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] Lazy import gpt_oss_triton_kernels_moe for mxfp4 (#23678)
Signed-off-by: mgoin <mgoin64@gmail.com>
Michael Goin committed
Aug 26, 2025 at 21:34 UTC
de02b07db4741cc9ed40b8262d7a67e6bce30211
1 file changed
+2
-2
vllm/model_executor/layers/quantization/mxfp4.py
+2
-2
@@ -10,8 +10,6 @@ from vllm.config import get_current_vllm_config
10
from vllm.logger import init_logger
11
from vllm.model_executor.layers.fused_moe import (FusedMoE, FusedMoEConfig,
12
FusedMoEMethodBase)
13
-from vllm.model_executor.layers.fused_moe.gpt_oss_triton_kernels_moe import (
14
- triton_kernel_moe_forward)
13
from vllm.model_executor.layers.linear import (LinearBase,
14
UnquantizedLinearMethod)
15
from vllm.model_executor.layers.quantization import QuantizationMethods
@@ -557,6 +555,8 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
555
)[0]
556
return trtllm_gen_output
557
else:
558
+ from vllm.model_executor.layers.fused_moe.gpt_oss_triton_kernels_moe import ( # noqa: E501
559
+ triton_kernel_moe_forward)
560
return triton_kernel_moe_forward(
561
hidden_states=x,
562
w1=self.w13_weight_triton_tensor,