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

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