[DSv4] Adding TRTLLM gen attention kernel (#43827)

This commit is contained in:
Yongye Zhu
2026-06-04 07:35:09 -07:00
committed by GitHub
parent 3e77036768
commit b5235fca2e
20 changed files with 2659 additions and 86 deletions
+10 -1
View File
@@ -75,6 +75,15 @@ class DeepseekV4SparseMLAAttentionImpl(SparseMLAAttentionImpl[FlashMLASparseMeta
"""
raise NotImplementedError
@classmethod
def init_layer_buffers(cls, layer: "DeepseekV4MLAAttention") -> None:
"""Register impl-specific buffers on the layer at construction.
No-op by default; FlashInfer overrides this to register its per-tensor
FP8 scale buffers.
"""
return None
class DeepseekV4FlashMLASparseBackend(FlashMLASparseBackend):
@staticmethod
@@ -83,7 +92,7 @@ class DeepseekV4FlashMLASparseBackend(FlashMLASparseBackend):
@staticmethod
def get_name() -> str:
return "V4_FLASHMLA_SPARSE"
return "FLASHMLA_SPARSE_DSV4"
@staticmethod
def get_impl_cls() -> type["DeepseekV4SparseMLAAttentionImpl"]: