From 7c7b2a8c97816b84c6c90ebe28e7e1a9ea334888 Mon Sep 17 00:00:00 2001
From: Hubert Lu <55214931+hubertlu-tw@users.noreply.github.com>
Date: Wed, 25 Mar 2026 16:37:26 -0700
Subject: [PATCH] [Bugfix] Lazy-import CuteDSL KDA kernel to fix AMD/ROCm
 startup crash (#21428)

---
 python/sglang/srt/layers/attention/linear/kda_backend.py | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)

diff --git a/python/sglang/srt/layers/attention/linear/kda_backend.py b/python/sglang/srt/layers/attention/linear/kda_backend.py
index a5f9ba2ae..fb27462b0 100644
--- a/python/sglang/srt/layers/attention/linear/kda_backend.py
+++ b/python/sglang/srt/layers/attention/linear/kda_backend.py
@@ -3,9 +3,6 @@ from typing import Tuple, Union
 import torch
 
 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.utils import (
     LinearAttnKernelBackend,
@@ -50,6 +47,10 @@ class KDAKernelDispatcher:
         elif decode_backend.is_cutedsl():
             if not is_cuda():
                 raise ValueError("KDA CuTe DSL backend requires CUDA")
+            from sglang.srt.layers.attention.linear.kernels.kda_cutedsl import (
+                CuteDSLKDAKernel,
+            )
+
             self.decode_kernel = CuteDSLKDAKernel()
         else:
             raise ValueError(