biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Spec Decode][V0] Fix spec decode correctness test in V0 eagle/medusa (#18175)

Browse Source 
Signed-off-by: wwl2755 <wangwenlong2755@gmail.com>
This commit is contained in:
wwl2755
2025-05-18 21:49:46 -05:00
committed by GitHub
parent d1211f8794
commit 9da1095daf
4 changed files with 21 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
vllm/sequence.py
View File

@@ -1330,6 +1330,8 @@ class HiddenStates(msgspec.Struct, array_like=True,
# may be "paused" then "resumed" later. This should only prune sequences
# which are confirmed to be aborted.
seq_ids = get_all_seq_ids(seq_group_metadata_list)
# Only keep sequence IDs that exist in self._seq_ids
seq_ids = [seq_id for seq_id in seq_ids if seq_id in self._seq_ids]
if seq_ids != self._seq_ids:
# Batch contents changed - prune removed sequences.
index = [self._seq_ids.index(seq_id) for seq_id in seq_ids]