Skip to content

server : do not re-verify replayed draft tokens after a checkpoint restore #28061

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
JayToltTech wants to merge 1 commit into from
Closed

server : do not re-verify replayed draft tokens after a checkpoint restore #28061

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
25 changes: 19 additions & 6 deletions tools/server/server-context.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3884,12 +3884,25 @@ struct server_context_impl {
common_sampler_ptr smpl_save(common_sampler_clone(slot.smpl.get()));

GGML_ASSERT(slot.spec_i_batch.size() == n_draft + 1);
const auto & synth_probs = common_speculative_get_synth_probs(spec.get());
auto accepted = synth_probs.empty()
? common_sampler_sample_and_accept_n(slot.smpl.get(), slot.ctx_tgt, slot.spec_i_batch, slot.spec_draft)
: server_sample_and_accept_synth(
slot.smpl.get(), slot.ctx_tgt, slot.spec_i_batch, slot.spec_draft,
synth_probs, slot.spec_synth_rng, slot.spec_is_replay);
std::vector<llama_token> accepted;
if (slot.spec_is_replay) {
// replayed tokens were accepted before the restore; re-verifying them can
// disagree when logits depend on batch shape, and each disagreement restores
// the same checkpoint again - the slot stops making progress
accepted = slot.spec_draft;
for (const llama_token id : accepted) {
common_sampler_accept(slot.smpl.get(), id, true);
}
accepted.push_back(common_sampler_sample(slot.smpl.get(), slot.ctx_tgt, slot.spec_i_batch.back()));
common_sampler_accept(slot.smpl.get(), accepted.back(), true);
} else {
const auto & synth_probs = common_speculative_get_synth_probs(spec.get());
accepted = synth_probs.empty()
? common_sampler_sample_and_accept_n(slot.smpl.get(), slot.ctx_tgt, slot.spec_i_batch, slot.spec_draft)
: server_sample_and_accept_synth(
slot.smpl.get(), slot.ctx_tgt, slot.spec_i_batch, slot.spec_draft,
synth_probs, slot.spec_synth_rng, slot.spec_is_replay);
}
slot.spec_i_batch.clear();

GGML_ASSERT(accepted.size() >= 1);
Expand Down