xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-10 03:54:56 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Bugfix][Attention] Fix FlashInfer MLA block size logic (#24692)

Browse Source 
Signed-off-by: Matthew Bonanni <mbonanni001@gmail.com>
This commit is contained in:
Matthew Bonanni 2025-09-11 18:39:42 -04:00 committed by GitHub
parent 7a70a71892
commit d4fd2768ef
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 11 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

13
vllm/platforms/cuda.py
View File

@ -146,6 +146,7 @@ class CudaPlatformBase(Platform):
# required block_size.
use_flashmla = False
use_cutlass_mla = False
use_flashinfer_mla = False
if envs.VLLM_ATTENTION_BACKEND is None:
# Default case
@ -164,6 +165,8 @@ class CudaPlatformBase(Platform):
use_flashmla = (envs.VLLM_ATTENTION_BACKEND == "FLASHMLA")
use_cutlass_mla = (
envs.VLLM_ATTENTION_BACKEND == "CUTLASS_MLA")
use_flashinfer_mla = (
envs.VLLM_ATTENTION_BACKEND == "FLASHINFER_MLA")
from vllm.attention.ops.flashmla import is_flashmla_supported
if use_flashmla and is_flashmla_supported()[0] \
@ -176,6 +179,11 @@ class CudaPlatformBase(Platform):
cache_config.block_size = 128
logger.info("Forcing kv cache block size to 128 for "
"CUTLASS_MLA backend.")
if use_flashinfer_mla and cache_config.block_size not in [32, 64]:
cache_config.block_size = 64
logger.info(
"Forcing kv cache block size to 64 for FlashInferMLA "
"backend.")
# lazy import to avoid circular import
from vllm.config import CUDAGraphMode
@ -228,8 +236,9 @@ class CudaPlatformBase(Platform):
use_cutlassmla = selected_backend == _Backend.CUTLASS_MLA or (
selected_backend is None and cls.is_device_capability(100)
and block_size == 128)
use_flashinfermla = (selected_backend == _Backend.FLASHINFER_MLA
and cls.has_device_capability(100))
use_flashinfermla = selected_backend == _Backend.FLASHINFER_MLA or (
selected_backend is None and cls.is_device_capability(100)
and block_size in [32, 64])
use_flashmla = selected_backend in [
_Backend.FLASHMLA, _Backend.FLASHMLA_VLLM_V1
] or (selected_backend is None and is_flashmla_supported()[0])