A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 05:30 UTC
[Attention] Register FLASHMLA_SPARSE (#26441)
Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>
Matthew Bonanni committed
Oct 8, 2025 at 18:28 UTC
2a03f93de9d1f8f3f52d55d32c508a2149d02eb2
2 files changed
+3
-1
vllm/attention/backends/registry.py
+2
@@ -21,6 +21,7 @@ class _Backend(enum.Enum):
21
TRITON_MLA = enum.auto()
22
CUTLASS_MLA = enum.auto()
23
FLASHMLA = enum.auto()
24
+ FLASHMLA_SPARSE = enum.auto()
25
FLASH_ATTN_MLA = enum.auto()
26
PALLAS = enum.auto()
27
IPEX = enum.auto()
@@ -43,6 +44,7 @@ BACKEND_MAP = {
44
_Backend.TRITON_MLA: "vllm.v1.attention.backends.mla.triton_mla.TritonMLABackend", # noqa: E501
45
_Backend.CUTLASS_MLA: "vllm.v1.attention.backends.mla.cutlass_mla.CutlassMLABackend", # noqa: E501
46
_Backend.FLASHMLA: "vllm.v1.attention.backends.mla.flashmla.FlashMLABackend", # noqa: E501
47
+ _Backend.FLASHMLA_SPARSE: "vllm.v1.attention.backends.mla.flashmla_sparse.FlashMLASparseBackend", # noqa: E501
48
_Backend.FLASH_ATTN_MLA: "vllm.v1.attention.backends.mla.flashattn_mla.FlashAttnMLABackend", # noqa: E501
49
_Backend.PALLAS: "vllm.v1.attention.backends.pallas.PallasAttentionBackend", # noqa: E501
50
_Backend.FLEX_ATTENTION: "vllm.v1.attention.backends.flex_attention.FlexAttentionBackend", # noqa: E501
vllm/v1/attention/backends/mla/flashmla_sparse.py
+1
-1
@@ -55,7 +55,7 @@ class FlashMLASparseBackend(AttentionBackend):
55
56
@staticmethod
57
def get_name() -> str:
58
- return "FLASHMLA_SPARSE_VLLM_V1"
58
+ return "FLASHMLA_SPARSE"
59
60
@staticmethod
61
def get_metadata_cls() -> type[AttentionMetadata]: