xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-04-11 17:27:03 +08:00
Code Issues Packages Projects Releases Wiki Activity

Sage Moore fixes for full cuda graph support for DeepEP+DeepGEMM LL

Browse Source 
Signed-off-by: Tyler Michael Smith <tyler@neuralmagic.com>
This commit is contained in:
Tyler Michael Smith 2025-06-24 11:21:52 -04:00
parent 26d34eb67e
commit e53382cc2e
1 changed files with 2 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
vllm/v1/attention/backends/mla/flashmla.py
View File

@ -75,7 +75,8 @@ class FlashMLAMetadataBuilder(MLACommonMetadataBuilder[FlashMLAMetadata]):
1, # MQA for the decode path
)
if self.runner.full_cuda_graph:
n = num_splits.size(0)
if self.runner.full_cuda_graph and (n-1) <= self.runner.cudagraph_batch_sizes[-1]:
# First time around (CUDAGraph capture), allocate the static buffer
if self.cg_buf_tile_scheduler_metadata is None:
self.cg_buf_tile_scheduler_metadata = tile_scheduler_metadata