[Attention][MLA] Make FLASHINFER_MLA the default MLA backend on Blackwell, and TRTLLM the default prefill (#32339)
Signed-off-by: Matthew Bonanni <mbonanni@redhat.com> Co-authored-by: Wentao Ye <44945378+yewentao256@users.noreply.github.com>
This commit is contained in:
co-authored by
Wentao Ye
parent
b89275d018
commit
8ebfacaa75
@@ -50,8 +50,8 @@ def _get_backend_priorities(
|
||||
if use_mla:
|
||||
if device_capability.major == 10:
|
||||
return [
|
||||
AttentionBackendEnum.CUTLASS_MLA,
|
||||
AttentionBackendEnum.FLASHINFER_MLA,
|
||||
AttentionBackendEnum.CUTLASS_MLA,
|
||||
AttentionBackendEnum.FLASH_ATTN_MLA,
|
||||
AttentionBackendEnum.FLASHMLA,
|
||||
AttentionBackendEnum.TRITON_MLA,
|
||||
@@ -183,12 +183,12 @@ class CudaPlatformBase(Platform):
|
||||
if vllm_config.attention_config.backend is None:
|
||||
# Default case
|
||||
if cls.is_device_capability_family(100) and not use_sparse:
|
||||
# Blackwell => Force CutlassMLA (unless sparse, i.e. DSv3.2).
|
||||
use_cutlass_mla = True
|
||||
# Blackwell => Force FlashInferMLA (unless sparse, i.e. DSv3.2).
|
||||
use_flashinfer_mla = True
|
||||
# Set the backend in AttentionConfig so it's used during
|
||||
# backend selection
|
||||
vllm_config.attention_config.backend = (
|
||||
AttentionBackendEnum.CUTLASS_MLA
|
||||
AttentionBackendEnum.FLASHINFER_MLA
|
||||
)
|
||||
else:
|
||||
# Not Blackwell
|
||||
|
||||
Reference in New Issue
Block a user