xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-01-25 18:24:33 +08:00
Code Issues Packages Projects Releases Wiki Activity

bugfix: set reorder_batch_threshold back to 1 when using FlashMLA and enable DCP

Browse Source 
Signed-off-by: FENP <32334296+FENP@users.noreply.github.com>
This commit is contained in:
FENP 2025-10-16 21:55:26 +08:00
parent d2740fafbf
commit a2d5ef088a
1 changed files with 13 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

13
vllm/v1/attention/backends/mla/common.py
View File

@ -558,6 +558,19 @@ class MLACommonMetadataBuilder(AttentionMetadataBuilder[M]):
self.dcp_world_size = 1
self.dcp_rank = 0
if (
self.dcp_world_size > 1
and self.__class__.reorder_batch_threshold > 1
and self.__class__.__name__ != "FlashAttnMLAMetadataBuilder"
):
logger.warning_once(
"DCP is enabled but not FlashAttnMLA is used. "
"Set query_len_support back to SINGLE_ONLY "
"and reorder_batch_threshold back to 1."
)
self.__class__.query_len_support = QueryLenSupport.SINGLE_ONLY
self.__class__.reorder_batch_threshold = 1
# Don't try to access the runner on AMD
if self.aot_schedule:
self.page_size = self.kv_cache_spec.block_size