[Refactor] Set fp4-gemm-backend=auto on SM100 and rename fp4-gemm-backend with flashinfer_ prefix (#17309)

python/sglang/srt/layers/quantization/compressed_tensors/schemes/compressed_tensors_w4a4_nvfp4.py

@@ -98,7 +98,7 @@ class CompressedTensorsW4A4Fp4(CompressedTensorsScheme):
             layer.weight_global_scale.max().to(torch.float32), requires_grad=False
         )
 
-        if get_fp4_gemm_runner_backend().is_trtllm():
+        if get_fp4_gemm_runner_backend().is_flashinfer_trtllm():
             # FlashInfer TRTLLM FP4 GEMM requires a different weight layout.
             # FlashInfer provides nvfp4_quantize to quantize + shuffle the
             # layout but we use our own quantization so we have to call

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

@@ -16,21 +16,35 @@ class Fp4GemmRunnerBackend(Enum):
     """Enum for FP4 GEMM runner backend selection."""
 
     AUTO = "auto"
-    CUDNN = "cudnn"
-    CUTLASS = "cutlass"
-    TRTLLM = "trtllm"
+    FLASHINFER_CUDNN = "flashinfer_cudnn"
+    FLASHINFER_CUTLASS = "flashinfer_cutlass"
+    FLASHINFER_TRTLLM = "flashinfer_trtllm"
 
     def is_auto(self) -> bool:
         return self == Fp4GemmRunnerBackend.AUTO
 
-    def is_cudnn(self) -> bool:
-        return self == Fp4GemmRunnerBackend.CUDNN
+    def is_flashinfer_cudnn(self) -> bool:
+        return self == Fp4GemmRunnerBackend.FLASHINFER_CUDNN
 
-    def is_cutlass(self) -> bool:
-        return self == Fp4GemmRunnerBackend.CUTLASS
+    def is_flashinfer_cutlass(self) -> bool:
+        return self == Fp4GemmRunnerBackend.FLASHINFER_CUTLASS
 
-    def is_trtllm(self) -> bool:
-        return self == Fp4GemmRunnerBackend.TRTLLM
+    def is_flashinfer_trtllm(self) -> bool:
+        return self == Fp4GemmRunnerBackend.FLASHINFER_TRTLLM
+
+    def get_flashinfer_backend(self) -> str:
+        """Get the backend string to pass to FlashInfer's mm_fp4 API.
+
+        This remaps SGLang's user-facing backend names to FlashInfer's API names.
+        Examples:
+            'flashinfer_trtllm' -> 'trtllm'
+            'flashinfer_cutlass' -> 'cutlass'
+            'flashinfer_cudnn' -> 'cudnn'
+        """
+        if self.value.startswith("flashinfer_"):
+            return self.value.removeprefix("flashinfer_")
+        else:
+            return self.value
 
 
 FP4_GEMM_RUNNER_BACKEND: Fp4GemmRunnerBackend | None = None

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

@@ -128,10 +128,9 @@ def fp4_gemm(
     out_features: int,
 ) -> torch.Tensor:
     fp4_backend = get_fp4_gemm_runner_backend()
-    # TODO(shuw@nvidia.com): Remove the "cutlass" default override after flashinfer 0.6.0
-    # and let flashinfer's auto backend selection handle it.
-    backend = fp4_backend.value if not fp4_backend.is_auto() else "cutlass"
     if enable_flashinfer_fp4_gemm:
+        # Use the remapping logic to convert SGLang backend names to FlashInfer API names
+        backend = fp4_backend.get_flashinfer_backend()
         return flashinfer_fp4_gemm(
             input, weight, input_sf, weight_sf, alpha, out_dtype, backend=backend
         )
@@ -1155,7 +1154,7 @@ class ModelOptFp4LinearMethod(LinearMethodBase):
         layer.input_scale_inv = Parameter(
             (1 / input_scale_2).to(torch.float32), requires_grad=False
         )
-        if get_fp4_gemm_runner_backend().is_trtllm():
+        if get_fp4_gemm_runner_backend().is_flashinfer_trtllm():
             # FlashInfer TRTLLM FP4 GEMM requires a different weight layout.
             # FlashInfer provides nvfp4_quantize to quantize + shuffle the
             # layout but we use our own quantization so we have to call

python/sglang/srt/server_args.py

@@ -194,9 +194,9 @@ FP8_GEMM_RUNNER_BACKEND_CHOICES = [
 
 FP4_GEMM_RUNNER_BACKEND_CHOICES = [
     "auto",
-    "cudnn",
-    "cutlass",
-    "trtllm",
+    "flashinfer_cudnn",
+    "flashinfer_cutlass",
+    "flashinfer_trtllm",
 ]
 
 MAMBA_SSM_DTYPE_CHOICES = ["float32", "bfloat16"]
@@ -3564,10 +3564,10 @@ class ServerArgs:
             default=ServerArgs.fp4_gemm_runner_backend,
             dest="fp4_gemm_runner_backend",
             help="Choose the runner backend for NVFP4 GEMM operations. "
-            "Options: 'auto' (default, selects between cudnn/cutlass based on CUDA/cuDNN version), "
-            "'cudnn' (cuDNN backend, optimal on CUDA 13+ with cuDNN 9.15+), "
-            "'cutlass' (CUTLASS backend, optimal on CUDA 12), "
-            "'trtllm' (TensorRT-LLM backend, requires different weight preparation with shuffling). "
+            "Options: 'auto' (default, selects between flashinfer_cudnn/flashinfer_cutlass based on CUDA/cuDNN version), "
+            "'flashinfer_cudnn' (FlashInfer cuDNN backend, optimal on CUDA 13+ with cuDNN 9.15+), "
+            "'flashinfer_cutlass' (FlashInfer CUTLASS backend, optimal on CUDA 12), "
+            "'flashinfer_trtllm' (FlashInfer TensorRT-LLM backend, requires different weight preparation with shuffling). "
             "NOTE: This replaces the deprecated environment variable "
             "SGLANG_FLASHINFER_FP4_GEMM_BACKEND.",
         )