[Fix] Fall back to triton MOE for GPT-OSS on Blackwell with driver >= 595 (#21780)

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

python/sglang/srt/server_args.py

@@ -42,6 +42,7 @@ from sglang.srt.utils.common import (
     get_device_name,
     get_device_sm,
     get_int_env_var,
+    get_nvidia_driver_version,
     get_quantization_config,
     human_readable_int,
     is_blackwell_supported,
@@ -1757,10 +1758,21 @@ class ServerArgs:
                     and is_triton_kernels_available()
                     and self.quantization is None
                 ):
-                    self.moe_runner_backend = "triton_kernel"
-                    logger.warning(
-                        "Detected GPT-OSS model, enabling triton_kernels MOE kernel."
-                    )
+                    # The triton_kernels package segfaults on Blackwell (B200)
+                    # with NVIDIA driver >= 595. Fall back to triton backend.
+                    if is_blackwell_supported() and get_nvidia_driver_version() >= (
+                        595,
+                    ):
+                        self.moe_runner_backend = "triton"
+                        logger.warning(
+                            "Detected GPT-OSS model on Blackwell with driver >= 595, "
+                            "using triton MOE kernel to avoid triton_kernels SIGSEGV."
+                        )
+                    else:
+                        self.moe_runner_backend = "triton_kernel"
+                        logger.warning(
+                            "Detected GPT-OSS model, enabling triton_kernels MOE kernel."
+                        )
 
             if self.moe_runner_backend == "triton_kernel":
                 assert (