[Kernel] Resubmit PR 48666 - Gemma4 FP8 KV FA4 head dim 512 backend selection (#53175)

Signed-off-by: Jhao-Ting Chen <jhaotingc@nvidia.com>
This commit is contained in:
Jhao-Ting Chen
2026-09-23 20:29:06 -07:00
committed by GitHub
parent e78e367c6e
commit 4fb767ff0a
47 changed files with 241 additions and 103 deletions
+2 -2
View File
@@ -86,7 +86,7 @@ class DeepseekV4SparseMLABackend(AttentionBackend):
]
@staticmethod
def get_supported_kernel_block_sizes() -> list[int | MultipleOf]:
def get_supported_kernel_block_sizes(kv_cache_spec=None) -> list[int | MultipleOf]:
return [64 if current_platform.is_device_capability_family(90) else 128]
@staticmethod
@@ -255,7 +255,7 @@ class FlashMLAMegaAttnBackend(DeepseekV4FlashMLABackend):
return "FLASHMLA_MEGA_ATTN_DSV41"
@staticmethod
def get_supported_kernel_block_sizes() -> list[int | MultipleOf]:
def get_supported_kernel_block_sizes(kv_cache_spec=None) -> list[int | MultipleOf]:
return [128]
@classmethod