Skip to content

speculative : auto-enable swa_full for SWA draft models #22660

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

Open
Ziqiao-git wants to merge 1 commit into
base: master
Choose a base branch
from
+16 −0
Open

speculative : auto-enable swa_full for SWA draft models #22660

Show file tree
Hide file tree
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
9 changes: 9 additions & 0 deletions common/speculative.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -987,6 +987,15 @@ common_speculative * common_speculative_init(
llama_context * ctx_tgt) {
llama_context * ctx_dft = nullptr;
if (params.draft.model) {
// for SWA draft models, force swa_full on the draft context so prefix
// reuse (seq_rm + seq_add) works beyond the SWA window — otherwise the
// draft must re-decode from the window edge on every long-context
// request and acceptance length degrades significantly
if (llama_model_n_swa(params.draft.model) > 0 && !params.draft.cparams.swa_full) {
LOG_INF("%s: draft model uses SWA — enabling swa_full for the draft context\n", __func__);
params.draft.cparams.swa_full = true;
}

ctx_dft = llama_init_from_model(params.draft.model, params.draft.cparams);
if (ctx_dft == nullptr) {
LOG_ERR("%s", "failed to create draft context\n");
Expand Down
7 changes: 7 additions & 0 deletions tools/server/server-context.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -797,6 +797,13 @@ struct server_context_impl {
return false;
}

// for SWA draft models, force swa_full on the draft context so prefix
// reuse works beyond the SWA window during speculation
if (llama_model_n_swa(model_dft.get()) > 0 && !params_dft.swa_full) {
SRV_INF("%s", "draft model uses SWA — enabling swa_full for the draft context\n");
params_dft.swa_full = true;
}

params_base.speculative.draft.model = model_dft.get();
params_base.speculative.draft.cparams = common_context_params_to_llama(params_dft);
}
Expand Down