(bugfix): block_size check for flex attn (#41363)
Co-authored-by: Matthew Bonanni <mbonanni@redhat.com>
This commit is contained in:
co-authored by
Matthew Bonanni
parent
dd5506a157
commit
4d5c89295b
@@ -35,6 +35,7 @@ from vllm.v1.attention.backend import (
|
||||
AttentionMetadataBuilder,
|
||||
AttentionType,
|
||||
CommonAttentionMetadata,
|
||||
MultipleOf,
|
||||
)
|
||||
from vllm.v1.kv_cache_interface import AttentionSpec, EncoderOnlyAttentionSpec
|
||||
|
||||
@@ -134,6 +135,10 @@ class FlexAttentionBackend(AttentionBackend):
|
||||
def get_supported_head_sizes(cls) -> list[int]:
|
||||
return []
|
||||
|
||||
@staticmethod
|
||||
def get_supported_kernel_block_sizes() -> list[int | MultipleOf]:
|
||||
return [MultipleOf(16)]
|
||||
|
||||
|
||||
# @torch.compile(fullgraph=True, mode="reduce-overhead")
|
||||
def physical_to_logical_mapping(
|
||||
|
||||
Reference in New Issue
Block a user