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 12, 19:45 UTC
[Attention] Add ROCM_AITER_MLA_SPARSE to attention backend registry (#29103)
Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>
Matthew Bonanni committed
Nov 20, 2025 at 23:24 UTC
11857a00b0a59183286eab393df4e13b20efec3a
2 files changed
+4
-4
vllm/attention/backends/registry.py
+3
@@ -52,6 +52,9 @@ class AttentionBackendEnum(Enum, metaclass=_AttentionBackendEnumMeta):
52
ROCM_AITER_FA = (
53
"vllm.v1.attention.backends.rocm_aiter_fa.AiterFlashAttentionBackend"
54
)
55
+ ROCM_AITER_MLA_SPARSE = (
56
+ "vllm.v1.attention.backends.mla.rocm_aiter_mla_sparse.ROCMAiterMLASparseBackend"
57
+ )
58
TORCH_SDPA = "" # this tag is only used for ViT
59
FLASHINFER = "vllm.v1.attention.backends.flashinfer.FlashInferBackend"
60
FLASHINFER_MLA = (
vllm/platforms/rocm.py
+1
-4
@@ -233,10 +233,7 @@ class RocmPlatform(Platform):
233
"Sparse MLA backend on ROCm only supports block size 1 for now."
234
)
235
logger.info_once("Using Sparse MLA backend on V1 engine.")
236
- return (
237
- "vllm.v1.attention.backends.mla.rocm_aiter_mla_sparse."
238
- "ROCMAiterMLASparseBackend"
239
- )
236
+ return AttentionBackendEnum.ROCM_AITER_MLA_SPARSE.get_path()
237
238
if use_mla:
239
if selected_backend is None: