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

llama : llama_perf + option to disable timings during decode (#9355)

Browse source 
* llama : llama_perf + option to disable timings during decode

ggml-ci

* common : add llama_arg

* Update src/llama.cpp

Co-authored-by: Xuan Son Nguyen <thichthat@gmail.com>

* perf : separate functions in the API

ggml-ci

* perf : safer pointer handling + naming update

ggml-ci

* minor : better local var name

* perf : abort on invalid sampler pointer

ggml-ci

---------

Co-authored-by: Xuan Son Nguyen <thichthat@gmail.com>
This commit is contained in:
Georgi Gerganov 2024-09-13 09:53:38 +03:00 • committed by GitHub
parent bd35cb0ae3
commit 0abc6a2c25
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
23 changed files with 135 additions and 91 deletions
Download patch file Download diff file Expand all files Collapse all files

8
common/arg.cpp
View file

@ -720,6 +720,14 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex,
params.prompt = value;
}
));
add_opt(llama_arg(
{"--no-perf"},
format("disable internal libllama performance timings (default: %s)", params.no_perf ? "true" : "false"),
[](gpt_params & params) {
params.no_perf = true;
params.sparams.no_perf = true;
}
).set_env("LLAMA_ARG_NO_PERF"));
add_opt(llama_arg(
{"-f", "--file"}, "FNAME",
"a file containing the prompt (default: none)",

3
common/common.cpp
View file

@ -820,7 +820,7 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
}
llama_kv_cache_clear(lctx);
llama_synchronize(lctx);
llama_perf_reset(lctx, LLAMA_PERF_TYPE_CONTEXT);
llama_perf_context_reset(lctx);
}
iparams.model = model;
@ -916,6 +916,7 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param
cparams.cb_eval_user_data = params.cb_eval_user_data;
cparams.offload_kqv = !params.no_kv_offload;
cparams.flash_attn = params.flash_attn;
cparams.no_perf = params.no_perf;
cparams.type_k = kv_cache_type_from_str(params.cache_type_k);
cparams.type_v = kv_cache_type_from_str(params.cache_type_v);

2
common/common.h
View file

@ -124,6 +124,7 @@ struct gpt_sampler_params {
float mirostat_eta = 0.10f; // learning rate
bool penalize_nl = false; // consider newlines as a repeatable token
bool ignore_eos = false;
bool no_perf = false; // disable performance metrics
std::vector<enum gpt_sampler_type> samplers = {
GPT_SAMPLER_TYPE_TOP_K,
@ -246,6 +247,7 @@ struct gpt_params {
bool simple_io = false; // improves compatibility with subprocesses and limited consoles
bool cont_batching = true; // insert new sequences for decoding on-the-fly
bool flash_attn = false; // flash attention
bool no_perf = false; // disable performance metrics
bool input_prefix_bos = false; // prefix BOS to user inputs, preceding input_prefix
bool logits_all = false; // return logits for all tokens in the batch

6
common/sampling.cpp
View file

@ -142,7 +142,7 @@ std::string gpt_sampler_params::print() const {
struct gpt_sampler * gpt_sampler_init(const struct llama_model * model, const struct gpt_sampler_params & params) {
llama_sampler_chain_params lparams = llama_sampler_chain_default_params();
lparams.no_perf = false; // TODO: control via params
lparams.no_perf = params.no_perf;
auto * result = new gpt_sampler {
/* .params = */ params,
@ -257,10 +257,10 @@ void gpt_perf_print(const struct llama_context * ctx, const struct gpt_sampler *
// TODO: measure grammar performance
if (gsmpl) {
llama_perf_print(gsmpl->chain, LLAMA_PERF_TYPE_SAMPLER_CHAIN);
llama_perf_sampler_print(gsmpl->chain);
}
if (ctx) {
llama_perf_print(ctx, LLAMA_PERF_TYPE_CONTEXT);
llama_perf_context_print(ctx);
}
}