extern/whisper.cpp
1
0
Fork 1
mirror of https://github.com/ggerganov/whisper.cpp.git synced 2025-08-16 02:28:16 +02:00
Code Issues Packages Projects Releases Wiki Activity

ggml : sync (ggml-alloc, GPU, eps, etc.) (#1220)

Browse Source 
* ggml : sync (ggml-alloc, GPU, eps, etc.)

* ggml : fix build

* wasm : fix build
This commit is contained in:
Georgi Gerganov
2023-09-05 13:54:40 +03:00
committed by GitHub
parent 6780c98e19
commit 59a3d0cb57
17 changed files with 11254 additions and 4095 deletions
Download Patch File Download Diff File Expand all files Collapse all files

94
examples/common.cpp
View File

@ -1,3 +1,5 @@
#define _USE_MATH_DEFINES // for M_PI
#include "common.h"
// third-party utilities
@ -13,53 +15,59 @@
#include <codecvt>
#include <sstream>
#ifndef M_PI
#define M_PI 3.14159265358979323846
#endif
#if defined(_MSC_VER)
#pragma warning(disable: 4244 4267) // possible loss of data
#endif
// Function to check if the next argument exists
std::string get_next_arg(int& i, int argc, char** argv, const std::string& flag, gpt_params& params) {
if (i + 1 < argc && argv[i + 1][0] != '-') {
return argv[++i];
} else {
fprintf(stderr, "error: %s requires one argument.\n", flag.c_str());
gpt_print_usage(argc, argv, params);
exit(0);
}
}
bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
for (int i = 1; i < argc; i++) {
std::string arg = argv[i];
if (arg == "-s" || arg == "--seed") {
params.seed = std::stoi(argv[++i]);
params.seed = std::stoi(get_next_arg(i, argc, argv, arg, params));
} else if (arg == "-t" || arg == "--threads") {
params.n_threads = std::stoi(argv[++i]);
params.n_threads = std::stoi(get_next_arg(i, argc, argv, arg, params));
} else if (arg == "-ngl" || arg == "--gpu-layers" || arg == "--n-gpu-layers") {
params.n_gpu_layers = std::stoi(get_next_arg(i, argc, argv, arg, params));
} else if (arg == "-p" || arg == "--prompt") {
params.prompt = argv[++i];
params.prompt = get_next_arg(i, argc, argv, arg, params);
} else if (arg == "-n" || arg == "--n_predict") {
params.n_predict = std::stoi(argv[++i]);
params.n_predict = std::stoi(get_next_arg(i, argc, argv, arg, params));
} else if (arg == "--top_k") {
params.top_k = std::max(1, std::stoi(argv[++i]));
params.top_k = std::stoi(get_next_arg(i, argc, argv, arg, params));
} else if (arg == "--top_p") {
params.top_p = std::stof(argv[++i]);
params.top_p = std::stof(get_next_arg(i, argc, argv, arg, params));
} else if (arg == "--temp") {
params.temp = std::stof(argv[++i]);
params.temp = std::stof(get_next_arg(i, argc, argv, arg, params));
} else if (arg == "--repeat-last-n") {
params.repeat_last_n = std::stof(argv[++i]);
params.repeat_last_n = std::stoi(get_next_arg(i, argc, argv, arg, params));
} else if (arg == "--repeat-penalty") {
params.repeat_penalty = std::stof(argv[++i]);
params.repeat_penalty = std::stof(get_next_arg(i, argc, argv, arg, params));
} else if (arg == "-b" || arg == "--batch_size") {
params.n_batch = std::stoi(argv[++i]);
params.n_batch= std::stoi(get_next_arg(i, argc, argv, arg, params));
} else if (arg == "-m" || arg == "--model") {
params.model = argv[++i];
params.model = get_next_arg(i, argc, argv, arg, params);
} else if (arg == "-i" || arg == "--interactive") {
params.interactive = true;
} else if (arg == "-ip" || arg == "--interactive-port") {
params.interactive = true;
params.interactive_port = std::stoi(argv[++i]);
params.interactive_port = std::stoi(get_next_arg(i, argc, argv, arg, params));
} else if (arg == "-h" || arg == "--help") {
gpt_print_usage(argc, argv, params);
exit(0);
} else if (arg == "-f" || arg == "--file") {
if (++i > argc) {
fprintf(stderr, "Invalid file param");
break;
}
get_next_arg(i, argc, argv, arg, params);
std::ifstream file(argv[i]);
if (!file) {
fprintf(stderr, "error: failed to open file '%s'\n", argv[i]);
@ -70,7 +78,7 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
params.prompt.pop_back();
}
} else if (arg == "-tt" || arg == "--token_test") {
params.token_test = argv[++i];
params.token_test = get_next_arg(i, argc, argv, arg, params);
}
else {
fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
@ -89,6 +97,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
fprintf(stderr, " -h, --help show this help message and exit\n");
fprintf(stderr, " -s SEED, --seed SEED RNG seed (default: -1)\n");
fprintf(stderr, " -t N, --threads N number of threads to use during computation (default: %d)\n", params.n_threads);
fprintf(stderr, " -ngl N, --gpu-layers N number of layers to offload to GPU on supported models (default: %d)\n", params.n_gpu_layers);
fprintf(stderr, " -p PROMPT, --prompt PROMPT\n");
fprintf(stderr, " prompt to start generation with (default: random)\n");
fprintf(stderr, " -f FNAME, --file FNAME\n");
@ -755,3 +764,46 @@ float similarity(const std::string & s0, const std::string & s1) {
return 1.0f - (dist / std::max(s0.size(), s1.size()));
}
bool sam_params_parse(int argc, char ** argv, sam_params & params) {
for (int i = 1; i < argc; i++) {
std::string arg = argv[i];
if (arg == "-s" || arg == "--seed") {
params.seed = std::stoi(argv[++i]);
} else if (arg == "-t" || arg == "--threads") {
params.n_threads = std::stoi(argv[++i]);
} else if (arg == "-m" || arg == "--model") {
params.model = argv[++i];
} else if (arg == "-i" || arg == "--inp") {
params.fname_inp = argv[++i];
} else if (arg == "-o" || arg == "--out") {
params.fname_out = argv[++i];
} else if (arg == "-h" || arg == "--help") {
sam_print_usage(argc, argv, params);
exit(0);
} else {
fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
sam_print_usage(argc, argv, params);
exit(0);
}
}
return true;
}
void sam_print_usage(int argc, char ** argv, const sam_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, " -s SEED, --seed SEED RNG seed (default: -1)\n");
fprintf(stderr, " -t N, --threads N number of threads to use during computation (default: %d)\n", params.n_threads);
fprintf(stderr, " -m FNAME, --model FNAME\n");
fprintf(stderr, " model path (default: %s)\n", params.model.c_str());
fprintf(stderr, " -i FNAME, --inp FNAME\n");
fprintf(stderr, " input file (default: %s)\n", params.fname_inp.c_str());
fprintf(stderr, " -o FNAME, --out FNAME\n");
fprintf(stderr, " output file (default: %s)\n", params.fname_out.c_str());
fprintf(stderr, "\n");
}

21
examples/common.h
View File

@ -11,7 +11,7 @@
#define COMMON_SAMPLE_RATE 16000
//
// CLI argument parsing
// GPT CLI argument parsing
//
struct gpt_params {
@ -33,6 +33,8 @@ struct gpt_params {
bool interactive = false;
int32_t interactive_port = -1;
int32_t n_gpu_layers = 0;
};
bool gpt_params_parse(int argc, char ** argv, gpt_params & params);
@ -155,3 +157,20 @@ bool vad_simple(
// compute similarity between two strings using Levenshtein distance
float similarity(const std::string & s0, const std::string & s1);
//
// SAM argument parsing
//
struct sam_params {
int32_t seed = -1; // RNG seed
int32_t n_threads = std::min(4, (int32_t) std::thread::hardware_concurrency());
std::string model = "models/sam-vit-b/ggml-model-f16.bin"; // model path
std::string fname_inp = "img.jpg";
std::string fname_out = "img.out";
};
bool sam_params_parse(int argc, char ** argv, sam_params & params);
void sam_print_usage(int argc, char ** argv, const sam_params & params);

17
examples/talk.wasm/gpt-2.cpp
View File

@ -191,9 +191,9 @@ bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab &
// create the ggml context
{
struct ggml_init_params params = {
.mem_size = ctx_size,
.mem_buffer = NULL,
.no_alloc = false,
/*.mem_size =*/ ctx_size,
/*.mem_buffer =*/ NULL,
/*.no_alloc =*/ false,
};
model.ctx = ggml_init(params);
@ -420,7 +420,6 @@ bool gpt2_eval(
struct ggml_context * ctx0 = ggml_init(params);
struct ggml_cgraph gf = {};
gf.n_threads = n_threads;
struct ggml_tensor * embd = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
memcpy(embd->data, embd_inp.data(), N*ggml_element_size(embd));
@ -442,7 +441,7 @@ bool gpt2_eval(
// norm
{
// [ 768, N]
cur = ggml_norm(ctx0, inpL);
cur = ggml_norm(ctx0, inpL, 1e-5f);
// cur = ln_1_g*cur + ln_1_b
// [ 768, N]
@ -589,7 +588,7 @@ bool gpt2_eval(
{
// norm
{
cur = ggml_norm(ctx0, inpFF);
cur = ggml_norm(ctx0, inpFF, 1e-5f);
// cur = ln_2_g*cur + ln_2_b
// [ 768, N]
@ -644,7 +643,7 @@ bool gpt2_eval(
// norm
{
// [ 768, N]
inpL = ggml_norm(ctx0, inpL);
inpL = ggml_norm(ctx0, inpL, 1e-5f);
// inpL = ln_f_g*inpL + ln_f_b
// [ 768, N]
@ -664,8 +663,8 @@ bool gpt2_eval(
//inpL = ggml_soft_max(ctx0, inpL);
// run the computation
ggml_build_forward_expand(&gf, inpL);
ggml_graph_compute (ctx0, &gf);
ggml_build_forward_expand (&gf, inpL);
ggml_graph_compute_with_ctx(ctx0, &gf, n_threads);
//if (n_past%100 == 0) {
// ggml_graph_print (&gf);

12
examples/talk/gpt-2.cpp
View File

@ -379,6 +379,7 @@ bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab &
// - embd_inp: the embeddings of the tokens in the context
// - embd_w: the predicted logits for the next token
//
// TODO: sync latest version from ggml repo
bool gpt2_eval(
const gpt2_model & model,
const int n_threads,
@ -420,7 +421,6 @@ bool gpt2_eval(
struct ggml_context * ctx0 = ggml_init(params);
struct ggml_cgraph gf = {};
gf.n_threads = n_threads;
struct ggml_tensor * embd = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
memcpy(embd->data, embd_inp.data(), N*ggml_element_size(embd));
@ -442,7 +442,7 @@ bool gpt2_eval(
// norm
{
// [ 768, N]
cur = ggml_norm(ctx0, inpL);
cur = ggml_norm(ctx0, inpL, 1e-5f);
// cur = ln_1_g*cur + ln_1_b
// [ 768, N]
@ -589,7 +589,7 @@ bool gpt2_eval(
{
// norm
{
cur = ggml_norm(ctx0, inpFF);
cur = ggml_norm(ctx0, inpFF, 1e-5f);
// cur = ln_2_g*cur + ln_2_b
// [ 768, N]
@ -644,7 +644,7 @@ bool gpt2_eval(
// norm
{
// [ 768, N]
inpL = ggml_norm(ctx0, inpL);
inpL = ggml_norm(ctx0, inpL, 1e-5f);
// inpL = ln_f_g*inpL + ln_f_b
// [ 768, N]
@ -664,8 +664,8 @@ bool gpt2_eval(
//inpL = ggml_soft_max(ctx0, inpL);
// run the computation
ggml_build_forward_expand(&gf, inpL);
ggml_graph_compute (ctx0, &gf);
ggml_build_forward_expand (&gf, inpL);
ggml_graph_compute_with_ctx(ctx0, &gf, n_threads);
//if (n_past%100 == 0) {
// ggml_graph_print (&gf);