523 lines
17 KiB
C++
523 lines
17 KiB
C++
#include "build-info.h"
|
|
|
|
#include "common.h"
|
|
#include "llama.h"
|
|
#include "grammar-parser.h"
|
|
|
|
#include <cmath>
|
|
#include <cstdio>
|
|
#include <string>
|
|
#include <vector>
|
|
|
|
struct seq_draft {
|
|
bool active = false;
|
|
bool drafting = false;
|
|
bool skip = false;
|
|
|
|
int i_batch_dft = 0;
|
|
std::vector<int> i_batch_tgt;
|
|
|
|
std::vector<llama_token> tokens;
|
|
|
|
struct llama_grammar * grammar = NULL;
|
|
|
|
std::vector<llama_token> last_tokens;
|
|
struct llama_sampling_context ctx_sampling;
|
|
};
|
|
|
|
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;
|
|
}
|
|
|
|
// max number of parallel drafting sequences (i.e. tree branches)
|
|
int n_seq_dft = 8;
|
|
|
|
#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
|
|
params.logits_all = true;
|
|
std::tie(model_tgt, ctx_tgt) = llama_init_from_gpt_params(params);
|
|
|
|
// load the draft model
|
|
params.model = params.model_draft;
|
|
params.n_gpu_layers = params.n_gpu_layers_draft;
|
|
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
|
|
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));
|
|
|
|
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);
|
|
const int n_vocab = llama_n_vocab(model_tgt);
|
|
//GGML_ASSERT(n_vocab == llama_n_vocab(model_dft));
|
|
|
|
// how many tokens to draft each time
|
|
int n_draft = params.n_draft;
|
|
|
|
int n_predict = 0;
|
|
int n_drafted = 0;
|
|
int n_accept = 0;
|
|
|
|
int n_past_tgt = inp.size();
|
|
int n_past_dft = inp.size();
|
|
|
|
// used to determine end of generation
|
|
bool has_eos = false;
|
|
|
|
// grammar stuff
|
|
struct llama_grammar * grammar = 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 = llama_grammar_init(grammar_rules.data(), grammar_rules.size(), parsed_grammar.symbol_ids.at("root"));
|
|
}
|
|
|
|
// target model sampling context
|
|
llama_sampling_context ctx_sampling = llama_sampling_context_init(params, grammar);
|
|
|
|
// TODO: move to llama_sampling_state
|
|
std::vector<llama_token_data> candidates;
|
|
candidates.reserve(n_vocab);
|
|
|
|
std::vector<llama_token> last_tokens;
|
|
last_tokens.resize(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);
|
|
}
|
|
|
|
// draft sequence data
|
|
std::vector<seq_draft> drafts(n_seq_dft);
|
|
|
|
for (int i = 0; i < n_seq_dft; ++i) {
|
|
{
|
|
auto & last_tokens = drafts[i].last_tokens;
|
|
|
|
last_tokens.resize(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);
|
|
}
|
|
}
|
|
|
|
drafts[i].ctx_sampling = llama_sampling_context_init(params, grammar);
|
|
}
|
|
|
|
llama_batch batch_dft = llama_batch_init(512, 0, 1);
|
|
llama_batch batch_tgt = llama_batch_init(512, 0, n_seq_dft);
|
|
|
|
const auto t_dec_start = ggml_time_us();
|
|
|
|
drafts[0].i_batch_tgt.resize(1);
|
|
drafts[0].i_batch_tgt[0] = 0;
|
|
|
|
while (true) {
|
|
for (int i = 0; i < n_seq_dft; ++i) {
|
|
if (!drafts[i].active) continue;
|
|
|
|
const auto & tokens = drafts[i].tokens;
|
|
|
|
LOG("draft %d: %s\n", i, LOG_TOKENS_TOSTR_PRETTY(ctx_dft, tokens));
|
|
}
|
|
|
|
int i_dft = 0;
|
|
int i_keep = 0;
|
|
|
|
while (true) {
|
|
LOG("sampling target: i_keep = %3d, i_dft = %3d, i_batch_tgt = %3d\n", i_keep, i_dft, drafts[i_keep].i_batch_tgt[i_dft]);
|
|
|
|
// sample from the target model
|
|
llama_token id = llama_sampling_sample(ctx_tgt, NULL, ctx_sampling, last_tokens, candidates, drafts[i_keep].i_batch_tgt[i_dft]);
|
|
|
|
// remember which tokens were sampled - used for repetition penalties during sampling
|
|
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;
|
|
|
|
// check if the target token matches any of the drafts
|
|
{
|
|
bool matches = false;
|
|
|
|
for (int i = 0; i < n_seq_dft; ++i) {
|
|
if (!drafts[i].active) continue;
|
|
|
|
if (i_dft < (int) drafts[i].tokens.size() && id == drafts[i].tokens[i_dft]) {
|
|
LOG("the sampled target token matches the %dth drafted token of sequence %d (%d, '%s') - accepted\n", i_dft, i, id, token_str.c_str());
|
|
|
|
i_keep = i;
|
|
matches = true;
|
|
} else {
|
|
drafts[i].active = false;
|
|
}
|
|
}
|
|
|
|
if (matches) {
|
|
++n_accept;
|
|
++n_past_tgt;
|
|
++n_past_dft;
|
|
++i_dft;
|
|
|
|
continue;
|
|
}
|
|
}
|
|
|
|
LOG("the sampled target token (%d, '%s') did not match, or we ran out of drafted tokens\n", id, token_str.c_str());
|
|
|
|
// TODO: simplify
|
|
{
|
|
LOG("keeping sequence %d\n", i_keep);
|
|
|
|
llama_kv_cache_seq_keep(ctx_dft, i_keep);
|
|
llama_kv_cache_seq_cp (ctx_dft, i_keep, 0, -1, -1);
|
|
llama_kv_cache_seq_keep(ctx_dft, 0);
|
|
|
|
llama_kv_cache_seq_rm (ctx_tgt, i_keep, n_past_tgt, -1);
|
|
llama_kv_cache_seq_keep(ctx_tgt, i_keep);
|
|
llama_kv_cache_seq_cp (ctx_tgt, i_keep, 0, -1, -1);
|
|
llama_kv_cache_seq_keep(ctx_tgt, 0);
|
|
}
|
|
|
|
for (int i = 0; i < n_seq_dft; ++i) {
|
|
drafts[i].active = false;
|
|
drafts[i].tokens.clear();
|
|
drafts[i].i_batch_tgt.clear();
|
|
}
|
|
// note: will be erased after the speculation phase
|
|
drafts[0].tokens.push_back(id);
|
|
drafts[0].i_batch_tgt.push_back(0);
|
|
|
|
{
|
|
batch_dft.n_tokens = 1;
|
|
|
|
batch_dft.token[0] = id;
|
|
batch_dft.pos[0] = n_past_dft;
|
|
batch_dft.n_seq_id[0] = 1;
|
|
batch_dft.seq_id[0][0] = 0;
|
|
batch_dft.logits[0] = true;
|
|
}
|
|
|
|
llama_kv_cache_seq_rm(ctx_dft, 0, n_past_dft, -1);
|
|
llama_decode(ctx_dft, batch_dft);
|
|
++n_past_dft;
|
|
|
|
break;
|
|
}
|
|
|
|
if (n_predict > params.n_predict || has_eos) {
|
|
break;
|
|
}
|
|
|
|
if (grammar) {
|
|
for (int i = 0; i < n_seq_dft; ++i) {
|
|
auto * grammar_dft = drafts[i].grammar;
|
|
if (grammar_dft) {
|
|
llama_grammar_free(grammar_dft);
|
|
}
|
|
|
|
grammar_dft = llama_grammar_copy(ctx_sampling.grammar);
|
|
|
|
LOG("copied target grammar to draft %d grammar\n", i);
|
|
}
|
|
}
|
|
|
|
int n_seq_cur = 1;
|
|
int n_past_cur = n_past_dft;
|
|
|
|
for (int i = 0; i < n_seq_dft; ++i) {
|
|
drafts[i].active = false;
|
|
drafts[i].drafting = false;
|
|
}
|
|
drafts[0].active = true;
|
|
drafts[0].drafting = true;
|
|
drafts[0].i_batch_dft = 0;
|
|
|
|
batch_tgt.n_tokens = 1;
|
|
batch_tgt.token[0] = drafts[0].tokens[0];
|
|
batch_tgt.pos[0] = n_past_tgt;
|
|
batch_tgt.n_seq_id[0] = 1;
|
|
batch_tgt.seq_id[0][0] = 0;
|
|
batch_tgt.logits[0] = true;
|
|
|
|
// sample n_draft tokens from the draft model using tree-based sampling
|
|
for (int i = 0; i < n_draft; ++i) {
|
|
batch_dft.n_tokens = 0;
|
|
|
|
for (int s = 0; s < n_seq_dft; ++s) {
|
|
drafts[s].skip = false;
|
|
}
|
|
|
|
for (int s = 0; s < n_seq_dft; ++s) {
|
|
if (!drafts[s].drafting || drafts[s].skip) continue;
|
|
|
|
auto & grammar = drafts[s].grammar;
|
|
auto & i_batch_dft = drafts[s].i_batch_dft;
|
|
|
|
float * logits = llama_get_logits_ith(ctx_dft, i_batch_dft);
|
|
|
|
// TODO: optimize
|
|
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 };
|
|
|
|
if (grammar != NULL) {
|
|
llama_sample_grammar(ctx_dft, &cur_p, grammar);
|
|
}
|
|
|
|
// computes softmax and sorts the candidates
|
|
llama_sample_softmax(ctx_dft, &cur_p);
|
|
|
|
for (int k = 0; k < 3; ++k) {
|
|
LOG(" - draft candidate %3d for seq %3d, pos %3d: %6d (%8.3f) '%s'\n",
|
|
k, s, i, cur_p.data[k].id, cur_p.data[k].p, llama_token_to_piece(ctx_dft, cur_p.data[k].id).c_str());
|
|
}
|
|
|
|
// TODO: make this configurable
|
|
if (cur_p.data[0].p < 0.1) {
|
|
//if (cur_p.data[0].p < 2*cur_p.data[1].p) {
|
|
LOG("stopping drafting for seq %3d, probability too low: %.3f < 2*%.3f\n", s, cur_p.data[0].p, cur_p.data[1].p);
|
|
drafts[s].drafting = false;
|
|
continue;
|
|
}
|
|
|
|
std::vector<int> sa(1, s);
|
|
|
|
for (int f = 1; f < 8; ++f) {
|
|
// TODO: make this configurable
|
|
if (n_seq_cur < n_seq_dft && cur_p.data[f].p > 0.10) {
|
|
LOG("splitting seq %3d into %3d\n", s, n_seq_cur);
|
|
|
|
llama_kv_cache_seq_rm(ctx_dft, n_seq_cur, -1, -1);
|
|
llama_kv_cache_seq_cp(ctx_dft, s, n_seq_cur, -1, -1);
|
|
|
|
for (int t = 0; t < batch_tgt.n_tokens; ++t) {
|
|
for (int p = 0; p < batch_tgt.n_seq_id[t]; ++p) {
|
|
if (batch_tgt.seq_id[t][p] == s) {
|
|
batch_tgt.seq_id[t][batch_tgt.n_seq_id[t]] = n_seq_cur;
|
|
batch_tgt.n_seq_id[t]++;
|
|
break;
|
|
}
|
|
}
|
|
}
|
|
|
|
drafts[n_seq_cur] = drafts[s];
|
|
drafts[n_seq_cur].skip = true;
|
|
// TODO: grammar
|
|
|
|
sa.push_back(n_seq_cur);
|
|
n_seq_cur++;
|
|
} else {
|
|
break;
|
|
}
|
|
}
|
|
|
|
// add drafted token for each sequence
|
|
for (int is = 0; is < (int) sa.size(); ++is) {
|
|
const llama_token id = cur_p.data[is].id;
|
|
|
|
int s = sa[is];
|
|
|
|
auto & drafted = drafts[s].tokens;
|
|
//auto & grammar = drafts[s].grammar;
|
|
|
|
auto & i_batch_dft = drafts[s].i_batch_dft;
|
|
auto & i_batch_tgt = drafts[s].i_batch_tgt;
|
|
|
|
drafted.push_back(id);
|
|
|
|
// add unique drafted tokens to the target batch
|
|
batch_tgt.token [batch_tgt.n_tokens] = id;
|
|
batch_tgt.pos [batch_tgt.n_tokens] = n_past_tgt + i + 1;
|
|
batch_tgt.n_seq_id[batch_tgt.n_tokens] = 1;
|
|
batch_tgt.seq_id [batch_tgt.n_tokens][0] = s;
|
|
batch_tgt.logits [batch_tgt.n_tokens] = true;
|
|
|
|
i_batch_tgt.push_back(batch_tgt.n_tokens);
|
|
|
|
batch_tgt.n_tokens++;
|
|
|
|
// no need to evaluate the last drafted token, since we won't use the result
|
|
if (i == n_draft - 1) {
|
|
drafts[s].drafting = false;
|
|
continue;
|
|
}
|
|
|
|
// add the token to the batch for batched decoding with the draft model
|
|
batch_dft.token [batch_dft.n_tokens] = id;
|
|
batch_dft.pos [batch_dft.n_tokens] = n_past_cur;
|
|
batch_dft.n_seq_id[batch_dft.n_tokens] = 1;
|
|
batch_dft.seq_id [batch_dft.n_tokens][0] = s;
|
|
batch_dft.logits [batch_dft.n_tokens] = true;
|
|
|
|
i_batch_dft = batch_dft.n_tokens;
|
|
|
|
batch_dft.n_tokens++;
|
|
}
|
|
}
|
|
|
|
// no sequence is drafting anymore
|
|
if (batch_dft.n_tokens == 0) {
|
|
break;
|
|
}
|
|
|
|
// evaluate the drafted tokens on the draft model
|
|
llama_decode(ctx_dft, batch_dft);
|
|
++n_past_cur;
|
|
++n_drafted;
|
|
|
|
// update grammar
|
|
for (int s = 0; s < n_seq_dft; ++s) {
|
|
if (!drafts[s].drafting) continue;
|
|
|
|
auto & drafted = drafts[s].tokens;
|
|
auto & grammar = drafts[s].grammar;
|
|
|
|
if (grammar != NULL) {
|
|
llama_grammar_accept_token(ctx_dft, grammar, drafted.back());
|
|
}
|
|
}
|
|
|
|
// !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
|
|
if (batch_tgt.n_tokens >= n_draft) {
|
|
break;
|
|
}
|
|
}
|
|
|
|
// evaluate the target model on the drafted tokens
|
|
{
|
|
llama_kv_cache_seq_keep(ctx_tgt, 0);
|
|
for (int s = 1; s < n_seq_dft; ++s) {
|
|
llama_kv_cache_seq_cp(ctx_tgt, 0, s, -1, -1);
|
|
}
|
|
|
|
//LOG("target batch: %s\n", LOG_BATCH_TOSTR_PRETTY(ctx_tgt, batch_tgt));
|
|
llama_decode(ctx_tgt, batch_tgt);
|
|
++n_past_tgt;
|
|
}
|
|
|
|
// the first token is always proposed by the traget model before the speculation loop so we erase it here
|
|
for (int i = 0; i < n_seq_dft; ++i) {
|
|
if (!drafts[i].active) continue;
|
|
|
|
drafts[i].tokens.erase(drafts[i].tokens.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));
|
|
|
|
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_batch_free(batch_dft);
|
|
|
|
llama_free(ctx_tgt);
|
|
llama_free_model(model_tgt);
|
|
|
|
llama_free(ctx_dft);
|
|
llama_free_model(model_dft);
|
|
|
|
if (grammar) {
|
|
llama_grammar_free(grammar);
|
|
|
|
for (int i = 0; i < n_seq_dft; ++i) {
|
|
llama_grammar_free(drafts[i].grammar);
|
|
}
|
|
}
|
|
llama_backend_free();
|
|
|
|
fprintf(stderr, "\n\n");
|
|
|
|
return 0;
|
|
}
|