xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-09 04:34:59 +08:00
Code Issues Packages Projects Releases Wiki Activity

Quick fix for IMA with the Prefix Prefill kernel during graph capture (#25983)

Browse Source 
Signed-off-by: Sage Moore <sage@neuralmagic.com>
This commit is contained in:
Sage Moore 2025-10-03 04:28:22 -07:00 committed by GitHub
parent aa5053e3fe
commit 5f2cacdb1e
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/attention/backends/rocm_attn.py
View File

@ -83,6 +83,14 @@ class RocmAttentionMetadataBuilder(
# max_model_len will cause graph capture to be extremely
# slow, so here we set it to 1.
attn_metadata.seq_lens.fill_(1)
if envs.VLLM_V1_USE_PREFILL_DECODE_ATTENTION:
# Here we set the query start locs to 0. This is to
# cover up an invalid memory access in the prefix_prefil kernel
# that we run into during graph capture (#25985)
common_attn_metadata.query_start_loc.zero_()
common_attn_metadata.query_start_loc_cpu.zero_()
return attn_metadata
def build(self,