xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-01-20 17:44:27 +08:00
Code Issues Packages Projects Releases Wiki Activity

[BugFix] Fix mixed penalties batch with async scheduling (#27910)

Browse Source 
Signed-off-by: Nick Hill <nhill@redhat.com>
This commit is contained in:
Nick Hill 2025-11-01 10:51:24 -07:00 committed by GitHub
parent af6e19f50f
commit c2ed069b32
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 8 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
vllm/v1/sample/ops/penalties.py
View File

@ -21,6 +21,14 @@ def apply_all_penalties(
"""
_, vocab_size = logits.shape
output_tokens_t = _convert_to_tensors(output_token_ids, vocab_size, logits.device)
# In the async scheduling case, rows that won't have penalties applied may contain
# -1 placeholder token ids. We must replace these with valid token ids so that the
# scatter done in apply_penalties is valid.
# NOTE(nick): The penalties implementation is currently quite inefficient and
# will be reworked anyhow.
output_tokens_t.masked_fill_(output_tokens_t == -1, vocab_size)
return apply_penalties(
logits,
prompt_token_ids,