2023-09-03 14:12:08 +02:00
|
|
|
#include "build-info.h"
|
|
|
|
|
|
|
|
#include "common.h"
|
|
|
|
#include "llama.h"
|
2023-09-05 07:46:17 +02:00
|
|
|
#include "grammar-parser.h"
|
2023-09-03 14:12:08 +02:00
|
|
|
|
|
|
|
#include <cmath>
|
|
|
|
#include <cstdio>
|
|
|
|
#include <string>
|
|
|
|
#include <vector>
|
|
|
|
|
|
|
|
int main(int argc, char ** argv) {
|
|
|
|
gpt_params params;
|
|
|
|
|
|
|
|
if (gpt_params_parse(argc, argv, params) == false) {
|
|
|
|
return 1;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (params.model_draft.empty()) {
|
|
|
|
fprintf(stderr, "%s: error: --model-draft is required\n", __func__);
|
|
|
|
return 1;
|
|
|
|
}
|
|
|
|
|
|
|
|
#ifndef LOG_DISABLE_LOGS
|
|
|
|
log_set_target(log_filename_generator("speculative", "log"));
|
|
|
|
LOG_TEE("Log start\n");
|
|
|
|
log_dump_cmdline(argc, argv);
|
|
|
|
#endif // LOG_DISABLE_LOGS
|
|
|
|
|
|
|
|
// init llama.cpp
|
|
|
|
llama_backend_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
|
2023-09-28 18:04:36 +02:00
|
|
|
params.logits_all = true;
|
2023-09-03 14:12:08 +02:00
|
|
|
std::tie(model_tgt, ctx_tgt) = llama_init_from_gpt_params(params);
|
|
|
|
|
|
|
|
// load the draft model
|
|
|
|
params.model = params.model_draft;
|
2023-09-13 08:50:46 +02:00
|
|
|
params.n_gpu_layers = params.n_gpu_layers_draft;
|
2023-09-03 14:12:08 +02:00
|
|
|
std::tie(model_dft, ctx_dft) = llama_init_from_gpt_params(params);
|
|
|
|
|
|
|
|
// tokenize the prompt
|
|
|
|
std::vector<llama_token> inp;
|
|
|
|
inp = ::llama_tokenize(ctx_tgt, params.prompt, true);
|
|
|
|
|
|
|
|
const int max_context_size = llama_n_ctx(ctx_tgt);
|
|
|
|
const int max_tokens_list_size = max_context_size - 4;
|
|
|
|
|
|
|
|
if ((int) inp.size() > max_tokens_list_size) {
|
|
|
|
fprintf(stderr, "%s: error: prompt too long (%d tokens, max %d)\n", __func__, (int) inp.size(), max_tokens_list_size);
|
|
|
|
return 1;
|
|
|
|
}
|
|
|
|
|
|
|
|
fprintf(stderr, "\n\n");
|
|
|
|
|
|
|
|
for (auto id : inp) {
|
|
|
|
fprintf(stderr, "%s", llama_token_to_piece(ctx_tgt, id).c_str());
|
|
|
|
}
|
|
|
|
|
|
|
|
fflush(stderr);
|
|
|
|
|
|
|
|
const int n_input = inp.size();
|
|
|
|
|
|
|
|
const auto t_enc_start = ggml_time_us();
|
|
|
|
|
|
|
|
// eval the prompt with both models
|
2023-09-28 21:42:38 +02:00
|
|
|
llama_decode(ctx_tgt, llama_batch_get_one( inp.data(), n_input - 1, 0, 0));
|
|
|
|
llama_decode(ctx_tgt, llama_batch_get_one(&inp.back(), 1, n_input - 1, 0));
|
|
|
|
llama_decode(ctx_dft, llama_batch_get_one( inp.data(), n_input, 0, 0));
|
2023-09-03 14:12:08 +02:00
|
|
|
|
|
|
|
const auto t_enc_end = ggml_time_us();
|
|
|
|
|
|
|
|
// the 2 models should have the same vocab
|
|
|
|
const int n_ctx = llama_n_ctx(ctx_tgt);
|
2023-09-28 21:42:38 +02:00
|
|
|
const int n_vocab = llama_n_vocab(model_tgt);
|
|
|
|
//GGML_ASSERT(n_vocab == llama_n_vocab(model_dft));
|
2023-09-03 14:12:08 +02:00
|
|
|
|
|
|
|
// how many tokens to draft each time
|
2023-09-14 18:14:44 +02:00
|
|
|
int n_draft = params.n_draft;
|
2023-09-03 14:12:08 +02:00
|
|
|
|
|
|
|
int n_predict = 0;
|
|
|
|
int n_drafted = 0;
|
|
|
|
int n_accept = 0;
|
|
|
|
|
|
|
|
int n_past_tgt = inp.size();
|
|
|
|
int n_past_dft = inp.size();
|
|
|
|
|
|
|
|
std::vector<llama_token> drafted;
|
|
|
|
|
|
|
|
std::vector<llama_token> last_tokens(n_ctx);
|
|
|
|
std::fill(last_tokens.begin(), last_tokens.end(), 0);
|
|
|
|
|
|
|
|
for (auto & id : inp) {
|
|
|
|
last_tokens.erase(last_tokens.begin());
|
|
|
|
last_tokens.push_back(id);
|
|
|
|
}
|
|
|
|
|
|
|
|
std::vector<llama_token_data> candidates;
|
|
|
|
candidates.reserve(n_vocab);
|
|
|
|
|
|
|
|
// used to determine end of generation
|
|
|
|
bool has_eos = false;
|
|
|
|
|
2023-09-05 07:46:17 +02:00
|
|
|
// grammar stuff
|
|
|
|
struct llama_grammar * grammar_dft = NULL;
|
|
|
|
struct llama_grammar * grammar_tgt = NULL;
|
|
|
|
|
|
|
|
grammar_parser::parse_state parsed_grammar;
|
|
|
|
|
|
|
|
// if requested - load the grammar, error checking is omitted for brevity
|
|
|
|
if (!params.grammar.empty()) {
|
|
|
|
parsed_grammar = grammar_parser::parse(params.grammar.c_str());
|
|
|
|
// will be empty (default) if there are parse errors
|
|
|
|
if (parsed_grammar.rules.empty()) {
|
|
|
|
return 1;
|
|
|
|
}
|
|
|
|
|
|
|
|
std::vector<const llama_grammar_element *> grammar_rules(parsed_grammar.c_rules());
|
|
|
|
grammar_tgt = llama_grammar_init(grammar_rules.data(), grammar_rules.size(), parsed_grammar.symbol_ids.at("root"));
|
|
|
|
}
|
|
|
|
|
2023-10-11 21:35:46 +02:00
|
|
|
llama_sampling_context ctx_sampling = llama_sampling_context_init(params, grammar_tgt);
|
|
|
|
|
2023-09-03 14:12:08 +02:00
|
|
|
const auto t_dec_start = ggml_time_us();
|
|
|
|
|
|
|
|
while (true) {
|
|
|
|
LOG("drafted: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx_dft, drafted));
|
|
|
|
|
|
|
|
int i_dft = 0;
|
2023-09-14 18:14:44 +02:00
|
|
|
|
2023-09-03 14:12:08 +02:00
|
|
|
while (true) {
|
2023-09-05 07:46:17 +02:00
|
|
|
// sample from the target model
|
2023-10-11 21:35:46 +02:00
|
|
|
llama_token id = llama_sampling_sample(ctx_tgt, NULL, ctx_sampling, last_tokens, candidates, i_dft);
|
2023-09-03 14:12:08 +02:00
|
|
|
|
2023-09-05 07:46:17 +02:00
|
|
|
// remember which tokens were sampled - used for repetition penalties during sampling
|
2023-09-03 14:12:08 +02:00
|
|
|
last_tokens.erase(last_tokens.begin());
|
|
|
|
last_tokens.push_back(id);
|
|
|
|
|
|
|
|
//LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx_tgt, last_tokens));
|
|
|
|
|
|
|
|
const std::string token_str = llama_token_to_piece(ctx_tgt, id);
|
|
|
|
printf("%s", token_str.c_str());
|
|
|
|
fflush(stdout);
|
|
|
|
|
|
|
|
if (id == llama_token_eos(ctx_tgt)) {
|
|
|
|
has_eos = true;
|
|
|
|
}
|
|
|
|
|
|
|
|
++n_predict;
|
|
|
|
|
2023-09-05 07:46:17 +02:00
|
|
|
// check if the draft matches the target
|
2023-09-03 14:12:08 +02:00
|
|
|
if (i_dft < (int) drafted.size() && id == drafted[i_dft]) {
|
2023-09-05 07:46:17 +02:00
|
|
|
LOG("the sampled target token matches the %dth drafted token (%d, '%s') - accepted\n", i_dft, id, token_str.c_str());
|
2023-09-03 14:12:08 +02:00
|
|
|
++n_accept;
|
|
|
|
++n_past_tgt;
|
|
|
|
++n_past_dft;
|
|
|
|
++i_dft;
|
|
|
|
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
|
|
|
// the drafted token was rejected or we are out of drafted tokens
|
2023-09-05 07:46:17 +02:00
|
|
|
|
|
|
|
if (i_dft < (int) drafted.size()) {
|
|
|
|
LOG("the %dth drafted token (%d, '%s') does not match the sampled target token (%d, '%s') - rejected\n",
|
|
|
|
i_dft, drafted[i_dft], llama_token_to_piece(ctx_dft, drafted[i_dft]).c_str(), id, token_str.c_str());
|
|
|
|
} else {
|
|
|
|
LOG("out of drafted tokens\n");
|
|
|
|
}
|
|
|
|
|
2023-10-03 20:04:01 +02:00
|
|
|
llama_kv_cache_seq_rm(ctx_dft, 0, n_past_dft, -1);
|
2023-09-28 21:42:38 +02:00
|
|
|
llama_decode(ctx_dft, llama_batch_get_one(&id, 1, n_past_dft, 0));
|
2023-09-03 14:12:08 +02:00
|
|
|
++n_past_dft;
|
|
|
|
|
2023-09-14 18:14:44 +02:00
|
|
|
// heuristic for n_draft
|
|
|
|
{
|
|
|
|
const int n_draft_cur = (int) drafted.size();
|
|
|
|
const bool all_accepted = i_dft == n_draft_cur;
|
|
|
|
|
|
|
|
LOG("n_draft = %d\n", n_draft);
|
|
|
|
LOG("n_draft_cur = %d\n", n_draft_cur);
|
|
|
|
LOG("i_dft = %d\n", i_dft);
|
|
|
|
LOG("all_accepted = %d\n", all_accepted);
|
|
|
|
|
|
|
|
if (all_accepted && n_draft == n_draft_cur) {
|
|
|
|
LOG(" - max drafted tokens accepted - n_draft += 8\n");
|
|
|
|
n_draft = std::min(30, n_draft + 8);
|
|
|
|
} else if (all_accepted) {
|
|
|
|
LOG(" - partially drafted tokens accepted - no change\n");
|
|
|
|
} else {
|
|
|
|
LOG(" - drafted token rejected - n_draft -= 1\n");
|
|
|
|
n_draft = std::max(2, n_draft - 1);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2023-09-03 14:12:08 +02:00
|
|
|
drafted.clear();
|
|
|
|
drafted.push_back(id);
|
|
|
|
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (n_predict > params.n_predict || has_eos) {
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
2023-09-05 07:46:17 +02:00
|
|
|
if (grammar_tgt) {
|
|
|
|
if (grammar_dft) {
|
|
|
|
llama_grammar_free(grammar_dft);
|
|
|
|
}
|
2023-10-11 21:35:46 +02:00
|
|
|
// Note: Hardcoded to sequence id 0, if this ever supports parallel generation
|
|
|
|
// that will need to change.
|
|
|
|
auto it = ctx_sampling.sequence_contexts.find(0);
|
|
|
|
GGML_ASSERT(it != ctx_sampling.sequence_contexts.end());
|
|
|
|
// This is necessary because each sequence id in sequence_contexts
|
|
|
|
// uses a copy of the original grammar.
|
|
|
|
grammar_dft = llama_grammar_copy(it->second.grammar);
|
2023-09-05 07:46:17 +02:00
|
|
|
|
|
|
|
LOG("copied target grammar to draft grammar\n");
|
|
|
|
}
|
|
|
|
|
|
|
|
// sample n_draft tokens from the draft model using greedy decoding
|
2023-09-03 14:12:08 +02:00
|
|
|
int n_past_cur = n_past_dft;
|
|
|
|
for (int i = 0; i < n_draft; ++i) {
|
|
|
|
float * logits = llama_get_logits(ctx_dft);
|
|
|
|
|
|
|
|
candidates.clear();
|
|
|
|
for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
|
|
|
|
candidates.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f});
|
|
|
|
}
|
|
|
|
|
|
|
|
llama_token_data_array cur_p = { candidates.data(), candidates.size(), false };
|
|
|
|
|
2023-09-05 07:46:17 +02:00
|
|
|
if (grammar_dft != NULL) {
|
|
|
|
llama_sample_grammar(ctx_dft, &cur_p, grammar_dft);
|
|
|
|
}
|
|
|
|
|
2023-09-03 14:12:08 +02:00
|
|
|
// computes softmax and sorts the candidates
|
|
|
|
llama_sample_softmax(ctx_dft, &cur_p);
|
|
|
|
|
|
|
|
for (int i = 0; i < 3; ++i) {
|
2023-09-05 07:46:17 +02:00
|
|
|
LOG(" - draft candidate %3d: %6d (%8.3f) '%s'\n", i, cur_p.data[i].id, cur_p.data[i].p, llama_token_to_piece(ctx_dft, cur_p.data[i].id).c_str());
|
2023-09-03 14:12:08 +02:00
|
|
|
}
|
|
|
|
|
2023-09-05 07:46:17 +02:00
|
|
|
// TODO: better logic?
|
2023-09-03 14:12:08 +02:00
|
|
|
if (cur_p.data[0].p < 2*cur_p.data[1].p) {
|
2023-09-05 07:46:17 +02:00
|
|
|
LOG("stopping drafting, probability too low: %.3f < 2*%.3f\n", cur_p.data[0].p, cur_p.data[1].p);
|
2023-09-03 14:12:08 +02:00
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
2023-09-05 07:46:17 +02:00
|
|
|
// drafted token
|
|
|
|
const llama_token id = cur_p.data[0].id;
|
|
|
|
|
|
|
|
drafted.push_back(id);
|
2023-09-03 14:12:08 +02:00
|
|
|
++n_drafted;
|
|
|
|
|
2023-09-05 07:46:17 +02:00
|
|
|
// no need to evaluate the last drafted token, since we won't use the result
|
|
|
|
if (i == n_draft - 1) {
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
|
|
|
// evaluate the drafted token on the draft model
|
2023-10-03 20:04:01 +02:00
|
|
|
llama_kv_cache_seq_rm(ctx_dft, 0, n_past_cur, -1);
|
2023-09-28 21:42:38 +02:00
|
|
|
llama_decode(ctx_dft, llama_batch_get_one(&drafted.back(), 1, n_past_cur, 0));
|
2023-09-05 07:46:17 +02:00
|
|
|
++n_past_cur;
|
|
|
|
|
|
|
|
if (grammar_dft != NULL) {
|
|
|
|
llama_grammar_accept_token(ctx_dft, grammar_dft, id);
|
2023-09-03 14:12:08 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// evaluate the target model on the drafted tokens
|
2023-10-03 20:04:01 +02:00
|
|
|
llama_kv_cache_seq_rm(ctx_tgt, 0, n_past_tgt, -1);
|
2023-09-28 21:42:38 +02:00
|
|
|
llama_decode(ctx_tgt, llama_batch_get_one(drafted.data(), drafted.size(), n_past_tgt, 0));
|
2023-09-03 14:12:08 +02:00
|
|
|
++n_past_tgt;
|
|
|
|
|
2023-09-05 07:46:17 +02:00
|
|
|
// the first token is always proposed by the traget model before the speculation loop
|
2023-09-03 14:12:08 +02:00
|
|
|
drafted.erase(drafted.begin());
|
|
|
|
}
|
|
|
|
|
|
|
|
auto t_dec_end = ggml_time_us();
|
|
|
|
|
|
|
|
LOG_TEE("\n\n");
|
|
|
|
|
|
|
|
LOG_TEE("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_TEE("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));
|
|
|
|
|
|
|
|
// TODO: make sure these numbers are computed correctly
|
|
|
|
LOG_TEE("\n");
|
|
|
|
LOG_TEE("n_draft = %d\n", n_draft);
|
|
|
|
LOG_TEE("n_predict = %d\n", n_predict);
|
|
|
|
LOG_TEE("n_drafted = %d\n", n_drafted);
|
|
|
|
LOG_TEE("n_accept = %d\n", n_accept);
|
|
|
|
LOG_TEE("accept = %.3f%%\n", 100.0f * n_accept / n_drafted);
|
|
|
|
|
|
|
|
LOG_TEE("\ndraft:\n");
|
|
|
|
llama_print_timings(ctx_dft);
|
|
|
|
|
|
|
|
LOG_TEE("\ntarget:\n");
|
|
|
|
llama_print_timings(ctx_tgt);
|
|
|
|
|
|
|
|
llama_free(ctx_tgt);
|
|
|
|
llama_free_model(model_tgt);
|
|
|
|
|
|
|
|
llama_free(ctx_dft);
|
|
|
|
llama_free_model(model_dft);
|
|
|
|
|
2023-09-05 07:46:17 +02:00
|
|
|
if (grammar_dft != NULL) {
|
|
|
|
llama_grammar_free(grammar_dft);
|
|
|
|
llama_grammar_free(grammar_tgt);
|
|
|
|
}
|
2023-09-03 14:12:08 +02:00
|
|
|
llama_backend_free();
|
|
|
|
|
|
|
|
fprintf(stderr, "\n\n");
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
}
|