root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-26 19:34:35 +00:00
Code Issues Actions 4 Packages Projects Releases Wiki Activity

server : fix parallel speculative decoding (#10513)

Browse Source 
ggml-ci
This commit is contained in:
Georgi Gerganov 2024-11-26 13:36:40 +02:00 committed by GitHub
parent 811872a59d
commit 84e1c33cde
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 35 additions and 36 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

71
examples/server/server.cpp
View File

@ -2267,49 +2267,48 @@ struct server_context {
continue; // continue loop of slots continue; // continue loop of slots
} }
llama_token id; llama_token id = common_sampler_sample(slot.smpl, ctx, slot.i_batch - i);
{ slot.i_batch = -1;
completion_token_output result;
id = common_sampler_sample(slot.smpl, ctx, slot.i_batch - i); common_sampler_accept(slot.smpl, id, true);
slot.i_batch = -1; slot.n_decoded += 1;
if (slot.n_decoded == 1) {
common_sampler_accept(slot.smpl, id, true); slot.t_start_generation = ggml_time_us();
slot.t_prompt_processing = (slot.t_start_generation - slot.t_start_process_prompt) / 1e3;
slot.n_decoded += 1; metrics.on_prompt_eval(slot);
if (slot.n_decoded == 1) {
slot.t_start_generation = ggml_time_us();
slot.t_prompt_processing = (slot.t_start_generation - slot.t_start_process_prompt) / 1e3;
metrics.on_prompt_eval(slot);
}
result.tok = id;
const auto * cur_p = common_sampler_get_candidates(slot.smpl);
for (size_t i = 0; i < (size_t) slot.params.sampling.n_probs; ++i) {
result.probs.push_back({
cur_p->data[i].id,
i >= cur_p->size ? 0.0f : cur_p->data[i].p,
});
}
if (!process_token(result, slot)) {
// release slot because of stop condition
slot.release();
slot.print_timings();
send_final_response(slot);
metrics.on_prediction(slot);
continue;
}
} }
// check if the slot supports speculative decoding completion_token_output result;
if (!slot.can_speculate()) { result.tok = id;
const auto * cur_p = common_sampler_get_candidates(slot.smpl);
for (size_t i = 0; i < (size_t) slot.params.sampling.n_probs; ++i) {
result.probs.push_back({
cur_p->data[i].id,
i >= cur_p->size ? 0.0f : cur_p->data[i].p,
});
}
if (!process_token(result, slot)) {
// release slot because of stop condition
slot.release();
slot.print_timings();
send_final_response(slot);
metrics.on_prediction(slot);
continue; continue;
} }
}
// do speculative decoding
for (auto & slot : slots) {
if (!slot.is_processing() || !slot.can_speculate()) {
continue;
}
llama_token id = slot.sampled;
struct common_speculative_params params_spec; struct common_speculative_params params_spec;
params_spec.n_draft = slot.params.speculative.n_max; params_spec.n_draft = slot.params.speculative.n_max;