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 10, 15:15 UTC
[ROCm][AITER] bugfix accuracy regression in ROCM_AITER_TRITON_MLA backend (#31816)
Signed-off-by: vllmellm <vllm.ellm@embeddedllm.com>
vllmellm committed
Jan 7, 2026 at 06:04 UTC
6409004b2656baa147a3ecc6577e5b24fc225541
1 file changed
+2
-10
vllm/v1/attention/backends/mla/aiter_triton_mla.py
+2
-10
@@ -1,13 +1,9 @@
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
-from vllm.v1.attention.backends.mla.common import MLACommonBackend
4
-from vllm.v1.attention.backends.mla.rocm_aiter_mla import (
5
- AiterMLAImpl,
6
- AiterMLAMetadataBuilder,
7
-)
3
+from vllm.v1.attention.backends.mla.rocm_aiter_mla import AiterMLABackend, AiterMLAImpl
4
5
10
-class AiterTritonMLABackend(MLACommonBackend):
6
+class AiterTritonMLABackend(AiterMLABackend):
7
@staticmethod
8
def get_name() -> str:
9
return "AITER_TRITON_MLA"
@@ -16,10 +12,6 @@ class AiterTritonMLABackend(MLACommonBackend):
12
def get_impl_cls() -> type["AiterTritonMLAImpl"]:
13
return AiterTritonMLAImpl
14
19
- @staticmethod
20
- def get_builder_cls() -> type["AiterMLAMetadataBuilder"]:
21
- return AiterMLAMetadataBuilder
22
-
15
16
class AiterTritonMLAImpl(AiterMLAImpl):
17
def __init__(