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
[ROCm] Fix for import when building with upstream triton for gfx1100 for gpt-oss serving (#29127)
Signed-off-by: Hongxia Yang <hongxia.yang@amd.com>
Hongxia Yang committed
Nov 20, 2025 at 22:30 UTC
3f5f36da3fefbae96960f60d41ccf8ac1155515e
1 file changed
+6
-6
vllm/model_executor/layers/quantization/utils/mxfp4_utils.py
+6
-6
@@ -39,15 +39,15 @@ def _swizzle_mxfp4(quant_tensor, scale, num_warps):
39
value_layout = StridedLayout
40
scale_layout = StridedLayout
41
elif current_platform.is_rocm():
42
- from triton_kernels.tensor_details.layout import (
43
- GFX950MXScaleLayout,
44
- StridedLayout,
45
- )
46
-
42
from vllm.platforms.rocm import on_gfx950
43
44
value_layout = StridedLayout
50
- scale_layout = GFX950MXScaleLayout if on_gfx950() else StridedLayout
45
+ if on_gfx950():
46
+ from triton_kernels.tensor_details.layout import GFX950MXScaleLayout
47
+
48
+ scale_layout = GFX950MXScaleLayout
49
+ else:
50
+ scale_layout = StridedLayout
51
else:
52
value_layout, value_layout_opts = layout.make_default_matmul_mxfp4_w_layout(
53
mx_axis=1