[ROCm][AMD] Enable preshuffled sparse indexing for 16-token blocks (#51216)
Signed-off-by: jamesETsmith <james.smith9113@gmail.com> Co-authored-by: Andreas Karatzas <akaratza@amd.com>
This commit is contained in:
co-authored by
Andreas Karatzas
parent
83ded8d839
commit
9b0ab5dd53
@@ -141,7 +141,7 @@ class DeepseekV32IndexerBackend(AttentionBackend):
|
||||
|
||||
@staticmethod
|
||||
def get_supported_kernel_block_sizes() -> list[int | MultipleOf]:
|
||||
return [1, 64] if current_platform.is_rocm() else [64]
|
||||
return [1, MultipleOf(16)] if current_platform.is_rocm() else [64]
|
||||
|
||||
@classmethod
|
||||
def get_supported_head_sizes(cls) -> list[int]:
|
||||
|
||||
@@ -275,7 +275,7 @@ class ROCMAiterMLASparseBackend(AttentionBackend):
|
||||
|
||||
@staticmethod
|
||||
def get_supported_kernel_block_sizes() -> list[int | MultipleOf]:
|
||||
return [1, 64]
|
||||
return [1, MultipleOf(16)]
|
||||
|
||||
@staticmethod
|
||||
def get_name() -> str:
|
||||
|
||||
Reference in New Issue
Block a user