xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-10 10:40:44 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Misc] Simplify the prefix caching logic on draft tokens (#20701)

Browse Source 
Signed-off-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>
This commit is contained in:
Woosuk Kwon 2025-07-09 14:48:35 -07:00 committed by GitHub
parent cd587c93ef
commit 7c12a765aa
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 10 additions and 11 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

16
vllm/v1/core/kv_cache_manager.py
View File

@ -190,7 +190,6 @@ class KVCacheManager:
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]:
@ -286,12 +285,17 @@ class KVCacheManager:
if not self.enable_caching or delay_cache_blocks:
return KVCacheBlocks(new_blocks)
# Speculated tokens might be rejected in the future, so we does
# not cache any speculated tokens. We only cache blocks with
# generated (accepted) tokens.
# NOTE(woosuk): We want to commit (cache) up to num_computed_tokens +
# num_new_tokens, but must exclude "non-committable" tokens (e.g.,
# draft tokens that could be rejected). Therefore, we cap the number
# at `request.num_tokens`, ensuring only "finalized" tokens are cached.
num_tokens_to_cache = min(num_computed_tokens + num_new_tokens,
request.num_tokens)
self.coordinator.cache_blocks(
request, self.req_to_block_hashes[request.request_id],
num_computed_tokens + num_new_tokens - num_draft_tokens)
request,
self.req_to_block_hashes[request.request_id],
num_tokens_to_cache,
)
return KVCacheBlocks(new_blocks)

5
vllm/v1/core/sched/scheduler.py
View File

@ -241,15 +241,10 @@ class Scheduler(SchedulerInterface):
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.