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 5, 22:15 UTC
[ROCm] Fix GPT-OSS import for triton 3.6 (#37453)
Signed-off-by: Gregory Shtrasberg <Gregory.Shtrasberg@amd.com>
Gregory Shtrasberg committed
Mar 27, 2026 at 13:00 UTC
b8665383dfda143f02378833326689cd79a1f1b3
1 file changed
+9
-2
vllm/model_executor/layers/quantization/utils/mxfp4_utils.py
+9
-2
@@ -48,9 +48,16 @@ def _swizzle_mxfp4(quant_tensor, scale, num_warps=8):
48
49
value_layout = StridedLayout
50
if on_gfx950():
51
- from triton_kernels.tensor_details.layout import GFX950MXScaleLayout
51
+ try:
52
+ # triton < 3.6
53
+ from triton_kernels.tensor_details.layout import GFX950MXScaleLayout
54
53
- scale_layout = GFX950MXScaleLayout
55
+ scale_layout = GFX950MXScaleLayout
56
+ except ImportError:
57
+ # triton >= 3.6
58
+ from triton_kernels.tensor_details.layout import CDNA4MXScaleLayout
59
+
60
+ scale_layout = CDNA4MXScaleLayout
61
else:
62
scale_layout = StridedLayout
63
else: