fix: default FP4 GEMM backend to flashinfer_cudnn on SM120 (Blackwell) (#20047)

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Co-authored-by: Brayden Zhong <b8zhong@uwaterloo.ca>

python/sglang/srt/layers/quantization/fp4_utils.py

@@ -5,6 +5,7 @@ from enum import Enum
 from typing import TYPE_CHECKING
 
 from sglang.srt.environ import envs
+from sglang.srt.utils.common import is_sm120_supported
 
 if TYPE_CHECKING:
     from sglang.srt.server_args import ServerArgs
@@ -75,6 +76,19 @@ def initialize_fp4_gemm_config(server_args: ServerArgs) -> None:
                 "Using server argument value."
             )
 
+    if backend == "auto":
+        if is_sm120_supported():
+            # flashinfer_cutlass produces NaN in dense MLP layers with
+            # heterogeneous batches on SM120 (Blackwell).  cudnn is stable.
+            # See: https://github.com/sgl-project/sglang/issues/20043
+            backend = "flashinfer_cudnn"
+            logger.info(
+                "SM120 (Blackwell) detected: auto-selecting "
+                "fp4-gemm-backend=flashinfer_cudnn"
+            )
+        else:
+            backend = "flashinfer_cutlass"
+
     FP4_GEMM_RUNNER_BACKEND = Fp4GemmRunnerBackend(backend)
 
 

python/sglang/srt/server_args.py

@@ -466,7 +466,7 @@ class ServerArgs:
     grammar_backend: Optional[str] = None
     mm_attention_backend: Optional[str] = None
     fp8_gemm_runner_backend: str = "auto"
-    fp4_gemm_runner_backend: str = "flashinfer_cutlass"
+    fp4_gemm_runner_backend: str = "auto"
     nsa_prefill_backend: Optional[str] = (
         None  # None = auto-detect based on hardware/kv_cache_dtype
     )
@@ -4308,8 +4308,8 @@ class ServerArgs:
             default=ServerArgs.fp4_gemm_runner_backend,
             dest="fp4_gemm_runner_backend",
             help="Choose the runner backend for NVFP4 GEMM operations. "
-            "Options: 'flashinfer_cutlass' (default), "
+            "Options: 'auto' (default; selects flashinfer_cudnn on SM120, flashinfer_cutlass otherwise), "
-            "'auto' (auto-selects between flashinfer_cudnn/flashinfer_cutlass based on CUDA/cuDNN version), "
+            "'flashinfer_cutlass' (CUTLASS backend), "
             "'flashinfer_cudnn' (FlashInfer cuDNN backend, optimal on CUDA 13+ with cuDNN 9.15+), "
             "'flashinfer_trtllm' (FlashInfer TensorRT-LLM backend, requires different weight preparation with shuffling). "
             "NOTE: This replaces the deprecated environment variable "