minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[Bugfix] Lazy-import CuteDSL KDA kernel to fix AMD/ROCm startup crash (#21428)

Browse Source
This commit is contained in:
Hubert Lu
2026-03-25 16:37:26 -07:00
committed by GitHub
parent 75682f1d2f
commit 7c7b2a8c97
1 changed files with 4 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/layers/attention/linear/kda_backend.py
+4 -3
View File
@@ -3,9 +3,6 @@ from typing import Tuple, Union
import torch import torch
from sglang.srt.layers.attention.hybrid_linear_attn_backend import MambaAttnBackendBase from sglang.srt.layers.attention.hybrid_linear_attn_backend import MambaAttnBackendBase
from sglang.srt.layers.attention.linear.kernels.kda_cutedsl import (
CuteDSLKDAKernel,
)
from sglang.srt.layers.attention.linear.kernels.kda_triton import TritonKDAKernel from sglang.srt.layers.attention.linear.kernels.kda_triton import TritonKDAKernel
from sglang.srt.layers.attention.linear.utils import ( from sglang.srt.layers.attention.linear.utils import (
LinearAttnKernelBackend, LinearAttnKernelBackend,
@@ -50,6 +47,10 @@ class KDAKernelDispatcher:
elif decode_backend.is_cutedsl(): elif decode_backend.is_cutedsl():
if not is_cuda(): if not is_cuda():
raise ValueError("KDA CuTe DSL backend requires CUDA") raise ValueError("KDA CuTe DSL backend requires CUDA")
from sglang.srt.layers.attention.linear.kernels.kda_cutedsl import (
CuteDSLKDAKernel,
)
self.decode_kernel = CuteDSLKDAKernel() self.decode_kernel = CuteDSLKDAKernel()
else: else:
raise ValueError( raise ValueError(