[ROCm] Small functional changes for gptoss (#25201)

Signed-off-by: jpvillam <jpvillam@amd.com>
Co-authored-by: jpvillam <jpvillam@amd.com>
This commit is contained in:
Juan Villamizar
2025-09-23 23:39:50 +00:00
committed by GitHub
co-authored by jpvillam
parent 5e25b12236
commit bde2a1a8a4
3 changed files with 26 additions and 6 deletions
@@ -1,6 +1,6 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from typing import Callable, Optional
from typing import Any, Callable, Optional
import torch
@@ -21,6 +21,10 @@ def _swizzle_mxfp4(quant_tensor, scale, num_warps):
from triton_kernels.tensor import FP4, convert_layout, wrap_torch_tensor
from triton_kernels.tensor_details import layout
from triton_kernels.tensor_details.layout import StridedLayout
value_layout_opts: dict[str, Any] = {}
scale_layout_opts: dict[str, Any] = {}
if (current_platform.is_cuda()
and current_platform.is_device_capability(90)
and not is_torch_equal_or_newer("2.8.1")):
@@ -28,8 +32,15 @@ def _swizzle_mxfp4(quant_tensor, scale, num_warps):
"Mxfp4 on hopper is running on torch < 2.8.1, "
"this cause swizling to be disabled, which may "
"cause performance degradation. Please upgrade to torch nightly")
value_layout, value_layout_opts = StridedLayout, dict()
scale_layout, scale_layout_opts = StridedLayout, dict()
value_layout = StridedLayout
scale_layout = StridedLayout
elif current_platform.is_rocm():
from triton_kernels.tensor_details.layout import (GFX950MXScaleLayout,
StridedLayout)
from vllm.platforms.rocm import on_gfx950
value_layout = StridedLayout
scale_layout = GFX950MXScaleLayout if on_gfx950() else StridedLayout
else:
value_layout, value_layout_opts = \
layout.make_default_matmul_mxfp4_w_layout(mx_axis=1)