root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-25 02:44:36 +00:00
Code Issues Actions 9 Packages Projects Releases Wiki Activity

examples : Fix llama-export-lora example (#8607)

Browse Source 
* fix export-lora example

* add more logging

* reject merging subset

* better check

* typo
This commit is contained in:
Xuan Son Nguyen 2024-07-23 23:48:37 +02:00 committed by GitHub
parent b841d07408
commit de280085e7
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
5 changed files with 378 additions and 436 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
Makefile
View File

@ -1322,7 +1322,7 @@ llama-finetune: examples/finetune/finetune.cpp \
$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
llama-export-lora: examples/export-lora/export-lora.cpp \ llama-export-lora: examples/export-lora/export-lora.cpp \
$(OBJ_GGML) common/log.h $(OBJ_ALL)
$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)

19
common/common.cpp
View File

@ -694,11 +694,6 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
params.lora_adapter.emplace_back(lora_adapter, std::stof(argv[i])); params.lora_adapter.emplace_back(lora_adapter, std::stof(argv[i]));
return true; return true;
} }
if (arg == "--lora-base") {
CHECK_ARG
params.lora_base = argv[i];
return true;
}
if (arg == "--control-vector") { if (arg == "--control-vector") {
CHECK_ARG CHECK_ARG
params.control_vectors.push_back({ 1.0f, argv[i], }); params.control_vectors.push_back({ 1.0f, argv[i], });
@ -1274,6 +1269,7 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
CHECK_ARG CHECK_ARG
params.out_file = argv[i]; params.out_file = argv[i];
params.cvector_outfile = argv[i]; params.cvector_outfile = argv[i];
params.lora_outfile = argv[i];
return true; return true;
} }
if (arg == "-ofreq" || arg == "--output-frequency") { if (arg == "-ofreq" || arg == "--output-frequency") {
@ -1583,9 +1579,8 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param
options.push_back({ "*", " --override-kv KEY=TYPE:VALUE", options.push_back({ "*", " --override-kv KEY=TYPE:VALUE",
"advanced option to override model metadata by key. may be specified multiple times.\n" "advanced option to override model metadata by key. may be specified multiple times.\n"
"types: int, float, bool, str. example: --override-kv tokenizer.ggml.add_bos_token=bool:false" }); "types: int, float, bool, str. example: --override-kv tokenizer.ggml.add_bos_token=bool:false" });
options.push_back({ "*", " --lora FNAME", "apply LoRA adapter (implies --no-mmap)" }); options.push_back({ "*", " --lora FNAME", "apply LoRA adapter (can be repeated to use multiple adapters)" });
options.push_back({ "*", " --lora-scaled FNAME S", "apply LoRA adapter with user defined scaling S (implies --no-mmap)" }); options.push_back({ "*", " --lora-scaled FNAME S", "apply LoRA adapter with user defined scaling S (can be repeated to use multiple adapters)" });
options.push_back({ "*", " --lora-base FNAME", "optional model to use as a base for the layers modified by the LoRA adapter" });
options.push_back({ "*", " --control-vector FNAME", "add a control vector\n" options.push_back({ "*", " --control-vector FNAME", "add a control vector\n"
"note: this argument can be repeated to add multiple control vectors" }); "note: this argument can be repeated to add multiple control vectors" });
options.push_back({ "*", " --control-vector-scaled FNAME SCALE", options.push_back({ "*", " --control-vector-scaled FNAME SCALE",
@ -1676,6 +1671,13 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param
options.push_back({ "cvector", " --pca-iter N", "number of iterations used for PCA (default: %d)", params.n_pca_iterations }); options.push_back({ "cvector", " --pca-iter N", "number of iterations used for PCA (default: %d)", params.n_pca_iterations });
options.push_back({ "cvector", " --method {pca,mean}", "dimensionality reduction method to be used (default: pca)" }); options.push_back({ "cvector", " --method {pca,mean}", "dimensionality reduction method to be used (default: pca)" });
options.push_back({ "export-lora" });
options.push_back({ "export-lora", "-m, --model", "model path from which to load base model (default '%s')", params.model.c_str() });
options.push_back({ "export-lora", " --lora FNAME", "path to LoRA adapter (can be repeated to use multiple adapters)" });
options.push_back({ "export-lora", " --lora-scaled FNAME S", "path to LoRA adapter with user defined scaling S (can be repeated to use multiple adapters)" });
options.push_back({ "*", "-t, --threads N", "number of threads to use during computation (default: %d)", params.n_threads });
options.push_back({ "export-lora", "-o, --output FNAME", "output file (default: '%s')", params.lora_outfile.c_str() });
printf("usage: %s [options]\n", argv[0]); printf("usage: %s [options]\n", argv[0]);
for (const auto & o : options) { for (const auto & o : options) {
@ -3166,7 +3168,6 @@ void yaml_dump_non_result_info(FILE * stream, const gpt_params & params, const l
} }
fprintf(stream, " - %s: %f\n", std::get<0>(la).c_str(), std::get<1>(la)); fprintf(stream, " - %s: %f\n", std::get<0>(la).c_str(), std::get<1>(la));
} }
fprintf(stream, "lora_base: %s\n", params.lora_base.c_str());
fprintf(stream, "main_gpu: %d # default: 0\n", params.main_gpu); fprintf(stream, "main_gpu: %d # default: 0\n", params.main_gpu);
fprintf(stream, "min_keep: %d # default: 0 (disabled)\n", sparams.min_keep); fprintf(stream, "min_keep: %d # default: 0 (disabled)\n", sparams.min_keep);
fprintf(stream, "mirostat: %d # default: 0 (disabled)\n", sparams.mirostat); fprintf(stream, "mirostat: %d # default: 0 (disabled)\n", sparams.mirostat);

3
common/common.h
View File

@ -128,7 +128,6 @@ struct gpt_params {
// TODO: avoid tuple, use struct // TODO: avoid tuple, use struct
std::vector<std::tuple<std::string, float>> lora_adapter; // lora adapter path with user defined scale std::vector<std::tuple<std::string, float>> lora_adapter; // lora adapter path with user defined scale
std::string lora_base = ""; // base model path for the lora adapter
std::vector<llama_control_vector_load_info> control_vectors; // control vector with user defined scale std::vector<llama_control_vector_load_info> control_vectors; // control vector with user defined scale
@ -255,6 +254,8 @@ struct gpt_params {
std::string cvector_negative_file = "examples/cvector-generator/negative.txt"; std::string cvector_negative_file = "examples/cvector-generator/negative.txt";
bool spm_infill = false; // suffix/prefix/middle pattern for infill bool spm_infill = false; // suffix/prefix/middle pattern for infill
std::string lora_outfile = "ggml-lora-merged-f16.gguf";
}; };
void gpt_params_handle_hf_token(gpt_params & params); void gpt_params_handle_hf_token(gpt_params & params);

15
examples/export-lora/README.md
View File

@ -6,12 +6,11 @@ Apply LORA adapters to base model and export the resulting model.
usage: llama-export-lora [options] usage: llama-export-lora [options]
options: options:
-h, --help show this help message and exit -m, --model model path from which to load base model (default '')
-m FNAME, --model-base FNAME model path from which to load base model (default '') --lora FNAME path to LoRA adapter (can be repeated to use multiple adapters)
-o FNAME, --model-out FNAME path to save exported model (default '') --lora-scaled FNAME S path to LoRA adapter with user defined scaling S (can be repeated to use multiple adapters)
-l FNAME, --lora FNAME apply LoRA adapter -t, --threads N number of threads to use during computation (default: 4)
-s FNAME S, --lora-scaled FNAME S apply LoRA adapter with user defined scaling S -o, --output FNAME output file (default: 'ggml-lora-merged-f16.gguf')
-t N, --threads N number of threads to use during computation (default: 4)
``` ```
For example: For example:
@ -20,7 +19,7 @@ For example:
./bin/llama-export-lora \ ./bin/llama-export-lora \
-m open-llama-3b-v2-q8_0.gguf \ -m open-llama-3b-v2-q8_0.gguf \
-o open-llama-3b-v2-q8_0-english2tokipona-chat.gguf \ -o open-llama-3b-v2-q8_0-english2tokipona-chat.gguf \
-l lora-open-llama-3b-v2-q8_0-english2tokipona-chat-LATEST.bin --lora lora-open-llama-3b-v2-q8_0-english2tokipona-chat-LATEST.bin
``` ```
Multiple LORA adapters can be applied by passing multiple `-l FN` or `-s FN S` command line parameters. Multiple LORA adapters can be applied by passing multiple `--lora FNAME` or `--lora-scaled FNAME S` command line parameters.

753
examples/export-lora/export-lora.cpp
View File

@ -1,465 +1,406 @@
#include "common.h" #include "common.h"
#include "ggml.h" #include "ggml.h"
#include "ggml-alloc.h" #include "ggml-alloc.h"
#include <map>
#include <vector> #include <vector>
#include <string> #include <string>
#include <thread> #include <thread>
#include <fstream>
struct lora_info { static bool g_verbose = false;
std::string filename;
static std::string get_kv_str(struct gguf_context * ctx_gguf, const std::string & key){
int id = gguf_find_key(ctx_gguf, key.c_str());
return id < 0 ? "" : std::string(gguf_get_val_str(ctx_gguf, id));
}
static float get_kv_f32(struct gguf_context * ctx_gguf, const std::string & key) {
int id = gguf_find_key(ctx_gguf, key.c_str());
return id < 0 ? 0.0f : gguf_get_val_f32(ctx_gguf, id);
}
static void zeros(std::ofstream & file, size_t n) {
char zero = 0;
for (size_t i = 0; i < n; ++i) {
file.write(&zero, 1);
}
}
static std::string ggml_ne_string(const ggml_tensor * t) {
std::string str;
for (int i = 0; i < GGML_MAX_DIMS; ++i) {
str += std::to_string(t->ne[i]);
if (i + 1 < GGML_MAX_DIMS) {
str += ", ";
}
}
return str;
}
static struct gguf_context * load_gguf(std::string & fname, struct ggml_context ** ctx_ggml) {
struct gguf_init_params params = {
/*.no_alloc = */ true,
/*.ctx = */ ctx_ggml,
};
struct gguf_context * ctx_gguf = gguf_init_from_file(fname.c_str(), params);
if (!ctx_gguf) {
throw std::runtime_error("failed to load input GGUF from " + fname);
}
return ctx_gguf;
}
static void replace_all(std::string & s, const std::string & search, const std::string & replace) {
std::string result;
for (size_t pos = 0; ; pos += search.length()) {
auto new_pos = s.find(search, pos);
if (new_pos == std::string::npos) {
result += s.substr(pos, s.size() - pos);
break;
}
result += s.substr(pos, new_pos - pos) + replace;
pos = new_pos;
}
s = std::move(result);
}
struct file_input {
struct ggml_context * ctx_meta = nullptr;
struct gguf_context * ctx_gguf = nullptr;
std::ifstream f_in;
std::map<std::string, ggml_tensor *> tensors;
float alpha;
float scale; float scale;
file_input(std::string & fname, float scale): f_in(fname, std::ios::binary), scale(scale) {
if (!f_in.is_open()) {
throw std::runtime_error("failed to open input gguf from " + fname);
}
ctx_gguf = load_gguf(fname, &ctx_meta);
alpha = get_kv_f32(ctx_gguf, "adapter.lora.alpha");
printf("%s: loaded gguf from %s\n", __func__, fname.c_str());
for (ggml_tensor * cur = ggml_get_first_tensor(ctx_meta); cur; cur = ggml_get_next_tensor(ctx_meta, cur)) {
std::string name(cur->name);
tensors[name] = cur;
if (g_verbose) {
printf("%s: %s\n", __func__, cur->name);
}
}
}
ggml_tensor * get_tensor(std::string name) {
if (tensors.find(name) == tensors.end()) {
return nullptr;
}
return tensors[name];
}
void read_tensor_data(std::string name, std::vector<uint8_t> & buf) {
if (tensors.find(name) == tensors.end()) {
throw std::runtime_error("cannot find tensor with name: " + name);
}
auto len = ggml_nbytes(tensors[name]);
if (buf.size() < len) {
buf.resize(len);
}
auto i_tensor_in = gguf_find_tensor(ctx_gguf, name.c_str()); // idx of tensor in the input file
auto offset = gguf_get_data_offset(ctx_gguf) + gguf_get_tensor_offset(ctx_gguf, i_tensor_in);
f_in.seekg(offset);
f_in.read((char* )buf.data(), len);
}
~file_input() {
gguf_free(ctx_gguf);
ggml_free(ctx_meta);
}
}; };
struct export_lora_params { struct lora_merge_ctx {
std::string fn_model_base; // input base model + adapters
std::string fn_model_out; file_input base_model;
std::vector<struct lora_info> lora; std::vector<std::unique_ptr<file_input>> adapters;
// for computing merged tensor
int n_threads; int n_threads;
}; ggml_backend_t backend = nullptr;
ggml_gallocr_t allocr = nullptr;
std::vector<uint8_t> read_buf;
struct lora_data { // output file
struct lora_info info; struct gguf_context * ctx_out;
std::vector<uint8_t> data; struct ggml_context * ctx_out_ggml;
struct ggml_context * ctx; std::ofstream fout;
uint32_t lora_r; lora_merge_ctx(
uint32_t lora_alpha; std::string & base_fname,
}; std::vector<std::tuple<std::string, float>> & lora_files,
std::string & outfile,
int n_threads) : base_model(base_fname, 0), n_threads(n_threads), fout(outfile, std::ios::binary) {
fout.exceptions(std::ofstream::failbit); // fail fast on write errors
struct llama_file { if (gguf_find_key(base_model.ctx_gguf, LLM_KV_SPLIT_COUNT) >= 0) {
// use FILE * so we don't have to re-open the file to mmap throw std::runtime_error("split model is not yet supported");
FILE * fp; }
size_t size;
llama_file(const char * fname, const char * mode) { for (auto lora_inp : lora_files) {
fp = std::fopen(fname, mode); auto fname = std::get<0>(lora_inp);
if (fp == NULL) { auto scale = std::get<1>(lora_inp);
size = 0; std::unique_ptr<file_input> adapter(new file_input(fname, scale));
check_metadata_lora(adapter.get());
adapters.push_back(std::move(adapter));
}
ctx_out = gguf_init_empty();
struct ggml_init_params params = {
/*.mem_size =*/ gguf_get_n_tensors(base_model.ctx_gguf)*ggml_tensor_overhead(),
/*.mem_buffer =*/ NULL,
/*.no_alloc =*/ true,
};
ctx_out_ggml = ggml_init(params);
backend = ggml_backend_cpu_init();
allocr = ggml_gallocr_new(ggml_backend_get_default_buffer_type(backend));
}
void check_metadata_lora(file_input * adapter) {
auto general_type = get_kv_str(adapter->ctx_gguf, "general.type");
if (general_type != "adapter") {
throw std::runtime_error("expect general.type to be 'adapter', but got: " + general_type);
}
auto adapter_type = get_kv_str(adapter->ctx_gguf, "adapter.type");
if (adapter_type != "lora") {
throw std::runtime_error("expect adapter.type to be 'lora', but got: " + adapter_type);
}
auto general_arch_base = get_kv_str(base_model.ctx_gguf, "general.architecture");
auto general_arch_lora = get_kv_str(adapter->ctx_gguf, "general.architecture");
if (general_arch_base != general_arch_lora) {
throw std::runtime_error("model arch and LoRA arch mismatch");
}
}
ggml_type get_out_tensor_type(struct ggml_tensor * t) {
if (t->type == GGML_TYPE_F32) {
return GGML_TYPE_F32;
} else { } else {
seek(0, SEEK_END); return GGML_TYPE_F16;
size = tell();
seek(0, SEEK_SET);
} }
} }
size_t tell() const { void run_merge() {
#ifdef _WIN32 // prepare metadata
__int64 ret = _ftelli64(fp); gguf_set_kv(ctx_out, base_model.ctx_gguf);
#else // output is forced to f16 for now
long ret = std::ftell(fp); gguf_set_val_u32(ctx_out, "general.file_type", LLAMA_FTYPE_MOSTLY_F16);
#endif
GGML_ASSERT(ret != -1); // this really shouldn't fail
return (size_t) ret;
}
void seek(size_t offset, int whence) { // check if all lora adapters have the same tensors
#ifdef _WIN32 // TODO: remove this when we can support merging subset of adapters. Ref: https://github.com/ggerganov/llama.cpp/pull/8607#discussion_r1686027777
int ret = _fseeki64(fp, (__int64) offset, whence); static const char * err_no_subset_adapter = "Input adapters do not have the same list of tensors. This is not yet supported. Please merge the adapter one-by-one instead of merging all at once.";
#else if (adapters.size() > 1) {
int ret = std::fseek(fp, (long) offset, whence); for (size_t i = 1; i < adapters.size(); ++i) {
#endif if (adapters[0]->tensors.size() != adapters[i]->tensors.size()) {
GGML_ASSERT(ret == 0); // same throw std::runtime_error(err_no_subset_adapter);
} }
for (auto & it : adapters[i]->tensors) {
void read_raw(void * ptr, size_t size) { if (adapters[0]->get_tensor(it.first) == nullptr) {
if (size == 0) { throw std::runtime_error(err_no_subset_adapter);
return;
} }
errno = 0;
std::size_t ret = std::fread(ptr, size, 1, fp);
if (ferror(fp)) {
die_fmt("read error: %s", strerror(errno));
} }
if (ret != 1) {
die("unexpectedly reached end of file");
} }
} }
std::uint32_t read_u32() { // if true, this tensor can be lora-merged. if false, we skip merging and just copy data to outfile
std::uint32_t ret; std::vector<std::pair<struct ggml_tensor *, bool>> base_tensors;
read_raw(&ret, sizeof(ret)); for (auto & it : base_model.tensors) {
return ret; bool t_a = true;
bool t_b = true;
for (auto & adapter : adapters) {
t_a &= nullptr != adapter->get_tensor(it.first + ".lora_a");
t_b &= nullptr != adapter->get_tensor(it.first + ".lora_b");
} }
auto base_tensor = it.second;
std::string read_string(std::uint32_t len) { struct ggml_tensor * out_tensor;
std::vector<char> chars(len); if (!t_a && !t_b) {
read_raw(chars.data(), len); // only copy
return std::string(chars.data(), len); out_tensor = ggml_dup_tensor(ctx_out_ggml, base_tensor);
} ggml_set_name(out_tensor, base_tensor->name);
base_tensors.push_back(std::make_pair(out_tensor, false));
void write_raw(const void * ptr, size_t size) { } else if (t_a && t_b) {
if (size == 0) { // need merging
return; out_tensor = ggml_dup_tensor(ctx_out_ggml, base_tensor);
} out_tensor->type = get_out_tensor_type(base_tensor);
errno = 0; ggml_set_name(out_tensor, base_tensor->name);
size_t ret = std::fwrite(ptr, size, 1, fp); base_tensors.push_back(std::make_pair(out_tensor, true));
if (ret != 1) {
die_fmt("write error: %s", strerror(errno));
}
}
void write_u32(std::uint32_t val) {
write_raw(&val, sizeof(val));
}
bool eof() {
return tell() >= size;
}
~llama_file() {
if (fp) {
std::fclose(fp);
}
}
};
static struct export_lora_params get_default_export_lora_params() {
struct export_lora_params result;
result.fn_model_base = "";
result.fn_model_out = "";
result.n_threads = GGML_DEFAULT_N_THREADS;
return result;
}
static void export_lora_print_usage(int /*argc*/, char ** argv, const struct export_lora_params * params) {
fprintf(stderr, "usage: %s [options]\n", argv[0]);
fprintf(stderr, "\n");
fprintf(stderr, "options:\n");
fprintf(stderr, " -h, --help show this help message and exit\n");
fprintf(stderr, " -m FNAME, --model-base FNAME model path from which to load base model (default '%s')\n", params->fn_model_base.c_str());
fprintf(stderr, " -o FNAME, --model-out FNAME path to save exported model (default '%s')\n", params->fn_model_out.c_str());
fprintf(stderr, " -l FNAME, --lora FNAME apply LoRA adapter\n");
fprintf(stderr, " -s FNAME S, --lora-scaled FNAME S apply LoRA adapter with user defined scaling S\n");
fprintf(stderr, " -t N, --threads N number of threads to use during computation (default: %d)\n", params->n_threads);
}
static bool export_lora_params_parse(int argc, char ** argv, struct export_lora_params * params) {
bool invalid_param = false;
std::string arg;
struct export_lora_params default_params = get_default_export_lora_params();
const std::string arg_prefix = "--";
for (int i = 1; i < argc; i++) {
arg = argv[i];
if (arg.compare(0, arg_prefix.size(), arg_prefix) == 0) {
std::replace(arg.begin(), arg.end(), '_', '-');
}
if (arg == "-m" || arg == "--model-base") {
if (++i >= argc) {
invalid_param = true;
break;
}
params->fn_model_base = argv[i];
} else if (arg == "-o" || arg == "--model-out") {
if (++i >= argc) {
invalid_param = true;
break;
}
params->fn_model_out = argv[i];
} else if (arg == "-l" || arg == "--lora") {
if (++i >= argc) {
invalid_param = true;
break;
}
struct lora_info lora;
lora.filename = argv[i];
lora.scale = 1.0f;
params->lora.push_back(lora);
} else if (arg == "-s" || arg == "--lora-scaled") {
if (++i >= argc) {
invalid_param = true;
break;
}
struct lora_info lora;
lora.filename = argv[i];
if (++i >= argc) {
invalid_param = true;
break;
}
lora.scale = std::stof(argv[i]);
params->lora.push_back(lora);
} else if (arg == "-t" || arg == "--threads") {
if (++i >= argc) {
invalid_param = true;
break;
}
params->n_threads = std::stoi(argv[i]);
if (params->n_threads <= 0) {
params->n_threads = std::thread::hardware_concurrency();
}
} else if (arg == "-h" || arg == "--help") {
export_lora_print_usage(argc, argv, &default_params);
exit(0);
} else { } else {
fprintf(stderr, "error: unknown argument: '%s'\n", arg.c_str()); throw std::runtime_error("tensor " + it.first + " missing either lora_a or lora_b");
export_lora_print_usage(argc, argv, &default_params); }
exit(1); gguf_add_tensor(ctx_out, out_tensor);
}
// placeholder for the meta data
{
size_t meta_size = gguf_get_meta_size(ctx_out);
zeros(fout, meta_size);
}
// process base model tensors
size_t n_merged = 0;
for (auto & it : base_tensors) {
if (it.second) {
merge_tensor(it.first);
n_merged++;
} else {
copy_tensor(it.first);
} }
} }
if (params->fn_model_base == default_params.fn_model_base) { // write output metadata
fprintf(stderr, "error: please specify a filename for model-base.\n"); {
export_lora_print_usage(argc, argv, &default_params); std::vector<uint8_t> data(gguf_get_meta_size(ctx_out));
exit(1); gguf_get_meta_data(ctx_out, data.data());
} fout.seekp(0);
if (params->fn_model_out == default_params.fn_model_out) { fout.write((const char *)data.data(), data.size());
fprintf(stderr, "error: please specify a filename for model-out.\n");
export_lora_print_usage(argc, argv, &default_params);
exit(1);
}
if (invalid_param) {
fprintf(stderr, "error: invalid parameter for argument: '%s'\n", arg.c_str());
export_lora_print_usage(argc, argv, &default_params);
exit(1);
}
return true;
}
static void free_lora(struct lora_data * lora) {
if (lora->ctx != NULL) {
ggml_free(lora->ctx);
}
delete lora;
}
static struct lora_data * load_lora(struct lora_info * info) {
struct lora_data * result = new struct lora_data;
result->info = *info;
result->ctx = NULL;
result->lora_r = 1;
result->lora_alpha = 1;
struct llama_file file(info->filename.c_str(), "rb");
if (file.fp == NULL) {
fprintf(stderr, "warning: Could not open lora adapter '%s'. Ignoring this adapter.\n",
info->filename.c_str());
free_lora(result);
return NULL;
} }
struct ggml_init_params params_ggml; printf("%s : merged %ld tensors with lora adapters\n", __func__, n_merged);
params_ggml.mem_size = ggml_tensor_overhead() * GGML_DEFAULT_GRAPH_SIZE; printf("%s : wrote %ld tensors to output file\n", __func__, base_tensors.size());
params_ggml.mem_buffer = NULL;
params_ggml.no_alloc = true;
result->ctx = ggml_init(params_ggml);
uint32_t magic = file.read_u32();
if (magic != LLAMA_FILE_MAGIC_GGLA) {
die_fmt("unexpected lora header file magic in '%s'", info->filename.c_str());
}
uint32_t version = file.read_u32();
if (version != 1) {
die_fmt("unexpected lora file version '%u' in '%s'", (unsigned) version, info->filename.c_str());
}
result->lora_r = file.read_u32();
result->lora_alpha = file.read_u32();
// read tensor infos from file
std::vector<char> name_buf;
std::vector<struct ggml_tensor *> tensors;
std::vector<size_t> tensors_offset;
size_t total_nbytes_pad = 0;
while(!file.eof()) {
int64_t ne[4] = {1,1,1,1};
uint32_t n_dims = file.read_u32();
uint32_t namelen = file.read_u32();
uint32_t type = file.read_u32();
for (uint32_t k = 0; k < n_dims; ++k) {
ne[k] = (int64_t)file.read_u32();
}
name_buf.clear();
name_buf.resize(namelen + 1, '\0');
file.read_raw(name_buf.data(), namelen);
file.seek((0-file.tell()) & 31, SEEK_CUR);
size_t offset = file.tell();
struct ggml_tensor * tensor = ggml_new_tensor(result->ctx, (enum ggml_type) type, n_dims, ne);
ggml_set_name(tensor, name_buf.data());
size_t nbytes = ggml_nbytes(tensor);
size_t nbytes_pad = ggml_nbytes_pad(tensor);
total_nbytes_pad += nbytes_pad;
tensors.push_back(tensor);
tensors_offset.push_back(offset);
file.seek(nbytes, SEEK_CUR);
}
// read tensor data
result->data.resize(total_nbytes_pad);
size_t data_offset = 0;
for (size_t i = 0; i < tensors.size(); ++i) {
struct ggml_tensor * tensor = tensors[i];
size_t offset = tensors_offset[i];
size_t nbytes = ggml_nbytes(tensor);
size_t nbytes_pad = ggml_nbytes_pad(tensor);
file.seek(offset, SEEK_SET);
tensor->data = result->data.data() + data_offset;
file.read_raw(tensor->data, nbytes);
data_offset += nbytes_pad;
}
return result;
}
static struct ggml_cgraph * build_graph_lora(
struct ggml_context * ctx,
struct ggml_tensor * tensor,
struct ggml_tensor * lora_a,
struct ggml_tensor * lora_b,
float scaling
) {
struct ggml_tensor * ab = ggml_mul_mat(ctx, lora_a, lora_b);
if (scaling != 1.0f) {
ab = ggml_scale(ctx, ab, scaling);
}
struct ggml_tensor * res = ggml_add_inplace(ctx, tensor, ab);
struct ggml_cgraph * gf = ggml_new_graph(ctx);
ggml_build_forward_expand (gf, res);
return gf;
}
static bool apply_lora(struct ggml_tensor * tensor, struct lora_data * lora, int n_threads) {
if (lora->ctx == NULL) {
return false;
}
std::string name = ggml_get_name(tensor);
std::string name_a = name + std::string(".loraA");
std::string name_b = name + std::string(".loraB");
struct ggml_tensor * lora_a = ggml_get_tensor(lora->ctx, name_a.c_str());
struct ggml_tensor * lora_b = ggml_get_tensor(lora->ctx, name_b.c_str());
if (lora_a == NULL || lora_b == NULL) {
return false;
} }
float scaling = lora->info.scale * (float)lora->lora_alpha / (float)lora->lora_r; void copy_tensor(struct ggml_tensor * base) {
printf("%s : %s [%s]\n", __func__, base->name, ggml_ne_string(base).c_str());
size_t len = ggml_nbytes(base);
base_model.read_tensor_data(base->name, read_buf);
fout.write((char* )read_buf.data(), len);
zeros(fout, GGML_PAD(len, GGUF_DEFAULT_ALIGNMENT) - len);
}
struct ggml_init_params params; void merge_tensor(struct ggml_tensor * base) {
params.mem_size = GGML_OBJECT_SIZE + ggml_graph_overhead() + ggml_tensor_overhead()*4 + GGML_MEM_ALIGN*5; std::string name_base(base->name);
params.mem_buffer = NULL; std::string name_lora_a = name_base + ".lora_a";
params.no_alloc = true; std::string name_lora_b = name_base + ".lora_b";
struct ggml_context * ctx = NULL;
struct ggml_gallocr * alloc = NULL;
struct ggml_cgraph * gf = NULL;
ctx = ggml_init(params); printf("%s : %s [%s]\n", __func__, base->name, ggml_ne_string(base).c_str());
alloc = ggml_gallocr_new(ggml_backend_cpu_buffer_type());
gf = build_graph_lora(ctx, tensor, lora_a, lora_b, scaling);
ggml_gallocr_alloc_graph(alloc, gf); // context for input tensor
std::vector<struct ggml_tensor *> inp_a(adapters.size());
std::vector<struct ggml_tensor *> inp_b(adapters.size());
struct ggml_init_params params {
/*.mem_size =*/ ggml_tensor_overhead()*(1+adapters.size()*2),
/*.mem_buffer =*/ NULL,
/*.no_alloc =*/ true,
};
struct ggml_context * ctx = ggml_init(params);
struct ggml_cplan cplan = ggml_graph_plan(gf, n_threads); // alloc tensors
static std::vector<uint8_t> data_work; struct ggml_tensor * inp = ggml_dup_tensor(ctx, base);
data_work.resize(cplan.work_size); for (size_t i = 0; i < adapters.size(); ++i) {
cplan.work_data = data_work.data(); auto t_a = adapters[i]->get_tensor(name_lora_a);
auto t_b = adapters[i]->get_tensor(name_lora_b);
inp_a[i] = ggml_dup_tensor(ctx, t_a);
inp_b[i] = ggml_dup_tensor(ctx, t_b);
}
ggml_backend_buffer_t buffer = ggml_backend_alloc_ctx_tensors(ctx, backend);
ggml_graph_compute(gf, &cplan); // load data to backend buffer
base_model.read_tensor_data(name_base, read_buf);
ggml_backend_tensor_set(inp, read_buf.data(), 0, ggml_nbytes(inp));
for (size_t i = 0; i < adapters.size(); ++i) {
adapters[i]->read_tensor_data(name_lora_a, read_buf);
ggml_backend_tensor_set(inp_a[i], read_buf.data(), 0, ggml_nbytes(inp_a[i]));
adapters[i]->read_tensor_data(name_lora_b, read_buf);
ggml_backend_tensor_set(inp_b[i], read_buf.data(), 0, ggml_nbytes(inp_b[i]));
}
// build graph
struct ggml_cgraph * gf;
{
static size_t buf_size = ggml_tensor_overhead()*GGML_DEFAULT_GRAPH_SIZE + ggml_graph_overhead();
static std::vector<uint8_t> buf(buf_size);
struct ggml_init_params params0 = {
/*.mem_size =*/ buf_size,
/*.mem_buffer =*/ buf.data(),
/*.no_alloc =*/ true,
};
struct ggml_context * ctx0 = ggml_init(params0);
gf = ggml_new_graph(ctx0);
struct ggml_tensor * cur = inp;
for (size_t i = 0; i < adapters.size(); ++i) {
struct ggml_tensor * a_T = ggml_cont(ctx0, ggml_transpose(ctx0, inp_a[i]));
struct ggml_tensor * delta = ggml_mul_mat(ctx0, a_T, inp_b[i]);
// scale
const float alpha = adapters[i]->alpha;
const float rank = (float) inp_b[i]->ne[0];
const float scale = alpha ? adapters[i]->scale * alpha / rank : adapters[i]->scale;
delta = ggml_scale(ctx0, delta, scale);
cur = ggml_add(ctx0, cur, delta);
printf("%s : + merging from adapter[%ld]\n", __func__, i);
printf("%s : input_scale=%f calculated_scale=%f rank=%d\n", __func__, adapters[i]->scale, scale, (int) inp_b[i]->ne[0]);
}
cur = ggml_cast(ctx0, cur, get_out_tensor_type(base));
ggml_build_forward_expand(gf, cur);
ggml_free(ctx0);
}
// compute
{
ggml_gallocr_alloc_graph(allocr, gf);
ggml_backend_cpu_set_n_threads(backend, n_threads);
ggml_backend_graph_compute(backend, gf);
}
// write data to output file
{
auto result = gf->nodes[gf->n_nodes - 1];
size_t len = ggml_nbytes(result);
if (read_buf.size() < len) {
read_buf.resize(len);
}
ggml_backend_tensor_get(result, read_buf.data(), 0, len);
fout.write((char* )read_buf.data(), len);
zeros(fout, GGML_PAD(len, GGUF_DEFAULT_ALIGNMENT) - len);
}
ggml_gallocr_free(alloc);
ggml_free(ctx); ggml_free(ctx);
return true; ggml_backend_buffer_free(buffer);
}
static void export_lora(struct export_lora_params * params) {
// load all loras
std::vector<struct lora_data *> loras;
for (size_t i = 0; i < params->lora.size(); ++i) {
struct lora_data * lora = load_lora(&params->lora[i]);
if (lora != NULL) {
loras.push_back(lora);
}
}
if (loras.size() == 0) {
fprintf(stderr, "warning: no lora adapters will be applied.\n");
} }
// open input file ~lora_merge_ctx() {
struct llama_file fin(params->fn_model_base.c_str(), "rb"); ggml_gallocr_free(allocr);
if (!fin.fp) { ggml_backend_free(backend);
die_fmt("Could not open file '%s'\n", params->fn_model_base.c_str()); gguf_free(ctx_out);
ggml_free(ctx_out_ggml);
} }
};
// open base model gguf, read tensors without their data static void print_usage(int argc, char ** argv, const gpt_params & params) {
struct ggml_context * ctx_in; gpt_params_print_usage(argc, argv, params);
struct gguf_init_params params_gguf;
params_gguf.no_alloc = true;
params_gguf.ctx = &ctx_in;
struct gguf_context * gguf_in = gguf_init_from_file(params->fn_model_base.c_str(), params_gguf);
// create new gguf printf("\nexample usage:\n");
struct gguf_context * gguf_out = gguf_init_empty(); printf("\n %s -m base-model.gguf --lora lora-file.gguf -o merged-model-f16.gguf\n", argv[0]);
printf("\nNOTE: output model is F16\n");
// copy meta data from base model: kv and tensors
gguf_set_kv(gguf_out, gguf_in);
int n_tensors = gguf_get_n_tensors(gguf_in);
for (int i=0; i < n_tensors; ++i) {
const char * name = gguf_get_tensor_name(gguf_in, i);
struct ggml_tensor * tensor = ggml_get_tensor(ctx_in, name);
gguf_add_tensor(gguf_out, tensor);
}
// create output file
struct llama_file fout(params->fn_model_out.c_str(), "wb");
if (!fout.fp) {
die_fmt("Could not create file '%s'\n", params->fn_model_out.c_str());
}
// write gguf meta data
std::vector<uint8_t> meta;
meta.resize(gguf_get_meta_size(gguf_out));
gguf_get_meta_data(gguf_out, meta.data());
fout.write_raw(meta.data(), meta.size());
std::vector<uint8_t> data;
std::vector<uint8_t> padding;
for (int i=0; i < n_tensors; ++i) {
const char * name = gguf_get_tensor_name(gguf_in, i);
struct ggml_tensor * tensor = ggml_get_tensor(ctx_in, name);
// read tensor data
data.resize(ggml_nbytes(tensor));
tensor->data = data.data();
size_t offset = gguf_get_tensor_offset(gguf_in, i);
fin.seek(offset + meta.size(), SEEK_SET);
fin.read_raw(data.data(), data.size());
// apply all loras
for (size_t k = 0; k < loras.size(); ++k) {
apply_lora(tensor, loras[k], params->n_threads);
}
// write tensor data + padding
padding.clear();
padding.resize(GGML_PAD(data.size(), gguf_get_alignment(gguf_out)) - data.size(), 0);
GGML_ASSERT(fout.tell() == offset + meta.size());
// fout.seek(offset + meta.size(), SEEK_SET);
fout.write_raw(data.data(), data.size());
fout.write_raw(padding.data(), padding.size());
if (i % 2 == 0) {
printf(".");
}
}
printf("\n"); printf("\n");
// close gguf
gguf_free(gguf_out);
gguf_free(gguf_in);
// free loras
for (size_t i = 0; i < loras.size(); ++i) {
free_lora(loras[i]);
}
} }
int main(int argc, char ** argv) { int main(int argc, char ** argv) {
struct export_lora_params params = get_default_export_lora_params(); gpt_params params;
if (!export_lora_params_parse(argc, argv, &params)) { if (!gpt_params_parse(argc, argv, params)) {
print_usage(argc, argv, params);
return 1; return 1;
} }
export_lora(&params); g_verbose = (params.verbosity == 1);
try {
lora_merge_ctx ctx(params.model, params.lora_adapter, params.lora_outfile, params.n_threads);
ctx.run_merge();
} catch (const std::exception & err) {
fprintf(stderr, "%s\n", err.what());
exit(EXIT_FAILURE);
}
printf("done, output file is %s\n", params.lora_outfile.c_str());
return 0; return 0;
} }