diff --git a/python/sglang/check_env.py b/python/sglang/check_env.py
index 516e90237..a9a8a62b2 100644
--- a/python/sglang/check_env.py
+++ b/python/sglang/check_env.py
@@ -194,22 +194,12 @@ class GPUEnv(BaseEnv):
         """
         Get CUDA driver version.
         """
-        versions = set()
-        try:
-            output = subprocess.check_output(
-                [
-                    "nvidia-smi",
-                    "--query-gpu=driver_version",
-                    "--format=csv,noheader,nounits",
-                ]
-            )
-            versions = set(output.decode().strip().split("\n"))
-            if len(versions) == 1:
-                return {"CUDA Driver Version": versions.pop()}
-            else:
-                return {"CUDA Driver Versions": ", ".join(sorted(versions))}
-        except subprocess.SubprocessError:
+        from sglang.srt.utils.common import get_nvidia_driver_version_str
+
+        ver = get_nvidia_driver_version_str()
+        if ver is None:
             return {"CUDA Driver Version": "Not Available"}
+        return {"CUDA Driver Version": ver}
 
     def get_topology(self):
         """
diff --git a/python/sglang/cli/killall.py b/python/sglang/cli/killall.py
index 39a10f053..1c886e1eb 100755
--- a/python/sglang/cli/killall.py
+++ b/python/sglang/cli/killall.py
@@ -77,19 +77,11 @@ def _run_smi(query, query_type="gpu"):
 
 
 def _get_smi_version():
-    """Return nvidia-smi driver version and CUDA version, or None on failure."""
-    try:
-        out = subprocess.check_output(
-            [
-                "nvidia-smi",
-                "--query-gpu=driver_version",
-                "--format=csv,noheader,nounits",
-            ],
-            text=True,
-            timeout=10,
-        )
-        driver = out.strip().splitlines()[0].strip() if out.strip() else "unknown"
-    except (subprocess.SubprocessError, FileNotFoundError, IndexError):
+    """Return nvidia-smi driver version and GPU name, or None on failure."""
+    from sglang.srt.utils.common import get_nvidia_driver_version_str
+
+    driver = get_nvidia_driver_version_str()
+    if driver is None:
         return None
     try:
         out = subprocess.check_output(
diff --git a/python/sglang/srt/server_args.py b/python/sglang/srt/server_args.py
index ce4cc7df4..6b4759f5c 100644
--- a/python/sglang/srt/server_args.py
+++ b/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 (
diff --git a/python/sglang/srt/utils/common.py b/python/sglang/srt/utils/common.py
index d7ea34c76..74bb6962e 100644
--- a/python/sglang/srt/utils/common.py
+++ b/python/sglang/srt/utils/common.py
@@ -3399,6 +3399,41 @@ def is_triton_kernels_available() -> bool:
     return importlib.util.find_spec("triton_kernels") is not None
 
 
+@lru_cache(maxsize=1)
+def get_nvidia_driver_version() -> tuple:
+    """Return the NVIDIA driver version as a tuple of ints, e.g. (595, 58, 3).
+    Returns (0,) on failure."""
+    version_str = get_nvidia_driver_version_str()
+    if version_str is None:
+        return (0,)
+    try:
+        return tuple(int(x) for x in version_str.split("."))
+    except ValueError:
+        return (0,)
+
+
+@lru_cache(maxsize=1)
+def get_nvidia_driver_version_str() -> str:
+    """Return the NVIDIA driver version string, e.g. '595.58.03'.
+    Returns None on failure."""
+    try:
+        result = subprocess.run(
+            [
+                "nvidia-smi",
+                "--query-gpu=driver_version",
+                "--format=csv,noheader,nounits",
+            ],
+            capture_output=True,
+            text=True,
+            check=True,
+            timeout=10,
+        )
+        version_str = result.stdout.strip().split("\n")[0].strip()
+        return version_str if version_str else None
+    except (subprocess.CalledProcessError, FileNotFoundError, ValueError):
+        return None
+
+
 def check_cuda_result(raw_output):
     import cuda.bindings.runtime as cuda_rt