[Diffusion] [NPU] Fix CI run (#18921)

python/sglang/multimodal_gen/runtime/managers/gpu_worker.py

@@ -154,8 +154,8 @@ class GPUWorker:
             output_batch.timings.record_memory_snapshot("mem_analysis", final_snapshot)
 
         # for details on max_memory_reserved: https://docs.pytorch.org/docs/stable/generated/torch.cuda.memory.max_memory_reserved.html
-        peak_reserved_bytes = torch.cuda.max_memory_reserved()
-        peak_allocated_bytes = torch.cuda.max_memory_allocated()
+        peak_reserved_bytes = torch.get_device_module().max_memory_reserved()
+        peak_allocated_bytes = torch.get_device_module().max_memory_allocated()
 
         output_batch.peak_memory_mb = peak_reserved_bytes / (1024**2)
         peak_reserved_gb = peak_reserved_bytes / (1024**3)

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

@@ -17,9 +17,9 @@
                 "per_frame_generation": null
             },
             "denoise_step_ms": {
-                "0": 195.27,
-                "1": 329.05,
-                "2": 545.43,
+                "0": 364.97,
+                "1": 542.66,
+                "2": 542.1,
                 "3": 541.3,
                 "4": 537.07,
                 "5": 537.21,

python/sglang/multimodal_gen/test/server/ascend/testcase_configs_npu.py

@@ -12,7 +12,7 @@ ONE_NPU_CASES: list[DiffusionTestCase] = [
         DiffusionServerArgs(
             model_path="/root/.cache/modelscope/hub/models/Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
             modality="video",
-            warmup=0,
+            warmup=True,
             custom_validator="video",
         ),
         DiffusionSamplingParams(