Skip to content

GitLab

Commit 0ce8bcfd authored by xuxzh1's avatar xuxzh1 🎱
Browse files

init

parent b0135f4b
Show whitespace changes
Inline Side-by-side
Showing with 1411 additions and 486 deletions
llm/llama.cpp/examples/embedding/embedding.cpp
View file @ 0ce8bcfd
......@@ -7,23 +7,30 @@
#pragma warning(disable: 4244 4267) // possible loss of data
#endif
static std::vector<std::string> split_lines(const std::string & s) {
std::string line;
static std::vector<std::string> split_lines(const std::string & s, const std::string & separator = "\n") {
std::vector<std::string> lines;
std::stringstream ss(s);
while (std::getline(ss, line)) {
lines.push_back(line);
size_t start = 0;
size_t end = s.find(separator);
while (end != std::string::npos) {
lines.push_back(s.substr(start, end - start));
start = end + separator.length();
end = s.find(separator, start);
}
lines.push_back(s.substr(start)); // Add the last part
return lines;
}
static void batch_add_seq(llama_batch & batch, const std::vector<int32_t> & tokens, int seq_id) {
for (size_t i = 0; i < tokens.size(); i++) {
llama_batch_add(batch, tokens[i], i, { seq_id }, i == tokens.size() - 1);
static void batch_add_seq(llama_batch & batch, const std::vector<int32_t> & tokens, llama_seq_id seq_id) {
size_t n_tokens = tokens.size();
for (size_t i = 0; i < n_tokens; i++) {
llama_batch_add(batch, tokens[i], i, { seq_id }, true);
}
}
static void batch_decode(llama_context * ctx, llama_batch & batch, float * output, int n_seq, int n_embd) {
static void batch_decode(llama_context * ctx, llama_batch & batch, float * output, int n_seq, int n_embd, int embd_norm) {
// clear previous kv_cache values (irrelevant for embeddings)
llama_kv_cache_clear(ctx);
......@@ -40,22 +47,10 @@ static void batch_decode(llama_context * ctx, llama_batch & batch, float * outpu
// try to get sequence embeddings - supported only when pooling_type is not NONE
const float * embd = llama_get_embeddings_seq(ctx, batch.seq_id[i][0]);
if (embd == NULL) {
embd = llama_get_embeddings_ith(ctx, i);
if (embd == NULL) {
fprintf(stderr, "%s: failed to get embeddings for token %d\n", __func__, i);
continue;
}
}
GGML_ASSERT(embd != NULL && "failed to get sequence embeddings");
float * out = output + batch.seq_id[i][0] * n_embd;
//TODO: I would also add a parameter here to enable normalization or not.
/*fprintf(stdout, "unnormalized_embedding:");
for (int hh = 0; hh < n_embd; hh++) {
fprintf(stdout, "%9.6f ", embd[hh]);
}
fprintf(stdout, "\n");*/
llama_embd_normalize(embd, out, n_embd);
llama_embd_normalize(embd, out, n_embd, embd_norm);
}
}
......@@ -63,6 +58,7 @@ int main(int argc, char ** argv) {
gpt_params params;
if (!gpt_params_parse(argc, argv, params)) {
gpt_params_print_usage(argc, argv, params);
return 1;
}
......@@ -79,18 +75,15 @@ int main(int argc, char ** argv) {
fprintf(stderr, "%s: seed = %u\n", __func__, params.seed);
std::mt19937 rng(params.seed);
if (params.random_prompt) {
params.prompt = string_random_prompt(rng);
}
llama_backend_init();
llama_numa_init(params.numa);
llama_model * model;
llama_context * ctx;
// load the model
std::tie(model, ctx) = llama_init_from_gpt_params(params);
llama_init_result llama_init = llama_init_from_gpt_params(params);
llama_model * model = llama_init.model;
llama_context * ctx = llama_init.context;
if (model == NULL) {
fprintf(stderr, "%s: error: unable to load model\n", __func__);
return 1;
......@@ -99,6 +92,12 @@ int main(int argc, char ** argv) {
const int n_ctx_train = llama_n_ctx_train(model);
const int n_ctx = llama_n_ctx(ctx);
const enum llama_pooling_type pooling_type = llama_pooling_type(ctx);
if (pooling_type == LLAMA_POOLING_TYPE_NONE) {
fprintf(stderr, "%s: error: pooling type NONE not supported\n", __func__);
return 1;
}
if (n_ctx > n_ctx_train) {
fprintf(stderr, "%s: warning: model was trained on only %d context tokens (%d specified)\n",
__func__, n_ctx_train, n_ctx);
......@@ -111,7 +110,7 @@ int main(int argc, char ** argv) {
}
// split the prompt into lines
std::vector<std::string> prompts = split_lines(params.prompt);
std::vector<std::string> prompts = split_lines(params.prompt, params.embd_sep);
// max batch size
const uint64_t n_batch = params.n_batch;
......@@ -171,7 +170,7 @@ int main(int argc, char ** argv) {
// encode if at capacity
if (batch.n_tokens + n_toks > n_batch) {
float * out = emb + p * n_embd;
batch_decode(ctx, batch, out, s, n_embd);
batch_decode(ctx, batch, out, s, n_embd, params.embd_normalize);
llama_batch_clear(batch);
p += s;
s = 0;
......@@ -184,15 +183,20 @@ int main(int argc, char ** argv) {
// final batch
float * out = emb + p * n_embd;
batch_decode(ctx, batch, out, s, n_embd);
batch_decode(ctx, batch, out, s, n_embd, params.embd_normalize);
if (params.embd_out.empty()) {
// print the first part of the embeddings or for a single prompt, the full embedding
fprintf(stdout, "\n");
for (int j = 0; j < n_prompts; j++) {
fprintf(stdout, "embedding %d: ", j);
for (int i = 0; i < (n_prompts > 1 ? std::min(16, n_embd) : n_embd); i++) {
if (params.embd_normalize == 0) {
fprintf(stdout, "%6.0f ", emb[j * n_embd + i]);
} else {
fprintf(stdout, "%9.6f ", emb[j * n_embd + i]);
}
}
fprintf(stdout, "\n");
}
......@@ -200,14 +204,58 @@ int main(int argc, char ** argv) {
if (n_prompts > 1) {
fprintf(stdout, "\n");
printf("cosine similarity matrix:\n\n");
for (int i = 0; i < n_prompts; i++) {
fprintf(stdout, "%6.6s ", prompts[i].c_str());
}
fprintf(stdout, "\n");
for (int i = 0; i < n_prompts; i++) {
for (int j = 0; j < n_prompts; j++) {
float sim = llama_embd_similarity_cos(emb + i * n_embd, emb + j * n_embd, n_embd);
fprintf(stdout, "%6.2f ", sim);
}
fprintf(stdout, "%1.10s", prompts[i].c_str());
fprintf(stdout, "\n");
}
}
}
if (params.embd_out == "json" || params.embd_out == "json+" || params.embd_out == "array") {
const bool notArray = params.embd_out != "array";
fprintf(stdout, notArray ? "{\n \"object\": \"list\",\n \"data\": [\n" : "[");
for (int j = 0;;) { // at least one iteration (one prompt)
if (notArray) fprintf(stdout, " {\n \"object\": \"embedding\",\n \"index\": %d,\n \"embedding\": ",j);
fprintf(stdout, "[");
for (int i = 0;;) { // at least one iteration (n_embd > 0)
fprintf(stdout, params.embd_normalize == 0 ? "%1.0f" : "%1.7f", emb[j * n_embd + i]);
i++;
if (i < n_embd) fprintf(stdout, ","); else break;
}
fprintf(stdout, notArray ? "]\n }" : "]");
j++;
if (j < n_prompts) fprintf(stdout, notArray ? ",\n" : ","); else break;
}
fprintf(stdout, notArray ? "\n ]" : "]\n");
if (params.embd_out == "json+" && n_prompts > 1) {
fprintf(stdout, ",\n \"cosineSimilarity\": [\n");
for (int i = 0;;) { // at least two iteration (n_prompts > 1)
fprintf(stdout, " [");
for (int j = 0;;) { // at least two iteration (n_prompts > 1)
float sim = llama_embd_similarity_cos(emb + i * n_embd, emb + j * n_embd, n_embd);
fprintf(stdout, "%6.2f", sim);
j++;
if (j < n_prompts) fprintf(stdout, ", "); else break;
}
fprintf(stdout, " ]");
i++;
if (i < n_prompts) fprintf(stdout, ",\n"); else break;
}
fprintf(stdout, "\n ]");
}
if (notArray) fprintf(stdout, "\n}\n");
}
// clean up
llama_print_timings(ctx);
......
llm/llama.cpp/examples/eval-callback/CMakeLists.txt
View file @ 0ce8bcfd
set(TARGET eval-callback)
set(TARGET llama-eval-callback)
add_executable(${TARGET} eval-callback.cpp)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_11)
set(TEST_TARGET test-eval-callback)
add_test(NAME ${TEST_TARGET} COMMAND eval-callback --hf-repo ggml-org/models --hf-file tinyllamas/stories260K.gguf --model stories260K.gguf --prompt hello --seed 42 -ngl 0)
add_test(NAME ${TEST_TARGET} COMMAND llama-eval-callback --hf-repo ggml-org/models --hf-file tinyllamas/stories260K.gguf --model stories260K.gguf --prompt hello --seed 42 -ngl 0)
set_property(TEST ${TEST_TARGET} PROPERTY LABELS eval-callback curl)
llm/llama.cpp/examples/eval-callback/README.md
View file @ 0ce8bcfd
......@@ -6,7 +6,7 @@ It simply prints to the console all operations and tensor data.
Usage:
```shell
eval-callback \
llama-eval-callback \
--hf-repo ggml-org/models \
--hf-file phi-2/ggml-model-q4_0.gguf \
--model phi-2-q4_0.gguf \
......
llm/llama.cpp/examples/eval-callback/eval-callback.cpp
View file @ 0ce8bcfd
......@@ -62,7 +62,7 @@ static void ggml_print_tensor(uint8_t * data, ggml_type type, const int64_t * ne
} else if (type == GGML_TYPE_I8) {
v = (float) *(int8_t *) &data[i];
} else {
GGML_ASSERT(false);
GGML_ABORT("fatal error");
}
printf("%12.4f", v);
sum += v;
......@@ -99,7 +99,7 @@ static bool ggml_debug(struct ggml_tensor * t, bool ask, void * user_data) {
char src1_str[128] = {0};
if (src1) {
sprintf(src1_str, "%s{%s}", src1->name, ggml_ne_string(src1).c_str());
snprintf(src1_str, sizeof(src1_str), "%s{%s}", src1->name, ggml_ne_string(src1).c_str());
}
printf("%s: %24s = (%s) %10s(%s{%s}, %s}) = {%s}\n", __func__,
......@@ -140,20 +140,18 @@ static bool run(llama_context * ctx, const gpt_params & params) {
}
int main(int argc, char ** argv) {
callback_data cb_data;
gpt_params params;
if (!gpt_params_parse(argc, argv, params)) {
gpt_params_print_usage(argc, argv, params);
return 1;
}
print_build_info();
std::mt19937 rng(params.seed);
if (params.random_prompt) {
params.prompt = string_random_prompt(rng);
}
llama_backend_init();
llama_numa_init(params.numa);
......@@ -165,9 +163,10 @@ int main(int argc, char ** argv) {
params.warmup = false;
// init
llama_model * model;
llama_context * ctx;
std::tie(model, ctx) = llama_init_from_gpt_params(params);
llama_init_result llama_init = llama_init_from_gpt_params(params);
llama_model * model = llama_init.model;
llama_context * ctx = llama_init.context;
if (model == nullptr || ctx == nullptr) {
fprintf(stderr, "%s : failed to init\n", __func__);
return 1;
......
llm/llama.cpp/examples/export-lora/CMakeLists.txt
View file @ 0ce8bcfd
set(TARGET export-lora)
set(TARGET llama-export-lora)
add_executable(${TARGET} export-lora.cpp)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
......
llm/llama.cpp/examples/export-lora/README.md
View file @ 0ce8bcfd
......@@ -3,24 +3,31 @@
Apply LORA adapters to base model and export the resulting model.
```
usage: export-lora [options]
usage: llama-export-lora [options]
options:
-h, --help show this help message and exit
-m FNAME, --model-base FNAME model path from which to load base model (default '')
-o FNAME, --model-out FNAME path to save exported model (default '')
-l FNAME, --lora FNAME apply LoRA adapter
-s FNAME S, --lora-scaled FNAME S apply LoRA adapter with user defined scaling S
-t N, --threads N number of threads to use during computation (default: 4)
-m, --model model path from which to load base model (default '')
--lora FNAME path to LoRA adapter (can be repeated to use multiple adapters)
--lora-scaled FNAME S path to LoRA adapter with user defined scaling S (can be repeated to use multiple adapters)
-t, --threads N number of threads to use during computation (default: 4)
-o, --output FNAME output file (default: 'ggml-lora-merged-f16.gguf')
```
For example:
```bash
./bin/export-lora \
./bin/llama-export-lora \
-m open-llama-3b-v2-q8_0.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.gguf
```
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:
```bash
./bin/llama-export-lora \
-m your_base_model.gguf \
-o your_merged_model.gguf \
--lora-scaled lora_task_A.gguf 0.5 \
--lora-scaled lora_task_B.gguf 0.5
```
llm/llama.cpp/examples/export-lora/export-lora.cpp
View file @ 0ce8bcfd
#include "common.h"
#include "ggml.h"
#include "ggml-alloc.h"
#include <map>
#include <vector>
#include <string>
#include <thread>
#include <fstream>
struct lora_info {
std::string filename;
float scale;
};
struct export_lora_params {
std::string fn_model_base;
std::string fn_model_out;
std::vector<struct lora_info> lora;
int n_threads;
};
static bool g_verbose = false;
struct lora_data {
struct lora_info info;
std::vector<uint8_t> data;
struct ggml_context * ctx;
uint32_t lora_r;
uint32_t lora_alpha;
};
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));
}
struct llama_file {
// use FILE * so we don't have to re-open the file to mmap
FILE * fp;
size_t size;
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);
}
llama_file(const char * fname, const char * mode) {
fp = std::fopen(fname, mode);
if (fp == NULL) {
size = 0;
} else {
seek(0, SEEK_END);
size = tell();
seek(0, SEEK_SET);
}
static void zeros(std::ofstream & file, size_t n) {
char zero = 0;
for (size_t i = 0; i < n; ++i) {
file.write(&zero, 1);
}
}
size_t tell() const {
#ifdef _WIN32
__int64 ret = _ftelli64(fp);
#else
long ret = std::ftell(fp);
#endif
GGML_ASSERT(ret != -1); // this really shouldn't fail
return (size_t) ret;
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 += ", ";
}
void seek(size_t offset, int whence) {
#ifdef _WIN32
int ret = _fseeki64(fp, (__int64) offset, whence);
#else
int ret = std::fseek(fp, (long) offset, whence);
#endif
GGML_ASSERT(ret == 0); // same
}
return str;
}
void read_raw(void * ptr, size_t size) {
if (size == 0) {
return;
}
errno = 0;
std::size_t ret = std::fread(ptr, size, 1, fp);
if (ferror(fp)) {
die_fmt("read error: %s", strerror(errno));
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);
}
if (ret != 1) {
die("unexpectedly reached end of file");
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);
}
std::uint32_t read_u32() {
std::uint32_t ret;
read_raw(&ret, sizeof(ret));
return ret;
}
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;
std::string read_string(std::uint32_t len) {
std::vector<char> chars(len);
read_raw(chars.data(), len);
return std::string(chars.data(), len);
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);
}
void write_raw(const void * ptr, size_t size) {
if (size == 0) {
return;
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);
}
errno = 0;
size_t ret = std::fwrite(ptr, size, 1, fp);
if (ret != 1) {
die_fmt("write error: %s", strerror(errno));
}
}
void write_u32(std::uint32_t val) {
write_raw(&val, sizeof(val));
ggml_tensor * get_tensor(std::string name) {
if (tensors.find(name) == tensors.end()) {
return nullptr;
}
bool eof() {
return tell() >= size;
return tensors[name];
}
~llama_file() {
if (fp) {
std::fclose(fp);
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);
}
};
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;
}
struct lora_merge_ctx {
// input base model + adapters
file_input base_model;
std::vector<std::unique_ptr<file_input>> adapters;
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);
}
// for computing merged tensor
int n_threads;
ggml_backend_t backend = nullptr;
ggml_gallocr_t allocr = nullptr;
std::vector<uint8_t> read_buf;
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 = "--";
// output file
struct gguf_context * ctx_out;
struct ggml_context * ctx_out_ggml;
std::ofstream fout;
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(), '_', '-');
}
lora_merge_ctx(
std::string & base_fname,
std::vector<llama_lora_adapter_info> & 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
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 {
fprintf(stderr, "error: unknown argument: '%s'\n", arg.c_str());
export_lora_print_usage(argc, argv, &default_params);
exit(1);
}
if (gguf_find_key(base_model.ctx_gguf, LLM_KV_SPLIT_COUNT) >= 0) {
throw std::runtime_error("split model is not yet supported");
}
if (params->fn_model_base == default_params.fn_model_base) {
fprintf(stderr, "error: please specify a filename for model-base.\n");
export_lora_print_usage(argc, argv, &default_params);
exit(1);
}
if (params->fn_model_out == default_params.fn_model_out) {
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);
for (auto & lora_inp : lora_files) {
auto fname = lora_inp.path;
auto scale = lora_inp.scale;
std::unique_ptr<file_input> adapter(new file_input(fname, scale));
check_metadata_lora(adapter.get());
adapters.push_back(std::move(adapter));
}
return true;
}
static void free_lora(struct lora_data * lora) {
if (lora->ctx != NULL) {
ggml_free(lora->ctx);
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));
}
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;
params_ggml.mem_size = ggml_tensor_overhead() * GGML_DEFAULT_GRAPH_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);
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);
}
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;
}
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);
}
static bool apply_lora(struct ggml_tensor * tensor, struct lora_data * lora, int n_threads) {
if (lora->ctx == NULL) {
return false;
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");
}
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;
struct ggml_init_params params;
params.mem_size = GGML_OBJECT_SIZE + ggml_graph_overhead() + ggml_tensor_overhead()*4 + GGML_MEM_ALIGN*5;
params.mem_buffer = NULL;
params.no_alloc = true;
struct ggml_context * ctx = NULL;
struct ggml_gallocr * alloc = NULL;
struct ggml_cgraph * gf = NULL;
ctx = ggml_init(params);
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);
struct ggml_cplan cplan = ggml_graph_plan(gf, n_threads);
static std::vector<uint8_t> data_work;
data_work.resize(cplan.work_size);
cplan.work_data = data_work.data();
ggml_graph_compute(gf, &cplan);
ggml_gallocr_free(alloc);
ggml_free(ctx);
return true;
}
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);
}
ggml_type get_out_tensor_type(struct ggml_tensor * t) {
if (t->type == GGML_TYPE_F32) {
return GGML_TYPE_F32;
} else {
return GGML_TYPE_F16;
}
}
void run_merge() {
// prepare metadata
gguf_set_kv(ctx_out, base_model.ctx_gguf);
// output is forced to f16 for now
gguf_set_val_u32(ctx_out, "general.file_type", LLAMA_FTYPE_MOSTLY_F16);
// check if all lora adapters have the same tensors
// TODO: remove this when we can support merging subset of adapters. Ref: https://github.com/ggerganov/llama.cpp/pull/8607#discussion_r1686027777
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.";
if (adapters.size() > 1) {
for (size_t i = 1; i < adapters.size(); ++i) {
if (adapters[0]->tensors.size() != adapters[i]->tensors.size()) {
throw std::runtime_error(err_no_subset_adapter);
}
for (auto & it : adapters[i]->tensors) {
if (adapters[0]->get_tensor(it.first) == nullptr) {
throw std::runtime_error(err_no_subset_adapter);
}
}
}
}
// mapping base tensor to out tensor (same shape with base, but different type)
// if out_tensor == nullptr, we only copy it
std::vector<std::pair<struct ggml_tensor *, struct ggml_tensor *>> base_to_out_tensors;
for (auto & it : base_model.tensors) {
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;
if (!t_a && !t_b) {
// only copy
struct ggml_tensor * cpy_tensor = ggml_dup_tensor(ctx_out_ggml, base_tensor);
ggml_set_name(cpy_tensor, base_tensor->name);
base_to_out_tensors.push_back(std::make_pair(cpy_tensor, nullptr));
gguf_add_tensor(ctx_out, cpy_tensor);
} else if (t_a && t_b) {
// need merging
struct ggml_tensor * out_tensor = ggml_new_tensor(
ctx_out_ggml, get_out_tensor_type(base_tensor), GGML_MAX_DIMS, base_tensor->ne);
ggml_set_name(out_tensor, base_tensor->name);
base_to_out_tensors.push_back(std::make_pair(base_tensor, out_tensor));
gguf_add_tensor(ctx_out, out_tensor);
} else {
throw std::runtime_error("tensor " + it.first + " missing either lora_a or lora_b");
}
if (loras.size() == 0) {
fprintf(stderr, "warning: no lora adapters will be applied.\n");
}
// open input file
struct llama_file fin(params->fn_model_base.c_str(), "rb");
if (!fin.fp) {
die_fmt("Could not open file '%s'\n", params->fn_model_base.c_str());
// placeholder for the meta data
{
size_t meta_size = gguf_get_meta_size(ctx_out);
zeros(fout, meta_size);
}
// open base model gguf, read tensors without their data
struct ggml_context * ctx_in;
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
struct gguf_context * gguf_out = gguf_init_empty();
// 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);
// process base model tensors
size_t n_merged = 0;
for (auto & it : base_to_out_tensors) {
if (it.second != nullptr) {
merge_tensor(it.first, it.second);
n_merged++;
} else {
copy_tensor(it.first);
}
}
// write output metadata
{
std::vector<uint8_t> data(gguf_get_meta_size(ctx_out));
gguf_get_meta_data(ctx_out, data.data());
fout.seekp(0);
fout.write((const char *)data.data(), data.size());
}
printf("%s : merged %ld tensors with lora adapters\n", __func__, n_merged);
printf("%s : wrote %ld tensors to output file\n", __func__, base_to_out_tensors.size());
}
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);
}
void merge_tensor(struct ggml_tensor * base, struct ggml_tensor * out) {
std::string name_base(base->name);
std::string name_lora_a = name_base + ".lora_a";
std::string name_lora_b = name_base + ".lora_b";
printf("%s : %s [%s]\n", __func__, base->name, ggml_ne_string(base).c_str());
// 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()*(2+adapters.size()*2),
/*.mem_buffer =*/ NULL,
/*.no_alloc =*/ true,
};
struct ggml_context * ctx = ggml_init(params);
// alloc tensors
struct ggml_tensor * inp_base = ggml_new_tensor(ctx, GGML_TYPE_F32, GGML_MAX_DIMS, base->ne);
for (size_t i = 0; i < adapters.size(); ++i) {
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);
// load base tensor to backend buffer
base_model.read_tensor_data(name_base, read_buf);
if (base->type != GGML_TYPE_F32) {
// optionally dequantize it
printf("%s : + dequantize base tensor from %s to F32\n", __func__, ggml_type_name(base->type));
auto nels = ggml_nelements(inp_base);
ggml_type_traits_t qtype = ggml_internal_get_type_traits(base->type);
std::vector<uint8_t> dequant_buf(nels * sizeof(float));
qtype.to_float(read_buf.data(), (float *)dequant_buf.data(), nels);
ggml_backend_tensor_set(inp_base, dequant_buf.data(), 0, dequant_buf.size());
} else {
ggml_backend_tensor_set(inp_base, read_buf.data(), 0, ggml_nbytes(inp_base));
}
// load lora tensors to backend buffer
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_base;
for (size_t i = 0; i < adapters.size(); ++i) {
struct ggml_tensor * a_T = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_cast(ctx0, inp_a[i], GGML_TYPE_F32)));
struct ggml_tensor * delta = ggml_mul_mat(ctx0, a_T, ggml_cast(ctx0, inp_b[i], GGML_TYPE_F32));
// 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, delta, cur);
printf("%s : + merging from adapter[%ld] type=%s\n", __func__, i, ggml_type_name(inp_a[i]->type));
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, out->type);
printf("%s : + output type is %s\n", __func__, ggml_type_name(out->type));
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);
}
// 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());
ggml_free(ctx);
ggml_backend_buffer_free(buffer);
}
// 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);
~lora_merge_ctx() {
ggml_gallocr_free(allocr);
ggml_backend_free(backend);
gguf_free(ctx_out);
ggml_free(ctx_out_ggml);
}
};
// 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());
static void print_usage(int argc, char ** argv, const gpt_params & params) {
gpt_params_print_usage(argc, argv, params);
if (i % 2 == 0) {
printf(".");
}
}
printf("\nexample usage:\n");
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");
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) {
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;
}
export_lora(&params);
g_verbose = (params.verbosity == 1);
try {
lora_merge_ctx ctx(params.model, params.lora_adapters, 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;
}
llm/llama.cpp/examples/gbnf-validator/CMakeLists.txt
View file @ 0ce8bcfd
set(TARGET gbnf-validator)
set(TARGET llama-gbnf-validator)
add_executable(${TARGET} gbnf-validator.cpp)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE common grammar-parser llama ${CMAKE_THREAD_LIBS_INIT})
target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_11)
llm/llama.cpp/examples/gbnf-validator/gbnf-validator.cpp
View file @ 0ce8bcfd
......@@ -7,6 +7,8 @@
#include <cstdio>
#include <cstdlib>
#include <sstream>
#include <fstream>
#include <string>
#include <vector>
......@@ -14,20 +16,25 @@ static bool llama_sample_grammar_string(struct llama_grammar * grammar, const st
auto decoded = decode_utf8(input_str, {});
const auto & code_points = decoded.first;
const llama_grammar_rules & rules = llama_grammar_get_rules (grammar);
llama_grammar_stacks & cur_stacks = llama_grammar_get_stacks(grammar);
size_t pos = 0;
for (auto it = code_points.begin(), end = code_points.end() - 1; it != end; ++it) {
auto prev_stacks = grammar->stacks;
llama_grammar_accept(grammar->rules, prev_stacks, *it, grammar->stacks);
if (grammar->stacks.empty()) {
const llama_grammar_stacks prev_stacks = llama_grammar_get_stacks(grammar); // copy
llama_grammar_accept(rules, prev_stacks, *it, cur_stacks);
if (cur_stacks.empty()) {
error_pos = pos;
error_msg = "Unexpected character '" + unicode_cpt_to_utf8(*it) + "'";
grammar->stacks = prev_stacks;
cur_stacks = prev_stacks;
return false;
}
++pos;
}
for (const auto & stack : grammar->stacks) {
for (const auto & stack : cur_stacks) {
if (stack.empty()) {
return true;
}
......@@ -69,13 +76,14 @@ int main(int argc, char** argv) {
return 1;
}
fseek(grammar_file, 0, SEEK_END);
size_t grammar_size = ftell(grammar_file);
fseek(grammar_file, 0, SEEK_SET);
std::string grammar_str(grammar_size, ' ');
fread(&grammar_str[0], 1, grammar_size, grammar_file);
fclose(grammar_file);
std::string grammar_str;
{
std::ifstream grammar_file(grammar_filename);
GGML_ASSERT(grammar_file.is_open() && "Failed to open grammar file");
std::stringstream buffer;
buffer << grammar_file.rdbuf();
grammar_str = buffer.str();
}
// Parse the GBNF grammar
auto parsed_grammar = grammar_parser::parse(grammar_str.c_str());
......@@ -98,22 +106,19 @@ int main(int argc, char** argv) {
auto grammar = llama_grammar_init(
grammar_rules.data(),
grammar_rules.size(), parsed_grammar.symbol_ids.at("root"));
if (grammar == nullptr) {
throw std::runtime_error("Failed to initialize llama_grammar");
}
// Read the input file
FILE* input_file = fopen(input_filename.c_str(), "r");
if (!input_file) {
fprintf(stdout, "Failed to open input file: %s\n", input_filename.c_str());
return 1;
std::string input_str;
{
std::ifstream input_file(input_filename);
GGML_ASSERT(input_file.is_open() && "Failed to open input file");
std::stringstream buffer;
buffer << input_file.rdbuf();
input_str = buffer.str();
}
fseek(input_file, 0, SEEK_END);
size_t input_size = ftell(input_file);
fseek(input_file, 0, SEEK_SET);
std::string input_str(input_size, ' ');
fread(&input_str[0], 1, input_size, input_file);
fclose(input_file);
// Validate the input string against the grammar
size_t error_pos;
std::string error_msg;
......
llm/llama.cpp/examples/gguf-hash/CMakeLists.txt 0 → 100644
View file @ 0ce8bcfd
set(TARGET llama-gguf-hash)
add_executable(${TARGET} gguf-hash.cpp)
install(TARGETS ${TARGET} RUNTIME)
# clibs dependencies
include_directories(deps/)
add_library(xxhash OBJECT deps/xxhash/xxhash.c deps/xxhash/xxhash.h)
target_link_libraries(${TARGET} PRIVATE xxhash)
add_library(sha1 OBJECT deps/sha1/sha1.c deps/sha1/sha1.h)
target_link_libraries(${TARGET} PRIVATE sha1)
add_library(sha256 OBJECT deps/sha256/sha256.c deps/sha256/sha256.h)
target_link_libraries(${TARGET} PRIVATE sha256)
target_link_libraries(${TARGET} PRIVATE ggml ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_11)
llm/llama.cpp/examples/gguf-hash/README.md 0 → 100644
View file @ 0ce8bcfd
# llama-gguf-hash
CLI to hash GGUF files to detect difference on a per model and per tensor level.
**Command line options:**
- `--help`: display help message
- `--xxh64`: use xhash 64bit hash mode (default)
- `--sha1`: use sha1
- `--uuid`: use uuid
- `--sha256`: use sha256
- `--all`: use all hash
- `--no-layer`: exclude per layer hash
- `--uuid`: generate UUIDv5 ID
- `-c`, `--check <manifest>`: verify against a manifest
## About
While most POSIX systems already have hash checking programs like sha256sum, it
is designed to check entire files. This is not ideal for our purpose if we want
to check for consistency of the tensor data even if the metadata content of the
gguf KV store has been updated.
This program is designed to hash a gguf tensor payload on a 'per tensor layer'
in addition to a 'entire tensor model' hash. The intent is that the entire
tensor layer can be checked first but if there is any detected inconsistencies,
then the per tensor hash can be used to narrow down the specific tensor layer
that has inconsistencies.
For Maintainers:
- Detection of tensor inconsistency during development and automated tests
- This is served by xxh64 which is fast
- This is also served by having per tensor layer to assist in narrowing down
the location of the faulty tensor layer
- This is also served by sha1 which is much slower but more widely supported
For Model Creators:
- Optional consistent UUID generation based on model tensor content
- This is served by UUIDv5 which is useful for databases keys
- llama.cpp UUIDv5 Namespace: `ef001206-dadc-5f6d-a15f-3359e577d4e5`
- Made via UUIDv5 URL namespace of `en.wikipedia.org/wiki/Llama.cpp`
For Model Users:
- Assurance of tensor layer integrity even if metadata was updated
- This is served by sha256 which is still considered very secure as of 2024
### Design Note
- The default behavior of this program if no arguments is provided is to hash
using xxhash's xxh32 mode because it is very fast and is primarily targeted
towards maintainers who may want to use this in automated tests.
- xxhash support xxh32 and xxh128 for 32bit hash and 128bit hash respectively
however we picked 64bit xxhash as most computers are 64bit as of 2024 and thus
would have a better affinity to calculating hash that is 64bit in size.
## Compile Example
```bash
cmake -B build -DCMAKE_BUILD_TYPE=Debug -DLLAMA_FATAL_WARNINGS=ON
make -C build clean
make -C build llama-gguf-hash VERBOSE=1
./build/bin/llama-gguf-hash test.gguf
./build/bin/llama-gguf-hash --xxh64 test.gguf
./build/bin/llama-gguf-hash --sha1 test.gguf
./build/bin/llama-gguf-hash --uuid test.gguf
./build/bin/llama-gguf-hash --sha256 test.gguf
```
## Generation and Verification Example
To generate we may use this command
```bash
./llama-gguf-hash --all test.gguf > test.gguf.manifest
```
Which would generate a manifest that looks like below, which contains multiple hash type and per tensor layer hashes as well
(This excludes UUID as that is an ID not a hash)
```bash
xxh64 f66e9cd66a4396a0 test.gguf:tensor_0
sha1 59f79ecefd8125a996fdf419239051a7e99e5f20 test.gguf:tensor_0
sha256 c0510d38fa060c46265e0160a85c7243096b01dd31c2f355bdbb5516b20de1bd test.gguf:tensor_0
xxh64 7d3a1f9ac04d0537 test.gguf:tensor_1
sha1 4765f592eacf096df4628ba59476af94d767080a test.gguf:tensor_1
sha256 8514cbcc73692a2c56bd7a33a022edd5ff819614bd23b19915d7224387f397a7 test.gguf:tensor_1
xxh64 a0af5d700049693b test.gguf:tensor_2
sha1 25cbfbad4513cc348e2c95ebdee69d6ff2fd8753 test.gguf:tensor_2
sha256 947e6b36e20f2cc95e1d2ce1c1669d813d574657ac6b5ac5196158d454d35180 test.gguf:tensor_2
xxh64 e83fddf559d7b6a6 test.gguf:tensor_3
sha1 a9cba73e2d90f2ee3dae2548caa42bef3fe6a96c test.gguf:tensor_3
sha256 423b044e016d8ac73c39f23f60bf01bedef5ecb03c0230accd824c91fe86f1a1 test.gguf:tensor_3
xxh64 1257733306b7992d test.gguf:tensor_4
sha1 d7bc61db93bb685ce9d598da89717c66729b7543 test.gguf:tensor_4
sha256 79737cb3912d4201384cf7f16a1a37ff7823f23ea796cb205b6ca361ab9e3ebf test.gguf:tensor_4
xxh64 d238d16ba4711e58 test.gguf:tensor_5
sha1 0706566c198fe1072f37e0a5135b4b5f23654c52 test.gguf:tensor_5
sha256 60949be8298eced0ecdde64487643d018407bd261691e061d9e9c3dbc9fd358b test.gguf:tensor_5
xxh64 3fbc3b65ab8c7f39 test.gguf:tensor_6
sha1 73922a0727226a409049f6fc3172a52219ca6f00 test.gguf:tensor_6
sha256 574f4c46ff384a3b9a225eb955d2a871847a2e8b3fa59387a8252832e92ef7b0 test.gguf:tensor_6
xxh64 c22021c29854f093 test.gguf:tensor_7
sha1 efc39cece6a951188fc41e354c73bbfe6813d447 test.gguf:tensor_7
sha256 4c0410cd3c500f078ae5b21e8dc9eb79e29112713b2ab58a882f82a3868d4d75 test.gguf:tensor_7
xxh64 936df61f5d64261f test.gguf:tensor_8
sha1 c2490296d789a4f34398a337fed8377d943d9f06 test.gguf:tensor_8
sha256 c4401313feeba0261275c3b25bd2d8fe40ce04e0f440c2980ed0e9674c30ff01 test.gguf:tensor_8
xxh64 93fd20c64421c081 test.gguf:tensor_9
sha1 7047ce1e78437a6884337a3751c7ee0421918a65 test.gguf:tensor_9
sha256 23d57cf0d7a6e90b0b3616b41300e0cd354781e812add854a5f95aa55f2bc514 test.gguf:tensor_9
xxh64 5a54d3aad816f302 test.gguf
sha1 d15be52c4ff213e823cb6dd13af7ee2f978e7042 test.gguf
sha256 7dd641b32f59b60dbd4b5420c4b0f6321ccf48f58f6ae201a3dbc4a58a27c6e4 test.gguf
```
We can then use the normal check command which will by default check for the highest security strength hash and verify against that:
```bash
$ ./llama-gguf-hash --check test.gguf.manifest test.gguf
manifest test.gguf.manifest sha256 sha1 xxh64
sha256 c0510d38fa060c46265e0160a85c7243096b01dd31c2f355bdbb5516b20de1bd test.gguf:tensor_0 - Ok
sha256 8514cbcc73692a2c56bd7a33a022edd5ff819614bd23b19915d7224387f397a7 test.gguf:tensor_1 - Ok
sha256 947e6b36e20f2cc95e1d2ce1c1669d813d574657ac6b5ac5196158d454d35180 test.gguf:tensor_2 - Ok
sha256 423b044e016d8ac73c39f23f60bf01bedef5ecb03c0230accd824c91fe86f1a1 test.gguf:tensor_3 - Ok
sha256 79737cb3912d4201384cf7f16a1a37ff7823f23ea796cb205b6ca361ab9e3ebf test.gguf:tensor_4 - Ok
sha256 60949be8298eced0ecdde64487643d018407bd261691e061d9e9c3dbc9fd358b test.gguf:tensor_5 - Ok
sha256 574f4c46ff384a3b9a225eb955d2a871847a2e8b3fa59387a8252832e92ef7b0 test.gguf:tensor_6 - Ok
sha256 4c0410cd3c500f078ae5b21e8dc9eb79e29112713b2ab58a882f82a3868d4d75 test.gguf:tensor_7 - Ok
sha256 c4401313feeba0261275c3b25bd2d8fe40ce04e0f440c2980ed0e9674c30ff01 test.gguf:tensor_8 - Ok
sha256 23d57cf0d7a6e90b0b3616b41300e0cd354781e812add854a5f95aa55f2bc514 test.gguf:tensor_9 - Ok
sha256 7dd641b32f59b60dbd4b5420c4b0f6321ccf48f58f6ae201a3dbc4a58a27c6e4 test.gguf - Ok
Verification results for test.gguf.manifest - Success
```
Or we may explicitly ask for a faster hash like:
```bash
$ ./llama-gguf-hash --check test.gguf.manifest --xxh64 test.gguf
manifest test.gguf.manifest sha256 sha1 xxh64
xxh64 f66e9cd66a4396a0 test.gguf:tensor_0 - Ok
xxh64 7d3a1f9ac04d0537 test.gguf:tensor_1 - Ok
xxh64 a0af5d700049693b test.gguf:tensor_2 - Ok
xxh64 e83fddf559d7b6a6 test.gguf:tensor_3 - Ok
xxh64 1257733306b7992d test.gguf:tensor_4 - Ok
xxh64 d238d16ba4711e58 test.gguf:tensor_5 - Ok
xxh64 3fbc3b65ab8c7f39 test.gguf:tensor_6 - Ok
xxh64 c22021c29854f093 test.gguf:tensor_7 - Ok
xxh64 936df61f5d64261f test.gguf:tensor_8 - Ok
xxh64 93fd20c64421c081 test.gguf:tensor_9 - Ok
xxh64 5a54d3aad816f302 test.gguf - Ok
Verification results for test.gguf.manifest - Success
```
Or maybe we want to just check that all the hash is valid:
```bash
$./llama-gguf-hash --check test.gguf.manifest --all test.gguf.manifest
manifest test.gguf.manifest sha256 sha1 xxh64
xxh64 f66e9cd66a4396a0 test.gguf:tensor_0 - Ok
sha1 59f79ecefd8125a996fdf419239051a7e99e5f20 test.gguf:tensor_0 - Ok
sha256 c0510d38fa060c46265e0160a85c7243096b01dd31c2f355bdbb5516b20de1bd test.gguf:tensor_0 - Ok
xxh64 7d3a1f9ac04d0537 test.gguf:tensor_1 - Ok
sha1 4765f592eacf096df4628ba59476af94d767080a test.gguf:tensor_1 - Ok
sha256 8514cbcc73692a2c56bd7a33a022edd5ff819614bd23b19915d7224387f397a7 test.gguf:tensor_1 - Ok
xxh64 a0af5d700049693b test.gguf:tensor_2 - Ok
sha1 25cbfbad4513cc348e2c95ebdee69d6ff2fd8753 test.gguf:tensor_2 - Ok
sha256 947e6b36e20f2cc95e1d2ce1c1669d813d574657ac6b5ac5196158d454d35180 test.gguf:tensor_2 - Ok
xxh64 e83fddf559d7b6a6 test.gguf:tensor_3 - Ok
sha1 a9cba73e2d90f2ee3dae2548caa42bef3fe6a96c test.gguf:tensor_3 - Ok
sha256 423b044e016d8ac73c39f23f60bf01bedef5ecb03c0230accd824c91fe86f1a1 test.gguf:tensor_3 - Ok
xxh64 1257733306b7992d test.gguf:tensor_4 - Ok
sha1 d7bc61db93bb685ce9d598da89717c66729b7543 test.gguf:tensor_4 - Ok
sha256 79737cb3912d4201384cf7f16a1a37ff7823f23ea796cb205b6ca361ab9e3ebf test.gguf:tensor_4 - Ok
xxh64 d238d16ba4711e58 test.gguf:tensor_5 - Ok
sha1 0706566c198fe1072f37e0a5135b4b5f23654c52 test.gguf:tensor_5 - Ok
sha256 60949be8298eced0ecdde64487643d018407bd261691e061d9e9c3dbc9fd358b test.gguf:tensor_5 - Ok
xxh64 3fbc3b65ab8c7f39 test.gguf:tensor_6 - Ok
sha1 73922a0727226a409049f6fc3172a52219ca6f00 test.gguf:tensor_6 - Ok
sha256 574f4c46ff384a3b9a225eb955d2a871847a2e8b3fa59387a8252832e92ef7b0 test.gguf:tensor_6 - Ok
xxh64 c22021c29854f093 test.gguf:tensor_7 - Ok
sha1 efc39cece6a951188fc41e354c73bbfe6813d447 test.gguf:tensor_7 - Ok
sha256 4c0410cd3c500f078ae5b21e8dc9eb79e29112713b2ab58a882f82a3868d4d75 test.gguf:tensor_7 - Ok
xxh64 936df61f5d64261f test.gguf:tensor_8 - Ok
sha1 c2490296d789a4f34398a337fed8377d943d9f06 test.gguf:tensor_8 - Ok
sha256 c4401313feeba0261275c3b25bd2d8fe40ce04e0f440c2980ed0e9674c30ff01 test.gguf:tensor_8 - Ok
xxh64 93fd20c64421c081 test.gguf:tensor_9 - Ok
sha1 7047ce1e78437a6884337a3751c7ee0421918a65 test.gguf:tensor_9 - Ok
sha256 23d57cf0d7a6e90b0b3616b41300e0cd354781e812add854a5f95aa55f2bc514 test.gguf:tensor_9 - Ok
xxh64 5a54d3aad816f302 test.gguf - Ok
sha1 d15be52c4ff213e823cb6dd13af7ee2f978e7042 test.gguf - Ok
sha256 7dd641b32f59b60dbd4b5420c4b0f6321ccf48f58f6ae201a3dbc4a58a27c6e4 test.gguf - Ok
Verification results for test.gguf.manifest - Success
```
## Crypto/Hash Libraries Used
These micro c libraries dependencies was installed via the [clib c package manager](https://github.com/clibs)
- https://github.com/Cyan4973/xxHash
- https://github.com/clibs/sha1/
- https://github.com/jb55/sha256.c
llm/llama.cpp/examples/gguf-hash/deps/rotate-bits/package.json 0 → 100644
View file @ 0ce8bcfd
{
"name": "rotate-bits",
"version": "0.1.1",
"repo": "jb55/rotate-bits.h",
"description": "rotate bits",
"keywords": ["rotl", "rotr"],
"src": ["rotate-bits.h"],
"license": "Public Domain",
"development": {
"thlorenz/tap.c": "*"
}
}
llm/llama.cpp/examples/gguf-hash/deps/rotate-bits/rotate-bits.h 0 → 100644
View file @ 0ce8bcfd
#ifndef __ROTATE_DEFS_H
#define __ROTATE_DEFS_H
#ifdef _MSC_VER
#include <stdlib.h>
#define ROTL32(v, n) _rotl((v), (n))
#define ROTL64(v, n) _rotl64((v), (n))
#define ROTR32(v, n) _rotr((v), (n))
#define ROTR64(v, n) _rotr64((v), (n))
#else
#include <stdint.h>
#define U8V(v) ((uint8_t)(v) & 0xFFU)
#define U16V(v) ((uint16_t)(v) & 0xFFFFU)
#define U32V(v) ((uint32_t)(v) & 0xFFFFFFFFU)
#define U64V(v) ((uint64_t)(v) & 0xFFFFFFFFFFFFFFFFU)
#define ROTL32(v, n) \
(U32V((uint32_t)(v) << (n)) | ((uint32_t)(v) >> (32 - (n))))
// tests fail if we don't have this cast...
#define ROTL64(v, n) \
(U64V((uint64_t)(v) << (n)) | ((uint64_t)(v) >> (64 - (n))))
#define ROTR32(v, n) ROTL32(v, 32 - (n))
#define ROTR64(v, n) ROTL64(v, 64 - (n))
#endif
#define ROTL8(v, n) \
(U8V((uint8_t)(v) << (n)) | ((uint8_t)(v) >> (8 - (n))))
#define ROTL16(v, n) \
(U16V((uint16_t)(v) << (n)) | ((uint16_t)(v) >> (16 - (n))))
#define ROTR8(v, n) ROTL8(v, 8 - (n))
#define ROTR16(v, n) ROTL16(v, 16 - (n))
#endif
llm/llama.cpp/examples/gguf-hash/deps/sha1/package.json 0 → 100644
View file @ 0ce8bcfd
{
"name": "sha1",
"version": "0.0.1",
"repo": "clibs/sha1",
"description": "sha1 hash algorithm",
"keywords": ["sha1", "hash"],
"license": "public domain",
"src": ["sha1.c", "sha1.h"]
}
llm/llama.cpp/examples/gguf-hash/deps/sha1/sha1.c 0 → 100644
View file @ 0ce8bcfd
/*
SHA-1 in C
By Steve Reid <steve@edmweb.com>
100% Public Domain
Test Vectors (from FIPS PUB 180-1)
"abc"
A9993E36 4706816A BA3E2571 7850C26C 9CD0D89D
"abcdbcdecdefdefgefghfghighijhijkijkljklmklmnlmnomnopnopq"
84983E44 1C3BD26E BAAE4AA1 F95129E5 E54670F1
A million repetitions of "a"
34AA973C D4C4DAA4 F61EEB2B DBAD2731 6534016F
*/
/* #define LITTLE_ENDIAN * This should be #define'd already, if true. */
/* #define SHA1HANDSOFF * Copies data before messing with it. */
#define SHA1HANDSOFF
#include <stdio.h>
#include <string.h>
/* for uint32_t */
#include <stdint.h>
#include "sha1.h"
#define rol(value, bits) (((value) << (bits)) | ((value) >> (32 - (bits))))
/* blk0() and blk() perform the initial expand. */
/* I got the idea of expanding during the round function from SSLeay */
#if BYTE_ORDER == LITTLE_ENDIAN
#define blk0(i) (block->l[i] = (rol(block->l[i],24)&0xFF00FF00) \
|(rol(block->l[i],8)&0x00FF00FF))
#elif BYTE_ORDER == BIG_ENDIAN
#define blk0(i) block->l[i]
#else
#error "Endianness not defined!"
#endif
#define blk(i) (block->l[i&15] = rol(block->l[(i+13)&15]^block->l[(i+8)&15] \
^block->l[(i+2)&15]^block->l[i&15],1))
/* (R0+R1), R2, R3, R4 are the different operations used in SHA1 */
#define R0(v,w,x,y,z,i) z+=((w&(x^y))^y)+blk0(i)+0x5A827999+rol(v,5);w=rol(w,30);
#define R1(v,w,x,y,z,i) z+=((w&(x^y))^y)+blk(i)+0x5A827999+rol(v,5);w=rol(w,30);
#define R2(v,w,x,y,z,i) z+=(w^x^y)+blk(i)+0x6ED9EBA1+rol(v,5);w=rol(w,30);
#define R3(v,w,x,y,z,i) z+=(((w|x)&y)|(w&x))+blk(i)+0x8F1BBCDC+rol(v,5);w=rol(w,30);
#define R4(v,w,x,y,z,i) z+=(w^x^y)+blk(i)+0xCA62C1D6+rol(v,5);w=rol(w,30);
/* Hash a single 512-bit block. This is the core of the algorithm. */
void SHA1Transform(
uint32_t state[5],
const unsigned char buffer[64]
)
{
uint32_t a, b, c, d, e;
typedef union
{
unsigned char c[64];
uint32_t l[16];
} CHAR64LONG16;
#ifdef SHA1HANDSOFF
CHAR64LONG16 block[1]; /* use array to appear as a pointer */
memcpy(block, buffer, 64);
#else
/* The following had better never be used because it causes the
* pointer-to-const buffer to be cast into a pointer to non-const.
* And the result is written through. I threw a "const" in, hoping
* this will cause a diagnostic.
*/
CHAR64LONG16 *block = (const CHAR64LONG16 *) buffer;
#endif
/* Copy context->state[] to working vars */
a = state[0];
b = state[1];
c = state[2];
d = state[3];
e = state[4];
/* 4 rounds of 20 operations each. Loop unrolled. */
R0(a, b, c, d, e, 0);
R0(e, a, b, c, d, 1);
R0(d, e, a, b, c, 2);
R0(c, d, e, a, b, 3);
R0(b, c, d, e, a, 4);
R0(a, b, c, d, e, 5);
R0(e, a, b, c, d, 6);
R0(d, e, a, b, c, 7);
R0(c, d, e, a, b, 8);
R0(b, c, d, e, a, 9);
R0(a, b, c, d, e, 10);
R0(e, a, b, c, d, 11);
R0(d, e, a, b, c, 12);
R0(c, d, e, a, b, 13);
R0(b, c, d, e, a, 14);
R0(a, b, c, d, e, 15);
R1(e, a, b, c, d, 16);
R1(d, e, a, b, c, 17);
R1(c, d, e, a, b, 18);
R1(b, c, d, e, a, 19);
R2(a, b, c, d, e, 20);
R2(e, a, b, c, d, 21);
R2(d, e, a, b, c, 22);
R2(c, d, e, a, b, 23);
R2(b, c, d, e, a, 24);
R2(a, b, c, d, e, 25);
R2(e, a, b, c, d, 26);
R2(d, e, a, b, c, 27);
R2(c, d, e, a, b, 28);
R2(b, c, d, e, a, 29);
R2(a, b, c, d, e, 30);
R2(e, a, b, c, d, 31);
R2(d, e, a, b, c, 32);
R2(c, d, e, a, b, 33);
R2(b, c, d, e, a, 34);
R2(a, b, c, d, e, 35);
R2(e, a, b, c, d, 36);
R2(d, e, a, b, c, 37);
R2(c, d, e, a, b, 38);
R2(b, c, d, e, a, 39);
R3(a, b, c, d, e, 40);
R3(e, a, b, c, d, 41);
R3(d, e, a, b, c, 42);
R3(c, d, e, a, b, 43);
R3(b, c, d, e, a, 44);
R3(a, b, c, d, e, 45);
R3(e, a, b, c, d, 46);
R3(d, e, a, b, c, 47);
R3(c, d, e, a, b, 48);
R3(b, c, d, e, a, 49);
R3(a, b, c, d, e, 50);
R3(e, a, b, c, d, 51);
R3(d, e, a, b, c, 52);
R3(c, d, e, a, b, 53);
R3(b, c, d, e, a, 54);
R3(a, b, c, d, e, 55);
R3(e, a, b, c, d, 56);
R3(d, e, a, b, c, 57);
R3(c, d, e, a, b, 58);
R3(b, c, d, e, a, 59);
R4(a, b, c, d, e, 60);
R4(e, a, b, c, d, 61);
R4(d, e, a, b, c, 62);
R4(c, d, e, a, b, 63);
R4(b, c, d, e, a, 64);
R4(a, b, c, d, e, 65);
R4(e, a, b, c, d, 66);
R4(d, e, a, b, c, 67);
R4(c, d, e, a, b, 68);
R4(b, c, d, e, a, 69);
R4(a, b, c, d, e, 70);
R4(e, a, b, c, d, 71);
R4(d, e, a, b, c, 72);
R4(c, d, e, a, b, 73);
R4(b, c, d, e, a, 74);
R4(a, b, c, d, e, 75);
R4(e, a, b, c, d, 76);
R4(d, e, a, b, c, 77);
R4(c, d, e, a, b, 78);
R4(b, c, d, e, a, 79);
/* Add the working vars back into context.state[] */
state[0] += a;
state[1] += b;
state[2] += c;
state[3] += d;
state[4] += e;
/* Wipe variables */
a = b = c = d = e = 0;
#ifdef SHA1HANDSOFF
memset(block, '\0', sizeof(block));
#endif
}
/* SHA1Init - Initialize new context */
void SHA1Init(
SHA1_CTX * context
)
{
/* SHA1 initialization constants */
context->state[0] = 0x67452301;
context->state[1] = 0xEFCDAB89;
context->state[2] = 0x98BADCFE;
context->state[3] = 0x10325476;
context->state[4] = 0xC3D2E1F0;
context->count[0] = context->count[1] = 0;
}
/* Run your data through this. */
void SHA1Update(
SHA1_CTX * context,
const unsigned char *data,
uint32_t len
)
{
uint32_t i;
uint32_t j;
j = context->count[0];
if ((context->count[0] += len << 3) < j)
context->count[1]++;
context->count[1] += (len >> 29);
j = (j >> 3) & 63;
if ((j + len) > 63)
{
memcpy(&context->buffer[j], data, (i = 64 - j));
SHA1Transform(context->state, context->buffer);
for (; i + 63 < len; i += 64)
{
SHA1Transform(context->state, &data[i]);
}
j = 0;
}
else
i = 0;
memcpy(&context->buffer[j], &data[i], len - i);
}
/* Add padding and return the message digest. */
void SHA1Final(
unsigned char digest[20],
SHA1_CTX * context
)
{
unsigned i;
unsigned char finalcount[8];
unsigned char c;
#if 0 /* untested "improvement" by DHR */
/* Convert context->count to a sequence of bytes
* in finalcount. Second element first, but
* big-endian order within element.
* But we do it all backwards.
*/
unsigned char *fcp = &finalcount[8];
for (i = 0; i < 2; i++)
{
uint32_t t = context->count[i];
int j;
for (j = 0; j < 4; t >>= 8, j++)
*--fcp = (unsigned char) t}
#else
for (i = 0; i < 8; i++)
{
finalcount[i] = (unsigned char) ((context->count[(i >= 4 ? 0 : 1)] >> ((3 - (i & 3)) * 8)) & 255); /* Endian independent */
}
#endif
c = 0200;
SHA1Update(context, &c, 1);
while ((context->count[0] & 504) != 448)
{
c = 0000;
SHA1Update(context, &c, 1);
}
SHA1Update(context, finalcount, 8); /* Should cause a SHA1Transform() */
for (i = 0; i < 20; i++)
{
digest[i] = (unsigned char)
((context->state[i >> 2] >> ((3 - (i & 3)) * 8)) & 255);
}
/* Wipe variables */
memset(context, '\0', sizeof(*context));
memset(&finalcount, '\0', sizeof(finalcount));
}
void SHA1(
char *hash_out,
const char *str,
uint32_t len)
{
SHA1_CTX ctx;
unsigned int ii;
SHA1Init(&ctx);
for (ii=0; ii<len; ii+=1)
SHA1Update(&ctx, (const unsigned char*)str + ii, 1);
SHA1Final((unsigned char *)hash_out, &ctx);
}
llm/llama.cpp/examples/gguf-hash/deps/sha1/sha1.h 0 → 100644
View file @ 0ce8bcfd
#ifndef SHA1_H
#define SHA1_H
/*
SHA-1 in C
By Steve Reid <steve@edmweb.com>
100% Public Domain
*/
#include "stdint.h"
#if defined(__cplusplus)
extern "C" {
#endif
typedef struct
{
uint32_t state[5];
uint32_t count[2];
unsigned char buffer[64];
} SHA1_CTX;
void SHA1Transform(
uint32_t state[5],
const unsigned char buffer[64]
);
void SHA1Init(
SHA1_CTX * context
);
void SHA1Update(
SHA1_CTX * context,
const unsigned char *data,
uint32_t len
);
void SHA1Final(
unsigned char digest[20],
SHA1_CTX * context
);
void SHA1(
char *hash_out,
const char *str,
uint32_t len);
#if defined(__cplusplus)
}
#endif
#endif /* SHA1_H */
llm/llama.cpp/examples/gguf-hash/deps/sha256/package.json 0 → 100644
View file @ 0ce8bcfd
{
"name": "sha256",
"version": "0.0.2",
"repo": "jb55/sha256.c",
"description": "sha256 in c",
"keywords": ["sha256", "sha2"],
"src": ["sha256.c", "sha256.h"],
"dependencies": {
"jb55/rotate-bits.h": "0.1.1"
},
"development": {
"thlorenz/tap.c": "*"
}
}
llm/llama.cpp/examples/gguf-hash/deps/sha256/sha256.c 0 → 100644
View file @ 0ce8bcfd
/* Crypto/Sha256.c -- SHA-256 Hash
2010-06-11 : Igor Pavlov : Public domain
This code is based on public domain code from Wei Dai's Crypto++ library. */
#include "rotate-bits/rotate-bits.h"
#include "sha256.h"
/* define it for speed optimization */
#define _SHA256_UNROLL
#define _SHA256_UNROLL2
void
sha256_init(sha256_t *p)
{
p->state[0] = 0x6a09e667;
p->state[1] = 0xbb67ae85;
p->state[2] = 0x3c6ef372;
p->state[3] = 0xa54ff53a;
p->state[4] = 0x510e527f;
p->state[5] = 0x9b05688c;
p->state[6] = 0x1f83d9ab;
p->state[7] = 0x5be0cd19;
p->count = 0;
}
#define S0(x) (ROTR32(x, 2) ^ ROTR32(x,13) ^ ROTR32(x, 22))
#define S1(x) (ROTR32(x, 6) ^ ROTR32(x,11) ^ ROTR32(x, 25))
#define s0(x) (ROTR32(x, 7) ^ ROTR32(x,18) ^ (x >> 3))
#define s1(x) (ROTR32(x,17) ^ ROTR32(x,19) ^ (x >> 10))
#define blk0(i) (W[i] = data[i])
#define blk2(i) (W[i&15] += s1(W[(i-2)&15]) + W[(i-7)&15] + s0(W[(i-15)&15]))
#define Ch(x,y,z) (z^(x&(y^z)))
#define Maj(x,y,z) ((x&y)|(z&(x|y)))
#define a(i) T[(0-(i))&7]
#define b(i) T[(1-(i))&7]
#define c(i) T[(2-(i))&7]
#define d(i) T[(3-(i))&7]
#define e(i) T[(4-(i))&7]
#define f(i) T[(5-(i))&7]
#define g(i) T[(6-(i))&7]
#define h(i) T[(7-(i))&7]
#ifdef _SHA256_UNROLL2
#define R(a,b,c,d,e,f,g,h, i) h += S1(e) + Ch(e,f,g) + K[i+j] + (j?blk2(i):blk0(i));\
d += h; h += S0(a) + Maj(a, b, c)
#define RX_8(i) \
R(a,b,c,d,e,f,g,h, i); \
R(h,a,b,c,d,e,f,g, (i+1)); \
R(g,h,a,b,c,d,e,f, (i+2)); \
R(f,g,h,a,b,c,d,e, (i+3)); \
R(e,f,g,h,a,b,c,d, (i+4)); \
R(d,e,f,g,h,a,b,c, (i+5)); \
R(c,d,e,f,g,h,a,b, (i+6)); \
R(b,c,d,e,f,g,h,a, (i+7))
#else
#define R(i) h(i) += S1(e(i)) + Ch(e(i),f(i),g(i)) + K[i+j] + (j?blk2(i):blk0(i));\
d(i) += h(i); h(i) += S0(a(i)) + Maj(a(i), b(i), c(i))
#ifdef _SHA256_UNROLL
#define RX_8(i) R(i+0); R(i+1); R(i+2); R(i+3); R(i+4); R(i+5); R(i+6); R(i+7);
#endif
#endif
static const uint32_t K[64] = {
0x428a2f98, 0x71374491, 0xb5c0fbcf, 0xe9b5dba5,
0x3956c25b, 0x59f111f1, 0x923f82a4, 0xab1c5ed5,
0xd807aa98, 0x12835b01, 0x243185be, 0x550c7dc3,
0x72be5d74, 0x80deb1fe, 0x9bdc06a7, 0xc19bf174,
0xe49b69c1, 0xefbe4786, 0x0fc19dc6, 0x240ca1cc,
0x2de92c6f, 0x4a7484aa, 0x5cb0a9dc, 0x76f988da,
0x983e5152, 0xa831c66d, 0xb00327c8, 0xbf597fc7,
0xc6e00bf3, 0xd5a79147, 0x06ca6351, 0x14292967,
0x27b70a85, 0x2e1b2138, 0x4d2c6dfc, 0x53380d13,
0x650a7354, 0x766a0abb, 0x81c2c92e, 0x92722c85,
0xa2bfe8a1, 0xa81a664b, 0xc24b8b70, 0xc76c51a3,
0xd192e819, 0xd6990624, 0xf40e3585, 0x106aa070,
0x19a4c116, 0x1e376c08, 0x2748774c, 0x34b0bcb5,
0x391c0cb3, 0x4ed8aa4a, 0x5b9cca4f, 0x682e6ff3,
0x748f82ee, 0x78a5636f, 0x84c87814, 0x8cc70208,
0x90befffa, 0xa4506ceb, 0xbef9a3f7, 0xc67178f2
};
static void
sha256_transform(uint32_t *state, const uint32_t *data)
{
uint32_t W[16] = {0};
unsigned j;
#ifdef _SHA256_UNROLL2
uint32_t a,b,c,d,e,f,g,h;
a = state[0];
b = state[1];
c = state[2];
d = state[3];
e = state[4];
f = state[5];
g = state[6];
h = state[7];
#else
uint32_t T[8];
for (j = 0; j < 8; j++)
T[j] = state[j];
#endif
for (j = 0; j < 64; j += 16)
{
#if defined(_SHA256_UNROLL) || defined(_SHA256_UNROLL2)
RX_8(0); RX_8(8);
#else
unsigned i;
for (i = 0; i < 16; i++) { R(i); }
#endif
}
#ifdef _SHA256_UNROLL2
state[0] += a;
state[1] += b;
state[2] += c;
state[3] += d;
state[4] += e;
state[5] += f;
state[6] += g;
state[7] += h;
#else
for (j = 0; j < 8; j++)
state[j] += T[j];
#endif
/* Wipe variables */
/* memset(W, 0, sizeof(W)); */
/* memset(T, 0, sizeof(T)); */
}
#undef S0
#undef S1
#undef s0
#undef s1
static void
sha256_write_byte_block(sha256_t *p)
{
uint32_t data32[16];
unsigned i;
for (i = 0; i < 16; i++)
data32[i] =
((uint32_t)(p->buffer[i * 4 ]) << 24) +
((uint32_t)(p->buffer[i * 4 + 1]) << 16) +
((uint32_t)(p->buffer[i * 4 + 2]) << 8) +
((uint32_t)(p->buffer[i * 4 + 3]));
sha256_transform(p->state, data32);
}
void
sha256_hash(unsigned char *buf, const unsigned char *data, size_t size)
{
sha256_t hash;
sha256_init(&hash);
sha256_update(&hash, data, size);
sha256_final(&hash, buf);
}
void
sha256_update(sha256_t *p, const unsigned char *data, size_t size)
{
uint32_t curBufferPos = (uint32_t)p->count & 0x3F;
while (size > 0)
{
p->buffer[curBufferPos++] = *data++;
p->count++;
size--;
if (curBufferPos == 64)
{
curBufferPos = 0;
sha256_write_byte_block(p);
}
}
}
void
sha256_final(sha256_t *p, unsigned char *digest)
{
uint64_t lenInBits = (p->count << 3);
uint32_t curBufferPos = (uint32_t)p->count & 0x3F;
unsigned i;
p->buffer[curBufferPos++] = 0x80;
while (curBufferPos != (64 - 8))
{
curBufferPos &= 0x3F;
if (curBufferPos == 0)
sha256_write_byte_block(p);
p->buffer[curBufferPos++] = 0;
}
for (i = 0; i < 8; i++)
{
p->buffer[curBufferPos++] = (unsigned char)(lenInBits >> 56);
lenInBits <<= 8;
}
sha256_write_byte_block(p);
for (i = 0; i < 8; i++)
{
*digest++ = (unsigned char)(p->state[i] >> 24);
*digest++ = (unsigned char)(p->state[i] >> 16);
*digest++ = (unsigned char)(p->state[i] >> 8);
*digest++ = (unsigned char)(p->state[i]);
}
sha256_init(p);
}
llm/llama.cpp/examples/gguf-hash/deps/sha256/sha256.h 0 → 100644
View file @ 0ce8bcfd
/* Sha256.h -- SHA-256 Hash
2010-06-11 : Igor Pavlov : Public domain */
#ifndef __CRYPTO_SHA256_H
#define __CRYPTO_SHA256_H
#include <stdlib.h>
#include <stdint.h>
#define SHA256_DIGEST_SIZE 32
typedef struct sha256_t
{
uint32_t state[8];
uint64_t count;
unsigned char buffer[64];
} sha256_t;
void sha256_init(sha256_t *p);
void sha256_update(sha256_t *p, const unsigned char *data, size_t size);
void sha256_final(sha256_t *p, unsigned char *digest);
void sha256_hash(unsigned char *buf, const unsigned char *data, size_t size);
#endif
llm/llama.cpp/examples/gguf-hash/deps/xxhash/clib.json 0 → 100644
View file @ 0ce8bcfd
{
"name": "xxhash",
"version": "0.8.2",
"repo": "Cyan4973/xxhash",
"description": "Extremely fast non-cryptographic hash algorithm",
"keywords": ["xxhash", "hashing"],
"license": "BSD-2-Clause",
"src": [
"xxhash.c",
"xxhash.h"
]
}
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment