Reland Cute-DSL FP4 dense GEMM (#23590)

Co-authored-by: b8zhong <b8zhong@users.noreply.github.com>

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

@@ -4,7 +4,7 @@ import logging
 from enum import Enum
 from typing import TYPE_CHECKING
 
-from sglang.srt.utils.common import is_sm120_supported
+from sglang.srt.utils.common import is_sm100_supported, is_sm120_supported
 
 if TYPE_CHECKING:
     from sglang.srt.server_args import ServerArgs
@@ -18,6 +18,7 @@ class Fp4GemmRunnerBackend(Enum):
     AUTO = "auto"
     CUTLASS = "cutlass"
     FLASHINFER_CUDNN = "flashinfer_cudnn"
+    FLASHINFER_CUTEDSL = "flashinfer_cutedsl"
     FLASHINFER_CUTLASS = "flashinfer_cutlass"
     FLASHINFER_TRTLLM = "flashinfer_trtllm"
 
@@ -36,6 +37,9 @@ class Fp4GemmRunnerBackend(Enum):
     def is_flashinfer_trtllm(self) -> bool:
         return self == Fp4GemmRunnerBackend.FLASHINFER_TRTLLM
 
+    def is_flashinfer_cutedsl(self) -> bool:
+        return self == Fp4GemmRunnerBackend.FLASHINFER_CUTEDSL
+
     def is_flashinfer(self) -> bool:
         return self.value.startswith("flashinfer_")
 
@@ -47,7 +51,10 @@ class Fp4GemmRunnerBackend(Enum):
             'flashinfer_trtllm' -> 'trtllm'
             'flashinfer_cutlass' -> 'cutlass'
             'flashinfer_cudnn' -> 'cudnn'
+            'flashinfer_cutedsl' -> 'cute-dsl'
         """
+        if self == Fp4GemmRunnerBackend.FLASHINFER_CUTEDSL:
+            return "cute-dsl"
         if self.value.startswith("flashinfer_"):
             return self.value.removeprefix("flashinfer_")
         else:
@@ -68,10 +75,8 @@ def initialize_fp4_gemm_config(server_args: ServerArgs) -> None:
             # 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"
-            )
+        elif is_sm100_supported():
+            backend = "flashinfer_cutedsl"
         else:
             backend = "flashinfer_cutlass"
 

python/sglang/srt/server_args.py

@@ -220,6 +220,7 @@ FP4_GEMM_RUNNER_BACKEND_CHOICES = [
     "auto",
     "cutlass",
     "flashinfer_cudnn",
+    "flashinfer_cutedsl",
     "flashinfer_cutlass",
     "flashinfer_trtllm",
 ]
@@ -5482,10 +5483,11 @@ 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 flashinfer_cudnn on SM120, flashinfer_cutlass otherwise), "
+            "Options: 'auto' (default; selects flashinfer_cudnn on SM120, flashinfer_cutedsl on SM100, flashinfer_cutlass otherwise), "
             "'cutlass' (SGLang CUTLASS kernel), "
             "'flashinfer_cutlass' (FlashInfer CUTLASS backend), "
             "'flashinfer_cudnn' (FlashInfer cuDNN backend, optimal on CUDA 13+ with cuDNN 9.15+), "
+            "'flashinfer_cutedsl' (FlashInfer CuTe DSL backend), "
             "'flashinfer_trtllm' (FlashInfer TensorRT-LLM backend, requires different weight preparation with shuffling). ",
         )
         parser.add_argument(

python/sglang/srt/utils/common.py

@@ -1229,6 +1229,11 @@ def configure_logger(server_args, prefix: str = ""):
     for name in ("httpx", "httpcore"):
         logging.getLogger(name).setLevel(logging.WARNING)
 
+    if is_flashinfer_available():
+        from flashinfer.jit.core import logger as flashinfer_logger
+
+        flashinfer_logger.setLevel(logging.ERROR)
+
 
 # source: https://github.com/vllm-project/vllm/blob/93b38bea5dd03e1b140ca997dfaadef86f8f1855/vllm/lora/utils.py#L9
 def replace_submodule(