Skip to content

server: preserve primary KV cache when MTP companion trim fails #1889

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
localweights wants to merge 1 commit into from
Closed

server: preserve primary KV cache when MTP companion trim fails #1889

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
19 changes: 17 additions & 2 deletions examples/server/server-context.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3953,8 +3953,23 @@ void server_context::batch_pending_prompt(const int32_t n_ubatch, const int32_t
auto * ctx_companion = slot.spec ? common_speculative_get_companion_ctx(slot.spec) : nullptr;
const bool target_trimmed = llama_kv_cache_seq_rm(ctx, slot.id, p0, -1);
const bool companion_trimmed = ctx_companion == nullptr || llama_kv_cache_seq_rm(ctx_companion, slot.id, p0, -1);
if (!target_trimmed || !companion_trimmed) {
// could not partially delete (likely using a non-Transformer model)

// When MTP companion ctx fails partial trim at p0 (common after
// generation due to unvalidated draft tokens leaving companion
// KV positions out of sync with primary), wipe just the
// companion entirely so it re-populates on next prefill. The
// primary cache is intact and reusable; clearing it discards
// hard-won prefill state for an issue confined to the draft
// ctx. This restores prefix-cache reuse for recurrent / MTP
// models that would otherwise force a full re-prefill on every
// request.
if (target_trimmed && ctx_companion != nullptr && !companion_trimmed) {
LLAMA_LOG_WARN("MTP companion ctx out-of-sync at p0=%d; wiping companion only, primary cache preserved\n", p0);
llama_kv_cache_seq_rm(ctx_companion, slot.id, -1, -1);
}
else if (!target_trimmed) {
// Primary trim failed (likely using a non-Transformer model
// that cannot partially delete). Full reset.
llama_kv_cache_seq_rm(ctx, slot.id, -1, -1);
if (ctx_companion != nullptr) {
llama_kv_cache_seq_rm(ctx_companion, slot.id, -1, -1);
Expand Down