Piecewise Cuda Graph Support for gpt-oss model (#13045)

Signed-off-by: Oasis-Git <ayw.sirius19@gmail.com>
2025-11-15 09:28:00 +08:00
parent 14dc052382
commit eae59b337e
5 changed files with 28 additions and 24 deletions
@@ -998,7 +998,12 @@ class ServerArgs:
                self.dtype = "bfloat16"

            if self.moe_runner_backend == "auto":
-                if is_blackwell_supported() and is_mxfp4_quant_format:
+                if self.enable_piecewise_cuda_graph:
+                    self.moe_runner_backend = "auto"
+                    logger.warning(
+                        "Enable piecewise CUDA graph, enabling auto MOE kernel."
+                    )
+                elif is_blackwell_supported() and is_mxfp4_quant_format:
                    self.moe_runner_backend = "flashinfer_mxfp4"
                    logger.warning(
                        "Detected SM100 and MXFP4 quantization format for GPT-OSS model, enabling FlashInfer MXFP4 MOE kernel."