vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

llama : refactor sampling v2 (#9294)

Browse source 
- Add `struct llama_sampler` and `struct llama_sampler_i`
- Add `llama_sampler_` API
- Add `llama_sampler_chain_` API for chaining multiple samplers
- Remove `LLAMA_API_INTERNAL`
- Add `llama_perf_` API and remove old `llama_print_timings` and `llama_reset_timings`
This commit is contained in:
Georgi Gerganov 2024-09-07 15:16:19 +03:00 • committed by GitHub
parent 947538acb8
commit df270ef745
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
48 changed files with 3497 additions and 2914 deletions
Download patch file Download diff file Expand all files Collapse all files

18
examples/llava/llava-cli.cpp
View file

@ -40,11 +40,11 @@ static bool eval_string(struct llama_context * ctx_llama, const char* str, int n
return true;
}
static const char * sample(struct llama_sampling_context * ctx_sampling,
static const char * sample(struct gpt_sampler * smpl,
struct llama_context * ctx_llama,
int * n_past) {
const llama_token id = llama_sampling_sample(ctx_sampling, ctx_llama, NULL);
llama_sampling_accept(ctx_sampling, ctx_llama, id, true);
const llama_token id = gpt_sampler_sample(smpl, ctx_llama, -1);
gpt_sampler_accept(smpl, id, true);
static std::string ret;
if (llama_token_is_eog(llama_get_model(ctx_llama), id)) {
ret = "</s>";
@ -191,15 +191,15 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_
LOG_TEE("\n");
struct llama_sampling_context * ctx_sampling = llama_sampling_init(params->sparams);
if (!ctx_sampling) {
struct gpt_sampler * smpl = gpt_sampler_init(ctx_llava->model, params->sparams);
if (!smpl) {
fprintf(stderr, "%s: failed to initialize sampling subsystem\n", __func__);
exit(1);
}
std::string response = "";
for (int i = 0; i < max_tgt_len; i++) {
const char * tmp = sample(ctx_sampling, ctx_llava->ctx_llama, &n_past);
const char * tmp = sample(smpl, ctx_llava->ctx_llama, &n_past);
response += tmp;
if (strcmp(tmp, "</s>") == 0) break;
if (strstr(tmp, "###")) break; // Yi-VL behavior
@ -211,7 +211,7 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_
fflush(stdout);
}
llama_sampling_free(ctx_sampling);
gpt_sampler_free(smpl);
printf("\n");
}
@ -310,7 +310,7 @@ int main(int argc, char ** argv) {
// process the prompt
process_prompt(ctx_llava, image_embed, &params, params.prompt);
llama_print_timings(ctx_llava->ctx_llama);
llama_perf_print(ctx_llava->ctx_llama, LLAMA_PERF_TYPE_CONTEXT);
llava_image_embed_free(image_embed);
ctx_llava->model = NULL;
llava_free(ctx_llava);
@ -327,7 +327,7 @@ int main(int argc, char ** argv) {
// process the prompt
process_prompt(ctx_llava, image_embed, &params, params.prompt);
llama_print_timings(ctx_llava->ctx_llama);
llama_perf_print(ctx_llava->ctx_llama, LLAMA_PERF_TYPE_CONTEXT);
llava_image_embed_free(image_embed);
ctx_llava->model = NULL;
llava_free(ctx_llava);

30
examples/llava/minicpmv-cli.cpp
View file

@ -163,11 +163,11 @@ static void process_image(struct llava_context * ctx_llava, struct llava_image_e
LOG_TEE("%s: image token past: %d\n", __func__, n_past);
}
static const char * sample(struct llama_sampling_context * ctx_sampling,
static const char * sample(struct gpt_sampler * smpl,
struct llama_context * ctx_llama,
int * n_past) {
const llama_token id = llama_sampling_sample(ctx_sampling, ctx_llama, NULL);
llama_sampling_accept(ctx_sampling, ctx_llama, id, true);
const llama_token id = gpt_sampler_sample(smpl, ctx_llama, -1);
gpt_sampler_accept(smpl, id, true);
static std::string ret;
if (llama_token_is_eog(llama_get_model(ctx_llama), id)) {
ret = "</s>";
@ -214,7 +214,7 @@ static struct llava_context * minicpmv_init(gpt_params * params, const std::stri
return ctx_llava;
}
static struct llama_sampling_context * llama_init(struct llava_context * ctx_llava, gpt_params * params, std::string prompt, int &n_past, bool is_first = false){
static struct gpt_sampler * llama_init(struct llava_context * ctx_llava, gpt_params * params, std::string prompt, int &n_past, bool is_first = false){
std::string user_prompt = prompt;
int has_minicpmv_projector = clip_is_minicpmv(ctx_llava->ctx_clip);
if (!is_first) {
@ -238,13 +238,13 @@ static struct llama_sampling_context * llama_init(struct llava_context * ctx_lla
LOG_TEE("\n");
struct llama_sampling_context * ctx_sampling = llama_sampling_init(params->sparams);
return ctx_sampling;
struct gpt_sampler * smpl = gpt_sampler_init(ctx_llava->model, params->sparams);
return smpl;
}
static const char * llama_loop(struct llava_context * ctx_llava,struct llama_sampling_context * ctx_sampling, int &n_past){
static const char * llama_loop(struct llava_context * ctx_llava,struct gpt_sampler * smpl, int &n_past){
const char * tmp = sample(ctx_sampling, ctx_llava->ctx_llama, &n_past);
const char * tmp = sample(smpl, ctx_llava->ctx_llama, &n_past);
return tmp;
}
@ -278,12 +278,12 @@ int main(int argc, char ** argv) {
if (!params.prompt.empty()) {
LOG_TEE("<user>%s\n", params.prompt.c_str());
LOG_TEE("<assistant>");
auto ctx_sampling = llama_init(ctx_llava, &params, params.prompt.c_str(), n_past, true);
auto smpl = llama_init(ctx_llava, &params, params.prompt.c_str(), n_past, true);
const int max_tgt_len = params.n_predict < 0 ? 256 : params.n_predict;
std::string response = "";
bool have_tmp = false;
for (int i = 0; i < max_tgt_len; i++) {
auto tmp = llama_loop(ctx_llava, ctx_sampling, n_past);
auto tmp = llama_loop(ctx_llava, smpl, n_past);
response += tmp;
if (strcmp(tmp, "</s>") == 0){
if(!have_tmp)continue;
@ -296,18 +296,18 @@ int main(int argc, char ** argv) {
fflush(stdout);
}
llama_sampling_free(ctx_sampling);
gpt_sampler_free(smpl);
}else {
while (true) {
LOG_TEE("<user>");
std::string prompt;
std::getline(std::cin, prompt);
LOG_TEE("<assistant>");
auto ctx_sampling = llama_init(ctx_llava, &params, prompt, n_past, true);
auto smpl = llama_init(ctx_llava, &params, prompt, n_past, true);
const int max_tgt_len = params.n_predict < 0 ? 256 : params.n_predict;
std::string response = "";
for (int i = 0; i < max_tgt_len; i++) {
auto tmp = llama_loop(ctx_llava, ctx_sampling, n_past);
auto tmp = llama_loop(ctx_llava, smpl, n_past);
response += tmp;
if (strcmp(tmp, "</s>") == 0) break;
if (strstr(tmp, "###")) break; // Yi-VL behavior
@ -315,11 +315,11 @@ int main(int argc, char ** argv) {
if (strstr(response.c_str(), "<user>")) break; // minicpm-v
fflush(stdout);
}
llama_sampling_free(ctx_sampling);
gpt_sampler_free(smpl);
}
}
printf("\n");
llama_print_timings(ctx_llava->ctx_llama);
llama_perf_print(ctx_llava->ctx_llama, LLAMA_PERF_TYPE_CONTEXT);
ctx_llava->model = NULL;
llava_free(ctx_llava);