speculative : refactor and add a simpler example (#10362)
* speculative : refactor and add a simpler example ggml-ci * speculative : clean-up and add comments and TODOs [no ci] * speculative : manage context in common_speculative ggml-ci * speculative : simplify ggml-ci * speculative : simplify (cont) ggml-ci * speculative : add --draft-min CLI arg * speculative : minor fixup * make : build fixes * speculative : do not redraft previous drafts ggml-ci * speculative : fix the draft sampling ggml-ci * speculative : fix compile warning * common : refactor args ggml-ci * common : change defaults [no ci] * common : final touches ggml-ci
This commit is contained in:
Georgi Gerganov
GitHub
parent
cce5a90075
commit
d9d54e498d
28 changed files with 1028 additions and 326 deletions
5
examples/speculative-simple/CMakeLists.txt
Normal file
5
examples/speculative-simple/CMakeLists.txt
Normal file
|
|
@ -0,0 +1,5 @@
|
|||
set(TARGET llama-speculative-simple)
|
||||
add_executable(${TARGET} speculative-simple.cpp)
|
||||
install(TARGETS ${TARGET} RUNTIME)
|
||||
target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
|
||||
target_compile_features(${TARGET} PRIVATE cxx_std_11)
|
||||
12
examples/speculative-simple/README.md
Normal file
12
examples/speculative-simple/README.md
Normal file
|
|
@ -0,0 +1,12 @@
|
|||
# llama.cpp/examples/speculative-simple
|
||||
|
||||
Demonstration of basic greedy speculative decoding
|
||||
|
||||
```bash
|
||||
./bin/llama-speculative-simple \
|
||||
-m ../models/qwen2.5-32b-coder-instruct/ggml-model-q8_0.gguf \
|
||||
-md ../models/qwen2.5-1.5b-coder-instruct/ggml-model-q4_0.gguf \
|
||||
-f test.txt -c 0 -ngl 99 --color \
|
||||
--sampling-seq k --top-k 1 -fa --temp 0.0 \
|
||||
-ngld 99 --draft-max 16 --draft-min 5 --draft-p-min 0.9
|
||||
```
|
||||
273
examples/speculative-simple/speculative-simple.cpp
Normal file
273
examples/speculative-simple/speculative-simple.cpp
Normal file
|
|
@ -0,0 +1,273 @@
|
|||
#include "arg.h"
|
||||
#include "common.h"
|
||||
#include "sampling.h"
|
||||
#include "speculative.h"
|
||||
#include "log.h"
|
||||
#include "llama.h"
|
||||
|
||||
#include <cstdio>
|
||||
#include <cstring>
|
||||
#include <string>
|
||||
#include <vector>
|
||||
|
||||
int main(int argc, char ** argv) {
|
||||
common_params params;
|
||||
|
||||
if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_SPECULATIVE)) {
|
||||
return 1;
|
||||
}
|
||||
|
||||
if (params.n_predict < -1) {
|
||||
LOG_ERR("%s: --n-predict must be >= -1\n", __func__);
|
||||
return 1;
|
||||
}
|
||||
|
||||
common_init();
|
||||
|
||||
if (params.speculative.model.empty()) {
|
||||
LOG_ERR("%s: --model-draft is required\n", __func__);
|
||||
return 1;
|
||||
}
|
||||
|
||||
// init llama.cpp
|
||||
llama_backend_init();
|
||||
llama_numa_init(params.numa);
|
||||
|
||||
llama_model * model_tgt = NULL;
|
||||
llama_model * model_dft = NULL;
|
||||
|
||||
llama_context * ctx_tgt = NULL;
|
||||
llama_context * ctx_dft = NULL;
|
||||
|
||||
// load the target model
|
||||
common_init_result llama_init_tgt = common_init_from_params(params);
|
||||
|
||||
model_tgt = llama_init_tgt.model;
|
||||
ctx_tgt = llama_init_tgt.context;
|
||||
|
||||
// load the draft model
|
||||
params.model = params.speculative.model;
|
||||
params.n_ctx = params.speculative.n_ctx;
|
||||
params.n_batch = params.speculative.n_ctx > 0 ? params.speculative.n_ctx : params.n_batch;
|
||||
params.n_gpu_layers = params.speculative.n_gpu_layers;
|
||||
|
||||
if (params.speculative.cpuparams.n_threads > 0) {
|
||||
params.cpuparams.n_threads = params.speculative.cpuparams.n_threads;
|
||||
}
|
||||
|
||||
params.cpuparams_batch.n_threads = params.speculative.cpuparams_batch.n_threads;
|
||||
common_init_result llama_init_dft = common_init_from_params(params);
|
||||
|
||||
model_dft = llama_init_dft.model;
|
||||
ctx_dft = llama_init_dft.context;
|
||||
|
||||
if (!common_speculative_are_compatible(ctx_tgt, ctx_dft)) {
|
||||
return 1;
|
||||
}
|
||||
|
||||
// Tokenize the prompt
|
||||
std::vector<llama_token> inp;
|
||||
inp = common_tokenize(ctx_tgt, params.prompt, true, true);
|
||||
|
||||
if (llama_n_ctx(ctx_tgt) < (int) inp.size()) {
|
||||
LOG_ERR("%s: the prompt exceeds the context size (%d tokens, ctx %d)\n", __func__, (int) inp.size(), llama_n_ctx(ctx_tgt));
|
||||
|
||||
return 1;
|
||||
}
|
||||
|
||||
if (llama_n_batch(ctx_tgt) < (int) inp.size()) {
|
||||
LOG_ERR("%s: the prompt exceeds the batch size (%d tokens, batch %d)\n", __func__, (int) inp.size(), llama_n_batch(ctx_tgt));
|
||||
|
||||
return 1;
|
||||
}
|
||||
|
||||
LOG("\n\n");
|
||||
|
||||
for (auto id : inp) {
|
||||
LOG("%s", common_token_to_piece(ctx_tgt, id).c_str());
|
||||
}
|
||||
|
||||
// how many tokens to draft each time
|
||||
int n_draft = params.speculative.n_max;
|
||||
int n_draft_min = params.speculative.n_min;
|
||||
|
||||
float p_min = params.speculative.p_min;
|
||||
|
||||
int n_predict = 0;
|
||||
int n_drafted = 0;
|
||||
int n_accept = 0;
|
||||
|
||||
// used to determine end of generation
|
||||
bool has_eos = false;
|
||||
|
||||
// ================================================
|
||||
// everything until here is standard initialization
|
||||
// the relevant stuff for speculative decoding starts here
|
||||
|
||||
const auto t_enc_start = ggml_time_us();
|
||||
|
||||
// target model sampling context
|
||||
struct common_sampler * smpl = common_sampler_init(model_tgt, params.sampling);
|
||||
|
||||
// eval the prompt
|
||||
llama_decode(ctx_tgt, llama_batch_get_one(inp.data(), inp.size() - 1));
|
||||
|
||||
// note: keep the last token separate!
|
||||
llama_token id_last = inp.back();
|
||||
|
||||
// all tokens currently in the target context
|
||||
auto prompt_tgt = std::vector<llama_token>(inp.begin(), inp.end() - 1);
|
||||
|
||||
int n_past = inp.size() - 1;
|
||||
|
||||
// init the speculator
|
||||
struct common_speculative_params params_spec;
|
||||
params_spec.n_draft = n_draft;
|
||||
params_spec.n_reuse = llama_n_ctx(ctx_dft) - n_draft;
|
||||
params_spec.p_min = p_min;
|
||||
|
||||
struct common_speculative * spec = common_speculative_init(ctx_dft);
|
||||
|
||||
llama_batch batch_tgt = llama_batch_init(llama_n_batch(ctx_tgt), 0, 1);
|
||||
|
||||
const auto t_enc_end = ggml_time_us();
|
||||
|
||||
const auto t_dec_start = ggml_time_us();
|
||||
|
||||
while (true) {
|
||||
// optionally, generate draft tokens that can be appended to the target batch
|
||||
//
|
||||
// this is the most important part of the speculation. the more probable tokens that are provided here
|
||||
// the better the performance will be. in theory, this computation can be performed asynchronously and even
|
||||
// offloaded to a remote device. it doesn't even have to be based on an LLM. instead, it can provide tokens
|
||||
// from a cache or lookup tables.
|
||||
//
|
||||
llama_tokens draft = common_speculative_gen_draft(spec, params_spec, prompt_tgt, id_last);
|
||||
|
||||
//LOG_DBG("draft: %s\n", string_from(ctx_dft, draft).c_str());
|
||||
|
||||
// always have a token to evaluate from before - id_last
|
||||
common_batch_clear(batch_tgt);
|
||||
common_batch_add (batch_tgt, id_last, n_past++, { 0 }, true);
|
||||
|
||||
// evaluate the target model on [id_last, draft0, draft1, ..., draftN-1]
|
||||
{
|
||||
// do not waste time on small drafts
|
||||
if (draft.size() < n_draft_min) {
|
||||
draft.clear();
|
||||
}
|
||||
|
||||
for (size_t i = 0; i < draft.size(); ++i) {
|
||||
common_batch_add(batch_tgt, draft[i], n_past + i, { 0 }, true);
|
||||
}
|
||||
|
||||
//LOG_DBG("target batch: %s\n", string_from(ctx_tgt, batch_tgt).c_str());
|
||||
|
||||
llama_decode(ctx_tgt, batch_tgt);
|
||||
}
|
||||
|
||||
// sample from the full target batch and return the accepted tokens based on the target sampler
|
||||
//
|
||||
// for each token to be accepted, the sampler would have to sample that same token
|
||||
// in such cases, instead of decoding the sampled token as we normally do, we simply continue with the
|
||||
// available logits from the batch and sample the next token until we run out of logits or the sampler
|
||||
// disagrees with the draft
|
||||
//
|
||||
const auto ids = common_sampler_sample_and_accept_n(smpl, ctx_tgt, draft);
|
||||
|
||||
//LOG_DBG("ids: %s\n", string_from(ctx_tgt, ids).c_str());
|
||||
|
||||
GGML_ASSERT(ids.size() > 0); // there will always be at least one accepted token
|
||||
|
||||
n_past += ids.size() - 1;
|
||||
n_drafted += batch_tgt.n_tokens - 1;
|
||||
n_accept += ids.size() - 1;
|
||||
|
||||
// process the accepted tokens and update contexts
|
||||
//
|
||||
// this is the standard token post-processing that we normally do
|
||||
// in this case, we do it for a group of accepted tokens at once
|
||||
//
|
||||
{
|
||||
llama_token id;
|
||||
std::string token_str;
|
||||
|
||||
for (size_t i = 0; i < ids.size(); ++i) {
|
||||
id = ids[i];
|
||||
|
||||
++n_predict;
|
||||
|
||||
if (llama_token_is_eog(model_tgt, id)) {
|
||||
has_eos = true;
|
||||
break;
|
||||
}
|
||||
|
||||
token_str = common_token_to_piece(ctx_tgt, id);
|
||||
|
||||
if (params.use_color && i + 1 < ids.size()) {
|
||||
LOG("\u001b[%dm%s\u001b[37m", (36 - 0 % 6), token_str.c_str());
|
||||
} else {
|
||||
LOG("%s", token_str.c_str());
|
||||
}
|
||||
}
|
||||
|
||||
if ((params.n_predict >= 0 && n_predict > params.n_predict) || has_eos) {
|
||||
break;
|
||||
}
|
||||
|
||||
LOG_DBG("accepted %d/%d draft tokens, the last target token is: (%d, '%s')\n", (int) ids.size() - 1, (int) draft.size(), id, token_str.c_str());
|
||||
|
||||
{
|
||||
LOG_DBG("clear kv cache from any extra tokens, n_past = %d\n", n_past);
|
||||
|
||||
llama_kv_cache_seq_rm(ctx_tgt, 0, n_past, -1);
|
||||
}
|
||||
|
||||
prompt_tgt.push_back(id_last);
|
||||
prompt_tgt.insert(prompt_tgt.end(), ids.begin(), ids.end() - 1);
|
||||
|
||||
// remember the last accepted token for the next iteration
|
||||
id_last = id;
|
||||
}
|
||||
}
|
||||
|
||||
auto t_dec_end = ggml_time_us();
|
||||
|
||||
const int n_input = inp.size();
|
||||
|
||||
LOG("\n\n");
|
||||
|
||||
LOG_INF("encoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_input, (t_enc_end - t_enc_start) / 1e6f, inp.size() / ((t_enc_end - t_enc_start) / 1e6f));
|
||||
LOG_INF("decoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_predict, (t_dec_end - t_dec_start) / 1e6f, n_predict / ((t_dec_end - t_dec_start) / 1e6f));
|
||||
|
||||
LOG_INF("\n");
|
||||
LOG_INF("n_draft = %d\n", n_draft);
|
||||
LOG_INF("n_predict = %d\n", n_predict);
|
||||
LOG_INF("n_drafted = %d\n", n_drafted);
|
||||
LOG_INF("n_accept = %d\n", n_accept);
|
||||
LOG_INF("accept = %.3f%%\n", 100.0f * n_accept / n_drafted);
|
||||
|
||||
LOG_INF("\n");
|
||||
LOG_INF("draft:\n\n");
|
||||
|
||||
llama_perf_context_print(ctx_dft);
|
||||
|
||||
LOG_INF("\n");
|
||||
LOG_INF("target:\n\n");
|
||||
common_perf_print(ctx_tgt, smpl);
|
||||
|
||||
common_sampler_free(smpl);
|
||||
common_speculative_free(spec);
|
||||
|
||||
llama_free(ctx_tgt);
|
||||
llama_free_model(model_tgt);
|
||||
|
||||
llama_free(ctx_dft);
|
||||
llama_free_model(model_dft);
|
||||
|
||||
llama_backend_free();
|
||||
|
||||
LOG("\n\n");
|
||||
|
||||
return 0;
|
||||
}
|
||||
Loading…
Add table
Add a link
Reference in a new issue