[Fix][XPU/ROCm/NPU] Defer sgl_kernel.quantization import in expert_pack (#36529)

python/sglang/srt/layers/quantization/expert_pack.py

@@ -7,7 +7,6 @@ from typing import Optional
 
 import torch
 import torch.nn.functional as F
-from sgl_kernel.quantization import ggml_moe_a8_vec
 
 from sglang.kernels.ops.moe.expert_pack_mxfp4 import (
     mxfp4_matvec,
@@ -157,6 +156,10 @@ class ExpertPackMoEMethod(FusedMoEMethodBase):
         weight_type: int,
         output_size: int,
     ) -> torch.Tensor:
+        # sgl_kernel.quantization is CUDA/MUSA-only; keep the import local so
+        # this module stays importable on other devices (see gguf.py:44-70).
+        from sgl_kernel.quantization import ggml_moe_a8_vec
+
         return ggml_moe_a8_vec(
             inputs,
             weights,

python/sglang/test/runners.py

@@ -424,16 +424,25 @@ class HFRunner:
                         f"before producing output"
                     )
 
-    def terminate(self):
+    def _stop_model_proc(self):
+        # Fire-and-forget terminate() leaves the child holding the accelerator
+        # during teardown; a follow-on SRTRunner on the same device can then
+        # deadlock in driver init (observed on Intel XPU B580).
         self.model_proc.terminate()
+        self.model_proc.join(timeout=30)
+        if self.model_proc.is_alive():
+            self.model_proc.kill()
+            self.model_proc.join()
         self.in_queue = self.out_queue = None
 
+    def terminate(self):
+        self._stop_model_proc()
+
     def __enter__(self):
         return self
 
     def __exit__(self, exc_type, exc_value, traceback):
-        self.model_proc.terminate()
-        self.in_queue = self.out_queue = None
+        self._stop_model_proc()
 
     @staticmethod
     def forward_generation_raw(