[NVIDIA] Deterministic inference backend order on Blackwell (#21450)

python/sglang/srt/server_args.py

@@ -913,6 +913,9 @@ class ServerArgs:
 
         # Set kernel backends.
         self._handle_sampling_backend()
+        # Must run before _handle_attention_backend_compatibility so the
+        # deterministic backend is set before auto-detection fills it in.
+        self._handle_deterministic_inference()
         self._handle_attention_backend_compatibility()
         self._handle_mamba_backend()
         self._handle_linear_attn_backend()
@@ -970,9 +973,6 @@ class ServerArgs:
         # Validate cache settings.
         self._handle_cache_compatibility()
 
-        # Handle deterministic inference.
-        self._handle_deterministic_inference()
-
         # Handle diffusion LLM inference.
         self._handle_dllm_inference()