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

[Attention] add DCP support for FLASH_ATTN_MLA backend (#24453)

Browse Source 
Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>
Co-authored-by: Matthew Bonanni <mbonanni@redhat.com>
This commit is contained in:
Lucas Wilkinson
2025-09-10 05:19:26 -04:00
committed by GitHub
parent 267c80d31f
commit 0ae43dbf8c
2 changed files with 19 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
vllm/v1/worker/gpu_model_runner.py
View File

@@ -440,6 +440,9 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
return
if self.reorder_batch_threshold is not None:
# NOTE(lucas): currently no backend supports the custom masking
# required for DCP with q_len > 1, so we assert here. Remove this
# assert once the custom mask is support is added to FA3.
if self.dcp_world_size > 1:
assert self.reorder_batch_threshold == 1, \
"DCP not support reorder_batch_threshold > 1 now."