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

move arg definitions out of gemma.h to app.h

This commit is contained in:
austinvhuang 2024-03-10 23:49:25 -04:00
parent 5d323c00fe
commit 60d054e041
4 changed files with 123 additions and 119 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
examples/hello_world/CMakeLists.txt
View File

@ -35,7 +35,7 @@ if (BUILD_MODE STREQUAL "local")
else() else()
FetchContent_Declare(gemma GIT_REPOSITORY https://github.com/google/gemma.cpp.git GIT_TAG 8c7b2cf61b9794b806de091685dc6739dd3db837) FetchContent_Declare(gemma GIT_REPOSITORY https://github.com/google/gemma.cpp.git GIT_TAG 8c7b2cf61b9794b806de091685dc6739dd3db837)
endif() endif()
FetchContent_MakeAvailabl(gemma) FetchContent_MakeAvailable(gemma)
if(NOT CMAKE_BUILD_TYPE) if(NOT CMAKE_BUILD_TYPE)
set(CMAKE_BUILD_TYPE "Release") set(CMAKE_BUILD_TYPE "Release")

3
examples/hello_world/run.cc
View File

@ -21,6 +21,9 @@
// copybara:import_next_line:gemma_cpp // copybara:import_next_line:gemma_cpp
#include "util/args.h" #include "util/args.h"
// copybara:end // copybara:end
// copybara:import_next_line:gemma_cpp
#include "util/app.h" // LoaderArgs
// copybara:end
#include "hwy/contrib/thread_pool/thread_pool.h" #include "hwy/contrib/thread_pool/thread_pool.h"
std::vector<int> tokenize( std::vector<int> tokenize(

118
gemma.h
View File

@ -71,124 +71,6 @@ struct RuntimeConfig {
int verbosity; int verbosity;
}; };
struct LoaderArgs : public ArgsBase<LoaderArgs> {
LoaderArgs(int argc, char* argv[]) { InitAndParse(argc, argv); }
static std::string ToLower(const std::string& text) {
std::string result = text;
std::transform(begin(result), end(result), begin(result),
[](unsigned char c) { return std::tolower(c); });
return result;
}
gcpp::Model ModelType() const {
const std::string model_type_lc = ToLower(model_type);
if (model_type_lc == "2b-pt" || model_type_lc == "2b-it") {
return gcpp::Model::GEMMA_2B;
} else {
return gcpp::Model::GEMMA_7B;
}
}
gcpp::ModelTraining ModelTraining() const {
const std::string model_type_lc = ToLower(model_type);
if (model_type_lc == "7b-pt" || model_type_lc == "2b-pt") {
return gcpp::ModelTraining::GEMMA_PT;
} else {
return gcpp::ModelTraining::GEMMA_IT;
}
}
// Returns error string or nullptr if OK.
const char* Validate() const {
const std::string model_type_lc = ToLower(model_type);
if (model_type.empty()) {
return "Missing --model flag, need to specify either 2b-pt, 7b-pt, "
"2b-it, or 7b-it.";
}
if (model_type_lc != "2b-pt" && model_type_lc != "7b-pt" &&
model_type_lc != "2b-it" && model_type_lc != "7b-it") {
return "Model type must be 2b-pt, 7b-pt, 2b-it, or "
"7b-it.";
}
if (tokenizer.path.empty()) {
return "Missing --tokenizer flag, a file for the tokenizer is required.";
}
if (compressed_weights.path.empty()) {
return "Missing --compressed_weights flag, a file for the compressed "
"model.";
}
return nullptr;
}
Path tokenizer;
Path weights; // uncompressed weights file location
Path compressed_weights; // compressed weights file location
std::string model_type;
template <class Visitor>
void ForEach(const Visitor& visitor) {
visitor(tokenizer, "tokenizer", Path(),
"Path name of tokenizer model file.\n Required argument.");
visitor(
compressed_weights, "compressed_weights", Path(),
"Path name of compressed weights file, regenerated from `--weights` "
"file if "
"the compressed weights file does not exist.\n Required argument.");
visitor(model_type, "model", std::string(),
"Model type\n 2b-it = 2B parameters, instruction-tuned\n "
"2b-pt = 2B parameters, pretrained\n 7b-it = 7B parameters "
"instruction-tuned\n 7b-pt = 7B parameters, pretrained\n"
" Required argument.");
visitor(weights, "weights", Path(),
"Path name of model weights (.sbs) file. Only required if "
"compressed_weights file is not present and needs to be "
"regenerated. This parameter is only required for compressing"
"new model weight exports, otherwise it is not needed.");
}
};
struct InferenceArgs : public ArgsBase<InferenceArgs> {
InferenceArgs(int argc, char* argv[]) { InitAndParse(argc, argv); }
size_t max_tokens;
size_t max_generated_tokens;
float temperature;
bool deterministic;
bool multiturn;
// Returns error string or nullptr if OK.
const char* Validate() const {
if (max_tokens > gcpp::kSeqLen) {
return "max_tokens is larger than the maximum sequence length (see "
"configs.h).";
}
if (max_generated_tokens > max_tokens) {
return "Maximum number of generated tokens is larger than the maximum "
"total tokens.";
}
return nullptr;
}
template <class Visitor>
void ForEach(const Visitor& visitor) {
visitor(max_tokens, "max_tokens", size_t{3072},
"Maximum number of tokens in prompt + generation.");
visitor(max_generated_tokens, "max_generated_tokens", size_t{2048},
"Maximum number of tokens to generate.");
visitor(temperature, "temperature", 1.0f, "Temperature for top-K", 2);
visitor(deterministic, "deterministic", false,
"Make top-k sampling deterministic", 2);
visitor(multiturn, "multiturn", false,
"Multiturn mode\n 0 = clear KV cache after every "
"interaction\n 1 = continue KV cache after every interaction\n "
" Default : 0 (conversation "
"resets every turn)");
}
};
struct GemmaInterface; struct GemmaInterface;
struct Gemma { struct Gemma {

119
util/app.h
View File

@ -98,6 +98,125 @@ class AppArgs : public ArgsBase<AppArgs> {
} }
}; };
struct LoaderArgs : public ArgsBase<LoaderArgs> {
LoaderArgs(int argc, char* argv[]) { InitAndParse(argc, argv); }
static std::string ToLower(const std::string& text) {
std::string result = text;
std::transform(begin(result), end(result), begin(result),
[](unsigned char c) { return std::tolower(c); });
return result;
}
gcpp::Model ModelType() const {
const std::string model_type_lc = ToLower(model_type);
if (model_type_lc == "2b-pt" || model_type_lc == "2b-it") {
return gcpp::Model::GEMMA_2B;
} else {
return gcpp::Model::GEMMA_7B;
}
}
gcpp::ModelTraining ModelTraining() const {
const std::string model_type_lc = ToLower(model_type);
if (model_type_lc == "7b-pt" || model_type_lc == "2b-pt") {
return gcpp::ModelTraining::GEMMA_PT;
} else {
return gcpp::ModelTraining::GEMMA_IT;
}
}
// Returns error string or nullptr if OK.
const char* Validate() const {
const std::string model_type_lc = ToLower(model_type);
if (model_type.empty()) {
return "Missing --model flag, need to specify either 2b-pt, 7b-pt, "
"2b-it, or 7b-it.";
}
if (model_type_lc != "2b-pt" && model_type_lc != "7b-pt" &&
model_type_lc != "2b-it" && model_type_lc != "7b-it") {
return "Model type must be 2b-pt, 7b-pt, 2b-it, or "
"7b-it.";
}
if (tokenizer.path.empty()) {
return "Missing --tokenizer flag, a file for the tokenizer is required.";
}
if (compressed_weights.path.empty()) {
return "Missing --compressed_weights flag, a file for the compressed "
"model.";
}
return nullptr;
}
Path tokenizer;
Path weights; // uncompressed weights file location
Path compressed_weights; // compressed weights file location
std::string model_type;
template <class Visitor>
void ForEach(const Visitor& visitor) {
visitor(tokenizer, "tokenizer", Path(),
"Path name of tokenizer model file.\n Required argument.");
visitor(
compressed_weights, "compressed_weights", Path(),
"Path name of compressed weights file, regenerated from `--weights` "
"file if "
"the compressed weights file does not exist.\n Required argument.");
visitor(model_type, "model", std::string(),
"Model type\n 2b-it = 2B parameters, instruction-tuned\n "
"2b-pt = 2B parameters, pretrained\n 7b-it = 7B parameters "
"instruction-tuned\n 7b-pt = 7B parameters, pretrained\n"
" Required argument.");
visitor(weights, "weights", Path(),
"Path name of model weights (.sbs) file. Only required if "
"compressed_weights file is not present and needs to be "
"regenerated. This parameter is only required for compressing"
"new model weight exports, otherwise it is not needed.");
}
};
struct InferenceArgs : public ArgsBase<InferenceArgs> {
InferenceArgs(int argc, char* argv[]) { InitAndParse(argc, argv); }
size_t max_tokens;
size_t max_generated_tokens;
float temperature;
bool deterministic;
bool multiturn;
// Returns error string or nullptr if OK.
const char* Validate() const {
if (max_tokens > gcpp::kSeqLen) {
return "max_tokens is larger than the maximum sequence length (see "
"configs.h).";
}
if (max_generated_tokens > max_tokens) {
return "Maximum number of generated tokens is larger than the maximum "
"total tokens.";
}
return nullptr;
}
template <class Visitor>
void ForEach(const Visitor& visitor) {
visitor(max_tokens, "max_tokens", size_t{3072},
"Maximum number of tokens in prompt + generation.");
visitor(max_generated_tokens, "max_generated_tokens", size_t{2048},
"Maximum number of tokens to generate.");
visitor(temperature, "temperature", 1.0f, "Temperature for top-K", 2);
visitor(deterministic, "deterministic", false,
"Make top-k sampling deterministic", 2);
visitor(multiturn, "multiturn", false,
"Multiturn mode\n 0 = clear KV cache after every "
"interaction\n 1 = continue KV cache after every interaction\n "
" Default : 0 (conversation "
"resets every turn)");
}
};
} // namespace gcpp } // namespace gcpp
#endif // THIRD_PARTY_GEMMA_CPP_UTIL_APP_H_ #endif // THIRD_PARTY_GEMMA_CPP_UTIL_APP_H_