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
[Bugfix] GPT OSS Attritbute error on H100 (#25228)
Signed-off-by: Varun Sundar Rabindranath <vsundarr@redhat.com> Co-authored-by: Varun Sundar Rabindranath <vsundarr@redhat.com>
Varun Sundar Rabindranath committed
Sep 19, 2025 at 15:14 UTC
7852b82b93d189a5374d403f9735d6849ba4c64e
2 files changed
+4
-4
vllm/model_executor/layers/fused_moe/config.py
+2
-2
@@ -1,7 +1,7 @@
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
from dataclasses import dataclass
4
-from typing import TYPE_CHECKING, Optional, Union
4
+from typing import Optional, Union
5
6
import torch
7
@@ -14,7 +14,7 @@ from vllm.model_executor.layers.quantization.utils.quant_utils import (
14
from vllm.utils import cdiv, has_triton_kernels
15
from vllm.utils.flashinfer import has_flashinfer_cutlass_fused_moe
16
17
-if TYPE_CHECKING and has_triton_kernels:
17
+if has_triton_kernels():
18
from triton_kernels.matmul_ogs import PrecisionConfig
19
20
logger = init_logger(__name__)
vllm/model_executor/layers/quantization/mxfp4.py
+2
-2
@@ -638,8 +638,8 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
638
return None
639
640
if self.mxfp4_backend == Mxfp4Backend.TRITON:
641
- w1_scale = layer.w13_precision_config
642
- w2_scale = layer.w2_precision_config
641
+ w1_scale = self.w13_precision_config
642
+ w2_scale = self.w2_precision_config
643
else:
644
w1_scale = layer.w13_weight_scale
645
w2_scale = layer.w2_weight_scale