Skip to content

[BugFix][Spec Decode] Improve Prefix Caching Logic in Speculative Decoding #18668

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

Merged
merged 2 commits into from
May 25, 2025
Merged

[BugFix][Spec Decode] Improve Prefix Caching Logic in Speculative Decoding #18668

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
d4dca3e
[BugFix][Spec Decode] Fix prefix caching for draft tokens
WoosukKwon May 24, 2025
f195107
fix
WoosukKwon May 24, 2025
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
3 changes: 2 additions & 1 deletion vllm/v1/core/kv_cache_manager.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -174,6 +174,7 @@ def allocate_slots(
num_new_tokens: int,
num_new_computed_tokens: int = 0,
new_computed_blocks: Optional[KVCacheBlocks] = None,
num_draft_tokens: int = 0,
num_lookahead_tokens: int = 0,
delay_cache_blocks: bool = False,
) -> Optional[KVCacheBlocks]:
Expand Down Expand Up @@ -273,7 +274,7 @@ def allocate_slots(
# generated (accepted) tokens.
self.single_type_manager.cache_blocks(
request, self.req_to_block_hashes[request.request_id],
num_computed_tokens + num_new_tokens - len(request.spec_token_ids))
num_computed_tokens + num_new_tokens - num_draft_tokens)

return KVCacheBlocks(new_blocks)

Expand Down
5 changes: 5 additions & 0 deletions vllm/v1/core/sched/scheduler.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -227,10 +227,15 @@ def schedule(self) -> SchedulerOutput:
req_index += 1
continue

num_draft_tokens = max(
num_new_tokens + request.num_computed_tokens -
request.num_tokens, 0)

while True:
new_blocks = self.kv_cache_manager.allocate_slots(
request,
num_new_tokens,
num_draft_tokens=num_draft_tokens,
num_lookahead_tokens=self.num_lookahead_tokens)
if new_blocks is None:
# The request cannot be scheduled.
Expand Down