed_weights_path,
Model model_type, hwy::ThreadPool& pool);
~Gemma(); // must be defined after GemmaInterface's dtor is defined. ~Gemma(); // must be defined after GemmaInterface's dtor is defined.
const sentencepiece::SentencePieceProcessor* Tokenizer() const; const sentencepiece::SentencePieceProcessor* Tokenizer() const;
std::unique_ptr<GemmaInterface> impl_; std::unique_ptr<GemmaInterface> impl_;

4
run.cc
View File

@ -234,8 +234,8 @@ void Run(LoaderArgs& loader, InferenceArgs& inference, AppArgs& app) {
[](uint64_t /*task*/, size_t thread) { PinThreadToCore(thread); }); [](uint64_t /*task*/, size_t thread) { PinThreadToCore(thread); });
} }
gcpp::Gemma model(loader.tokenizer, loader.compressed_weights, loader.weights, gcpp::Gemma model(loader.tokenizer, loader.compressed_weights,
loader.ModelType(), loader.ModelTraining(), pool); loader.ModelType(), pool);
auto kv_cache = CreateKVCache(loader.ModelType()); auto kv_cache = CreateKVCache(loader.ModelType());