[Diffusion] [NPU] [CI] fix CI performance (#19486)

python/sglang/multimodal_gen/runtime/platforms/npu.py

@@ -124,3 +124,8 @@ class NPUPlatformBase(Platform):
     @classmethod
     def get_device_communicator_cls(cls) -> str:
         return "sglang.multimodal_gen.runtime.distributed.device_communicators.cuda_communicator.CudaCommunicator"  # noqa
+
+    @classmethod
+    def enable_dit_layerwise_offload_for_wan_by_default(cls) -> bool:
+        """The performance of the layerwise_offload feature depends on the device's memory size and the memory size occupied by the model. Use --dit-layerwise-offload True if it suitable for your case."""
+        return False

python/sglang/multimodal_gen/test/server/ascend/perf_baselines_npu.json

@@ -12,7 +12,7 @@
                 "LatentPreparationStage": 0.25,
                 "TimestepPreparationStage": 2.9,
                 "DenoisingStage": 26188.0,
-                "DecodingStage": 142.03,
+                "DecodingStage": 320.03,
                 "per_frame_generation": null
             },
             "denoise_step_ms": {