[diffusion] CI: enable warmup as default (#19010)

Author: mickqian

python/sglang/multimodal_gen/test/scripts/gen_perf_baselines.py

@@ -65,8 +65,9 @@ def _build_server_extra_args(case: DiffusionTestCase) -> str:
         a += f" --ring-degree {server_args.ring_degree}"
     if server_args.lora_path:
         a += f" --lora-path {server_args.lora_path}"
-    if server_args.warmup:
-        a += " --warmup"
+
+    # default warmup
+    a += " --warmup"
 
     for extra_arg in server_args.extras:
         a += f" {extra_arg}"

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

@@ -12,7 +12,6 @@
         DiffusionServerArgs(
             model_path="/root/.cache/modelscope/hub/models/Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
             modality="video",
-            warmup=True,
             custom_validator="video",
         ),
         DiffusionSamplingParams(

python/sglang/multimodal_gen/test/server/perf_baselines.json

@@ -523,28 +523,28 @@
         },
         "zimage_image_t2i": {
             "stages_ms": {
-                "ConditioningStage": 0.01,
-                "DecodingStage": 33.93,
-                "TextEncodingStage": 409.74,
                 "InputValidationStage": 0.03,
-                "DenoisingStage": 808.69,
-                "LatentPreparationStage": 0.14,
-                "TimestepPreparationStage": 1.33
+                "TextEncodingStage": 403.47,
+                "ConditioningStage": 0.01,
+                "TimestepPreparationStage": 1.41,
+                "LatentPreparationStage": 0.11,
+                "DenoisingStage": 756.21,
+                "DecodingStage": 29.41
             },
             "denoise_step_ms": {
-                "0": 23.4,
-                "1": 28.26,
-                "2": 100.72,
-                "3": 100.48,
-                "4": 99.62,
-                "5": 99.83,
-                "6": 99.67,
-                "7": 99.3,
-                "8": 100.83
-            },
-            "expected_e2e_ms": 1461.08,
-            "expected_avg_denoise_ms": 89.36,
-            "expected_median_denoise_ms": 99.67
+                "0": 22.29,
+                "1": 75.04,
+                "2": 93.82,
+                "3": 93.34,
+                "4": 93.38,
+                "5": 93.58,
+                "6": 94.01,
+                "7": 93.97,
+                "8": 94.32
+            },
+            "expected_e2e_ms": 1192.92,
+            "expected_avg_denoise_ms": 83.75,
+            "expected_median_denoise_ms": 93.58
         },
         "zimage_image_t2i_multi_lora": {
             "stages_ms": {
@@ -596,31 +596,6 @@
             "expected_avg_denoise_ms": 144.49,
             "expected_median_denoise_ms": 156.16
         },
-        "zimage_image_t2i_warmup": {
-            "stages_ms": {
-                "InputValidationStage": 0.02,
-                "TextEncodingStage": 456.16,
-                "ConditioningStage": 0.01,
-                "TimestepPreparationStage": 0.98,
-                "LatentPreparationStage": 0.06,
-                "DenoisingStage": 889.42,
-                "DecodingStage": 37.81
-            },
-            "denoise_step_ms": {
-                "0": 16.49,
-                "1": 94.63,
-                "2": 109.65,
-                "3": 110.05,
-                "4": 109.39,
-                "5": 110.58,
-                "6": 109.52,
-                "7": 110.54,
-                "8": 115.24
-            },
-            "expected_e2e_ms": 1184.45,
-            "expected_avg_denoise_ms": 98.46,
-            "expected_median_denoise_ms": 109.65
-        },
         "qwen_image_edit_ti2i": {
             "stages_ms": {
                 "LatentPreparationStage": 0.16,

python/sglang/multimodal_gen/test/server/test_server_common.py

@@ -87,8 +87,8 @@ def diffusion_server(case: DiffusionTestCase) -> ServerContext:
     if server_args.lora_path:
         extra_args += f" --lora-path {server_args.lora_path}"
 
-    if server_args.warmup:
-        extra_args += f" --warmup"
+    # default warmup
+    extra_args += f" --warmup"
 
     for arg in server_args.extras:
         extra_args += f" {arg}"