[diffusion] CI: deprecate WarmupRunner in CI (#18038)

Co-authored-by: Mick <mickjagger19@icloud.com>

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

@@ -186,6 +186,8 @@ class GPUWorker:
 
             # TODO: extract to avoid duplication
             if req.perf_dump_path is not None or envs.SGLANG_DIFFUSION_STAGE_LOGGING:
+                # Avoid logging warmup perf records that share the same request_id.
+                if not req.is_warmup:
                     PerformanceLogger.log_request_summary(timings=output_batch.timings)
         except Exception as e:
             logger.error(

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

@@ -10,8 +10,6 @@ from openai import OpenAI
 
 from sglang.multimodal_gen.test.server.test_server_utils import (
     ServerManager,
-    WarmupRunner,
-    download_image_from_url,
     get_generate_fn,
 )
 from sglang.multimodal_gen.test.server.testcase_configs import (
@@ -20,7 +18,6 @@ from sglang.multimodal_gen.test.server.testcase_configs import (
 )
 from sglang.multimodal_gen.test.test_utils import (
     get_dynamic_server_port,
-    is_image_url,
     wait_for_req_perf_record,
 )
 
@@ -68,8 +65,8 @@ 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.enable_warmup:
+    if server_args.warmup:
-        a += " --enable-warmup"
+        a += " --warmup"
     return a
 
 
@@ -110,35 +107,6 @@ def _run_case(case: DiffusionTestCase) -> dict:
     try:
         sp = case.sampling_params
         output_size = os.environ.get("SGLANG_TEST_OUTPUT_SIZE", sp.output_size)
-        w = WarmupRunner(
-            port=ctx.port,
-            model=case.server_args.model_path,
-            prompt=sp.prompt or "A colorful raccoon icon",
-            output_size=output_size,
-            output_format=sp.output_format,
-        )
-        if case.server_args.warmup > 0:
-            if sp.image_path and sp.prompt:
-                image_path_list = sp.image_path
-                if not isinstance(image_path_list, list):
-                    image_path_list = [image_path_list]
-                new_image_path_list = []
-                for p in image_path_list:
-                    if is_image_url(p):
-                        new_image_path_list.append(download_image_from_url(str(p)))
-                    else:
-                        pp = Path(p)
-                        if not pp.exists():
-                            raise FileNotFoundError(str(pp))
-                        new_image_path_list.append(pp)
-                w.run_edit_warmups(
-                    count=case.server_args.warmup,
-                    edit_prompt=sp.prompt,
-                    image_path=new_image_path_list,
-                )
-            else:
-                w.run_text_warmups(case.server_args.warmup)
-
         client = _openai_client(ctx.port)
         gen = get_generate_fn(
             model_path=case.server_args.model_path,

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

@@ -31,11 +31,7 @@
             0.6,
             0.8,
             1.0
-        ],
+        ]
-        "warmup_requests": {
-            "text": 1,
-            "image_edit": 0
-        }
     },
     "scenarios": {
         "qwen_image_t2i": {
@@ -53,7 +49,7 @@
                 "DecodingStage": 751.1
             },
             "denoise_step_ms": {
-                "0": 1077.77,
+                "0": 400.06,
                 "1": 345.13,
                 "2": 413.8,
                 "3": 405.49,
@@ -182,8 +178,8 @@
                 "DecodingStage": 653.03
             },
             "denoise_step_ms": {
-                "0": 165.27,
+                "0": 58.88,
-                "1": 58.88,
+                "1": 165.27,
                 "2": 166.85,
                 "3": 166.51,
                 "4": 166.77,
@@ -249,8 +245,8 @@
                 "DecodingStage": 489.8
             },
             "denoise_step_ms": {
-                "0": 511.3,
+                "0": 68.7,
-                "1": 132.57,
+                "1": 511.3,
                 "2": 541.19,
                 "3": 518.93,
                 "4": 541.2,
@@ -470,8 +466,8 @@
                 "DecodingStage": 381.56
             },
             "denoise_step_ms": {
-                "0": 166.27,
+                "0": 73.27,
-                "1": 59.6,
+                "1": 166.6,
                 "2": 167.31,
                 "3": 168.7,
                 "4": 168.83,
@@ -703,56 +699,56 @@
                 "DecodingStage": 52.24
             },
             "denoise_step_ms": {
-                "0": 239.28,
+                "0": 227.68,
-                "1": 285.07,
+                "1": 277.41,
-                "2": 286.65,
+                "2": 276.7,
-                "3": 299.09,
+                "3": 291.52,
-                "4": 191.11,
+                "4": 52.8,
-                "5": 101.8,
+                "5": 6.58,
-                "6": 51.77,
+                "6": 231.58,
-                "7": 147.26,
+                "7": 52.55,
-                "8": 101.68,
+                "8": 7.69,
-                "9": 51.4,
+                "9": 230.59,
-                "10": 150.58,
+                "10": 52.58,
-                "11": 106.56,
+                "11": 7.14,
-                "12": 53.7,
+                "12": 6.95,
-                "13": 7.33,
+                "13": 234.71,
-                "14": 148.41,
+                "14": 53.28,
-                "15": 102.08,
+                "15": 7.09,
-                "16": 52.04,
+                "16": 6.63,
-                "17": 7.3,
+                "17": 233.93,
-                "18": 145.91,
+                "18": 52.71,
-                "19": 100.31,
+                "19": 6.64,
-                "20": 51.68,
+                "20": 6.5,
-                "21": 7.27,
+                "21": 231.37,
-                "22": 145.92,
+                "22": 52.28,
-                "23": 104.35,
+                "23": 6.61,
-                "24": 54.33,
+                "24": 6.48,
-                "25": 7.46,
+                "25": 232.86,
-                "26": 150.96,
+                "26": 54.92,
-                "27": 104.39,
+                "27": 7.51,
-                "28": 53.21,
+                "28": 7.19,
-                "29": 7.22,
+                "29": 233.51,
-                "30": 148.3,
+                "30": 52.97,
-                "31": 102.06,
+                "31": 6.72,
-                "32": 51.88,
+                "32": 7.02,
-                "33": 7.31,
+                "33": 233.14,
-                "34": 145.8,
+                "34": 52.47,
-                "35": 101.95,
+                "35": 6.66,
-                "36": 52.15,
+                "36": 6.52,
-                "37": 7.42,
+                "37": 233.84,
-                "38": 148.87,
+                "38": 51.49,
-                "39": 103.01,
+                "39": 6.87,
-                "40": 52.21,
+                "40": 6.74,
-                "41": 7.28,
+                "41": 233.75,
-                "42": 147.39,
+                "42": 52.65,
-                "43": 103.89,
+                "43": 6.62,
-                "44": 51.44,
+                "44": 6.55,
-                "45": 7.12,
+                "45": 233.45,
-                "46": 144.78,
+                "46": 52.33,
-                "47": 100.34,
+                "47": 6.55,
-                "48": 195.54,
+                "48": 232.58,
-                "49": 246.97
+                "49": 52.84
             },
             "expected_e2e_ms": 5982.78,
             "expected_avg_denoise_ms": 104.84,
@@ -770,7 +766,7 @@
                 "per_frame_generation": null
             },
             "denoise_step_ms": {
-                "0": 879.71,
+                "0": 240.71,
                 "1": 248.13,
                 "2": 246.48,
                 "3": 247.87,
@@ -837,7 +833,7 @@
                 "per_frame_generation": null
             },
             "denoise_step_ms": {
-                "0": 879.71,
+                "0": 224.71,
                 "1": 248.13,
                 "2": 246.48,
                 "3": 247.87,
@@ -971,7 +967,7 @@
                 "per_frame_generation": null
             },
             "denoise_step_ms": {
-                "0": 1069.91,
+                "0": 266.91,
                 "1": 211.32,
                 "2": 206.59,
                 "3": 208.12,
@@ -1237,22 +1233,22 @@
         "fastwan2_2_ti2v_5b": {
             "stages_ms": {
                 "InputValidationStage": 300.00,
-                "TextEncodingStage": 2327.87,
+                "TextEncodingStage": 843.86,
                 "ConditioningStage": 0.01,
                 "TimestepPreparationStage": 58.66,
                 "LatentPreparationStage": 28.55,
-                "DmdDenoisingStage": 4438.3,
+                "DmdDenoisingStage": 499.34,
-                "DecodingStage": 14177.77,
+                "DecodingStage": 1924.01,
                 "per_frame_generation": null
             },
             "denoise_step_ms": {
-                "0": 2022.21,
+                "0": 164.76,
-                "1": 1263.17,
+                "1": 165.6,
-                "2": 1149.59
+                "2": 165.84
             },
-            "expected_e2e_ms": 21133.36,
+            "expected_e2e_ms": 7722.91,
-            "expected_avg_denoise_ms": 1478.32,
+            "expected_avg_denoise_ms": 165.42,
-            "expected_median_denoise_ms": 1263.17
+            "expected_median_denoise_ms": 165.66
         },
         "fast_hunyuan_video": {
             "stages_ms": {
@@ -1266,12 +1262,12 @@
                 "per_frame_generation": null
             },
             "denoise_step_ms": {
-                "0": 2518.62,
+                "0": 485.99,
-                "1": 578.59,
+                "1": 1399.84,
-                "2": 1485.76,
+                "2": 1399.91,
-                "3": 1490.86,
+                "3": 1397.79,
-                "4": 1489.93,
+                "4": 1400.61,
-                "5": 1487.02
+                "5": 1402.53
             },
             "expected_e2e_ms": 16672.15,
             "expected_avg_denoise_ms": 1608.46,
@@ -1290,7 +1286,7 @@
                 "per_frame_generation": null
             },
             "denoise_step_ms": {
-                "0": 15659.6,
+                "0": 1525.6,
                 "1": 1582.6,
                 "2": 1597.84,
                 "3": 1601.34,
@@ -1687,7 +1683,7 @@
                 "per_frame_generation": null
             },
             "denoise_step_ms": {
-                "0": 1964.07,
+                "0": 233.29,
                 "1": 265.02,
                 "2": 257.83,
                 "3": 260.27,
@@ -1756,7 +1752,7 @@
                 "per_frame_generation": null
             },
             "denoise_step_ms": {
-                "0": 7828.3,
+                "0": 6680.62,
                 "1": 3765.8,
                 "2": 3774.63,
                 "3": 3772.93,

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

@@ -8,7 +8,6 @@ If the actual run is significantly better than the baseline, the improved cases
 from __future__ import annotations
 
 import os
-from pathlib import Path
 from typing import Any, Callable
 
 import openai
@@ -25,8 +24,6 @@ from sglang.multimodal_gen.test.server.test_server_utils import (
     PerformanceValidator,
     ServerContext,
     ServerManager,
-    WarmupRunner,
-    download_image_from_url,
     get_generate_fn,
 )
 from sglang.multimodal_gen.test.server.testcase_configs import (
@@ -37,7 +34,6 @@ from sglang.multimodal_gen.test.server.testcase_configs import (
 )
 from sglang.multimodal_gen.test.test_utils import (
     get_dynamic_server_port,
-    is_image_url,
     wait_for_req_perf_record,
 )
 
@@ -91,8 +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.enable_warmup:
+    if server_args.warmup:
-        extra_args += f" --enable-warmup"
+        extra_args += f" --warmup"
 
     # Build custom environment variables
     env_vars = {}
@@ -115,41 +111,6 @@ def diffusion_server(case: DiffusionTestCase) -> ServerContext:
         output_size = os.environ.get(
             "SGLANG_TEST_OUTPUT_SIZE", sampling_params.output_size
         )
-        warmup = WarmupRunner(
-            port=ctx.port,
-            model=server_args.model_path,
-            prompt=sampling_params.prompt or "A colorful raccoon icon",
-            output_size=output_size,
-            output_format=sampling_params.output_format,
-        )
-        if server_args.warmup > 0:
-            if sampling_params.image_path and case.sampling_params.prompt:
-                # Handle URL or local path
-                image_path_list = sampling_params.image_path
-                if not isinstance(image_path_list, list):
-                    image_path_list = [image_path_list]
-
-                new_image_path_list = []
-                for image_path in image_path_list:
-                    if is_image_url(image_path):
-                        new_image_path_list.append(
-                            download_image_from_url(str(image_path))
-                        )
-                    else:
-                        path_obj = Path(image_path)
-                        if not path_obj.exists():
-                            pytest.skip(f"{case.id}: file missing: {image_path}")
-                        new_image_path_list.append(path_obj)
-
-                image_path_list = new_image_path_list
-
-                warmup.run_edit_warmups(
-                    count=server_args.warmup,
-                    edit_prompt=sampling_params.prompt,
-                    image_path=image_path_list,
-                )
-            else:
-                warmup.run_text_warmups(server_args.warmup)
     except Exception as exc:
         logger.error("Warm-up failed for %s: %s", case.id, exc)
         ctx.cleanup()

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

@@ -18,7 +18,7 @@ from typing import Any, Callable, Sequence
 from urllib.request import urlopen
 
 import pytest
-from openai import Client, OpenAI
+from openai import Client
 
 from sglang.multimodal_gen.benchmarks.compare_perf import calculate_upper_bound
 from sglang.multimodal_gen.runtime.platforms import current_platform
@@ -451,81 +451,6 @@ class ServerManager:
             return ""
 
 
-class WarmupRunner:
-    """Handles warmup requests for a server."""
-
-    def __init__(
-        self,
-        port: int,
-        model: str,
-        prompt: str,
-        output_size: str,
-        output_format: str = None,
-    ):
-        self.client = OpenAI(
-            api_key="sglang-anything",
-            base_url=f"http://localhost:{port}/v1",
-        )
-        self.model = model
-        self.prompt = prompt
-        self.output_size = output_size
-        self.output_format = output_format
-
-    def run_text_warmups(self, count: int) -> None:
-        """Run text-to-image warmup requests."""
-        if count <= 0:
-            return
-
-        logger.info("[server-test] Running %s text warm-up(s)", count)
-        for _ in range(count):
-            result = self.client.images.generate(
-                model=self.model,
-                prompt=self.prompt,
-                n=1,
-                size=self.output_size,
-                response_format="b64_json",
-            )
-            validate_image(result.data[0].b64_json)
-
-    def run_edit_warmups(
-        self,
-        count: int,
-        edit_prompt: str,
-        image_path: Path,
-    ) -> None:
-        """Run image-edit warmup requests."""
-        if count <= 0:
-            return
-
-        if not isinstance(image_path, list):
-            image_path = [image_path]
-
-        for image in image_path:
-            if not image.exists():
-                logger.warning(
-                    "[server-test] Skipping edit warmup: image missing at %s", image
-                )
-                return
-
-        logger.info("[server-test] Running %s edit warm-up(s)", count)
-        for _ in range(count):
-            images = [open(image, "rb") for image in image_path]
-            try:
-                result = self.client.images.edit(
-                    model=self.model,
-                    image=images,
-                    prompt=edit_prompt,
-                    n=1,
-                    size=self.output_size,
-                    response_format="b64_json",
-                    output_format=self.output_format,
-                )
-            finally:
-                for img in images:
-                    img.close()
-            validate_image(result.data[0].b64_json)
-
-
 class PerformanceValidator:
     """Validates performance metrics against expectations."""
 

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

@@ -98,7 +98,6 @@ class BaselineConfig:
 
     scenarios: dict[str, ScenarioConfig]
     step_fractions: Sequence[float]
-    warmup_defaults: dict[str, int]
     tolerances: ToleranceConfig
     improvement_threshold: float
 
@@ -127,7 +126,6 @@ class BaselineConfig:
         return cls(
             scenarios=scenarios,
             step_fractions=tuple(data["sampling"]["step_fractions"]),
-            warmup_defaults=data["sampling"].get("warmup_requests", {}),
             tolerances=tolerances,
             improvement_threshold=data.get("improvement_reporting", {}).get(
                 "threshold", 0.2
@@ -142,7 +140,6 @@ class DiffusionServerArgs:
     model_path: str  # HF repo or local path
     modality: str = "image"  # "image" or "video" or "3d"
 
-    warmup: int = 1  # number of warmup requests
     custom_validator: str | None = None  # optional custom validator name
     # resources
     num_gpus: int = 1
@@ -161,7 +158,7 @@ class DiffusionServerArgs:
         None  # Second LoRA adapter path for multi-LoRA testing
     )
     # misc
-    enable_warmup: bool = False
+    warmup: bool = False
 
     dit_layerwise_offload: bool = False
     dit_offload_prefetch_size: int | float | None = None
@@ -325,6 +322,7 @@ ONE_GPU_CASES_A: list[DiffusionTestCase] = [
         DiffusionServerArgs(
             model_path="Qwen/Qwen-Image",
             modality="image",
+            warmup=True,
         ),
         T2I_sampling_params,
     ),
@@ -334,22 +332,21 @@ ONE_GPU_CASES_A: list[DiffusionTestCase] = [
             model_path="Qwen/Qwen-Image",
             modality="image",
             enable_cache_dit=True,
+            warmup=True,
         ),
         T2I_sampling_params,
     ),
     DiffusionTestCase(
         "flux_image_t2i",
         DiffusionServerArgs(
-            model_path="black-forest-labs/FLUX.1-dev",
+            model_path="black-forest-labs/FLUX.1-dev", modality="image", warmup=True
-            modality="image",
         ),
         T2I_sampling_params,
     ),
     DiffusionTestCase(
         "flux_2_image_t2i",
         DiffusionServerArgs(
-            model_path="black-forest-labs/FLUX.2-dev",
+            model_path="black-forest-labs/FLUX.2-dev", modality="image", warmup=True
-            modality="image",
         ),
         T2I_sampling_params,
     ),
@@ -358,6 +355,7 @@ ONE_GPU_CASES_A: list[DiffusionTestCase] = [
         DiffusionServerArgs(
             model_path="black-forest-labs/FLUX.2-klein-4B",
             modality="image",
+            warmup=True,
         ),
         T2I_sampling_params,
     ),
@@ -371,21 +369,21 @@ ONE_GPU_CASES_A: list[DiffusionTestCase] = [
             modality="image",
             dit_layerwise_offload=True,
             dit_offload_prefetch_size=2,
+            warmup=True,
         ),
         T2I_sampling_params,
     ),
     DiffusionTestCase(
         "zimage_image_t2i",
         DiffusionServerArgs(
-            model_path="Tongyi-MAI/Z-Image-Turbo",
+            model_path="Tongyi-MAI/Z-Image-Turbo", modality="image", warmup=True
-            modality="image",
         ),
         T2I_sampling_params,
     ),
     DiffusionTestCase(
         "zimage_image_t2i_warmup",
         DiffusionServerArgs(
-            model_path="Tongyi-MAI/Z-Image-Turbo", modality="image", enable_warmup=True
+            model_path="Tongyi-MAI/Z-Image-Turbo", modality="image", warmup=True
         ),
         T2I_sampling_params,
     ),
@@ -397,6 +395,7 @@ ONE_GPU_CASES_A: list[DiffusionTestCase] = [
             modality="image",
             lora_path="reverentelusarca/elusarca-anime-style-lora-z-image-turbo",
             second_lora_path="tarn59/pixel_art_style_lora_z_image_turbo",
+            warmup=True,
         ),
         T2I_sampling_params,
     ),
@@ -404,32 +403,28 @@ ONE_GPU_CASES_A: list[DiffusionTestCase] = [
     DiffusionTestCase(
         "qwen_image_edit_ti2i",
         DiffusionServerArgs(
-            model_path="Qwen/Qwen-Image-Edit",
+            model_path="Qwen/Qwen-Image-Edit", modality="image", warmup=True
-            modality="image",
         ),
         TI2I_sampling_params,
     ),
     DiffusionTestCase(
         "qwen_image_edit_2509_ti2i",
         DiffusionServerArgs(
-            model_path="Qwen/Qwen-Image-Edit-2509",
+            model_path="Qwen/Qwen-Image-Edit-2509", modality="image", warmup=True
-            modality="image",
         ),
         MULTI_IMAGE_TI2I_sampling_params,
     ),
     DiffusionTestCase(
         "qwen_image_edit_2511_ti2i",
         DiffusionServerArgs(
-            model_path="Qwen/Qwen-Image-Edit-2511",
+            model_path="Qwen/Qwen-Image-Edit-2511", modality="image", warmup=True
-            modality="image",
         ),
         TI2I_sampling_params,
     ),
     DiffusionTestCase(
         "qwen_image_layered_i2i",
         DiffusionServerArgs(
-            model_path="Qwen/Qwen-Image-Layered",
+            model_path="Qwen/Qwen-Image-Layered", modality="image", warmup=True
-            modality="image",
         ),
         MULTI_FRAME_I2I_sampling_params,
     ),
@@ -442,8 +437,8 @@ ONE_GPU_CASES_B: list[DiffusionTestCase] = [
         DiffusionServerArgs(
             model_path="Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
             modality="video",
-            warmup=0,
             custom_validator="video",
+            warmup=True,
         ),
         DiffusionSamplingParams(
             prompt=T2V_PROMPT,
@@ -454,9 +449,9 @@ ONE_GPU_CASES_B: list[DiffusionTestCase] = [
         DiffusionServerArgs(
             model_path="Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
             modality="video",
-            warmup=0,
             custom_validator="video",
             text_encoder_cpu_offload=True,
+            warmup=True,
         ),
         DiffusionSamplingParams(
             prompt=T2V_PROMPT,
@@ -468,8 +463,8 @@ ONE_GPU_CASES_B: list[DiffusionTestCase] = [
         DiffusionServerArgs(
             model_path="Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
             modality="video",
-            warmup=0,
             custom_validator="video",
+            warmup=True,
         ),
         DiffusionSamplingParams(
             prompt=T2V_PROMPT,
@@ -484,10 +479,10 @@ ONE_GPU_CASES_B: list[DiffusionTestCase] = [
         DiffusionServerArgs(
             model_path="Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
             modality="video",
-            warmup=0,
             custom_validator="video",
             num_gpus=1,
             dynamic_lora_path="Cseti/Wan-LoRA-Arcane-Jinx-v1",
+            warmup=True,
         ),
         DiffusionSamplingParams(
             prompt="csetiarcane Nfj1nx with blue hair, a woman walking in a cyberpunk city at night",
@@ -508,8 +503,7 @@ ONE_GPU_CASES_B: list[DiffusionTestCase] = [
     DiffusionTestCase(
         "flux_2_ti2i",
         DiffusionServerArgs(
-            model_path="black-forest-labs/FLUX.2-dev",
+            model_path="black-forest-labs/FLUX.2-dev", modality="image", warmup=True
-            modality="image",
         ),
         TI2I_sampling_params,
     ),
@@ -518,8 +512,8 @@ ONE_GPU_CASES_B: list[DiffusionTestCase] = [
         DiffusionServerArgs(
             model_path="FastVideo/FastHunyuan-diffusers",
             modality="video",
-            warmup=0,
             custom_validator="video",
+            warmup=True,
         ),
         DiffusionSamplingParams(
             prompt=T2V_PROMPT,
@@ -531,8 +525,8 @@ ONE_GPU_CASES_B: list[DiffusionTestCase] = [
         DiffusionServerArgs(
             model_path="Wan-AI/Wan2.2-TI2V-5B-Diffusers",
             modality="video",
-            warmup=0,
             custom_validator="video",
+            warmup=True,
         ),
         TI2V_sampling_params,
     ),
@@ -541,8 +535,8 @@ ONE_GPU_CASES_B: list[DiffusionTestCase] = [
         DiffusionServerArgs(
             model_path="FastVideo/FastWan2.2-TI2V-5B-FullAttn-Diffusers",
             modality="video",
-            warmup=0,
             custom_validator="video",
+            warmup=True,
         ),
         TI2V_sampling_params,
     ),
@@ -556,8 +550,8 @@ if not current_platform.is_hip():
             DiffusionServerArgs(
                 model_path="IPostYellow/TurboWan2.1-T2V-1.3B-Diffusers",
                 modality="video",
-                warmup=0,
                 custom_validator="video",
+                warmup=True,
             ),
             DiffusionSamplingParams(
                 prompt=T2V_PROMPT,
@@ -571,8 +565,8 @@ TWO_GPU_CASES_A = [
         DiffusionServerArgs(
             model_path="Wan-AI/Wan2.2-I2V-A14B-Diffusers",
             modality="video",
-            warmup=0,
             custom_validator="video",
+            warmup=True,
         ),
         TI2V_sampling_params,
     ),
@@ -581,9 +575,9 @@ TWO_GPU_CASES_A = [
         DiffusionServerArgs(
             model_path="Wan-AI/Wan2.2-T2V-A14B-Diffusers",
             modality="video",
-            warmup=0,
             custom_validator="video",
             num_gpus=2,
+            warmup=True,
         ),
         DiffusionSamplingParams(
             prompt=T2V_PROMPT,
@@ -595,10 +589,10 @@ TWO_GPU_CASES_A = [
         DiffusionServerArgs(
             model_path="Wan-AI/Wan2.2-T2V-A14B-Diffusers",
             modality="video",
-            warmup=0,
             custom_validator="video",
             num_gpus=2,
             lora_path="Cseti/wan2.2-14B-Arcane_Jinx-lora-v1",
+            warmup=True,
         ),
         DiffusionSamplingParams(
             prompt="Nfj1nx with blue hair, a woman walking in a cyberpunk city at night",
@@ -608,10 +602,10 @@ TWO_GPU_CASES_A = [
         "wan2_1_t2v_14b_2gpu",
         DiffusionServerArgs(
             model_path="Wan-AI/Wan2.1-T2V-14B-Diffusers",
-            warmup=0,
             modality="video",
             num_gpus=2,
             custom_validator="video",
+            warmup=True,
         ),
         DiffusionSamplingParams(
             prompt=T2V_PROMPT,
@@ -623,10 +617,10 @@ TWO_GPU_CASES_A = [
         DiffusionServerArgs(
             model_path="Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
             modality="video",
-            warmup=0,
             custom_validator="video",
             num_gpus=2,
             cfg_parallel=True,
+            warmup=True,
         ),
         DiffusionSamplingParams(
             prompt=T2V_PROMPT,
@@ -642,10 +636,10 @@ if not current_platform.is_hip():
             DiffusionServerArgs(
                 model_path="IPostYellow/TurboWan2.2-I2V-A14B-Diffusers",
                 modality="video",
-                warmup=0,
                 custom_validator="video",
                 num_gpus=2,
                 tp_size=2,
+                warmup=True,
             ),
             TURBOWAN_I2V_sampling_params,
         )
@@ -656,10 +650,10 @@ TWO_GPU_CASES_B = [
         "wan2_1_i2v_14b_480P_2gpu",
         DiffusionServerArgs(
             model_path="Wan-AI/Wan2.1-I2V-14B-480P-Diffusers",
-            warmup=0,
             modality="video",
             custom_validator="video",
             num_gpus=2,
+            warmup=True,
         ),
         TI2V_sampling_params,
     ),
@@ -669,10 +663,10 @@ TWO_GPU_CASES_B = [
         DiffusionServerArgs(
             model_path="Wan-AI/Wan2.1-I2V-14B-720P-Diffusers",
             modality="video",
-            warmup=0,
             custom_validator="video",
             num_gpus=2,
             lora_path="starsfriday/Wan2.1-Divine-Power-LoRA",
+            warmup=True,
         ),
         TI2V_sampling_params,
     ),
@@ -681,9 +675,9 @@ TWO_GPU_CASES_B = [
         DiffusionServerArgs(
             model_path="Wan-AI/Wan2.1-I2V-14B-720P-Diffusers",
             modality="video",
-            warmup=0,
             custom_validator="video",
             num_gpus=2,
+            warmup=True,
         ),
         TI2V_sampling_params,
     ),
@@ -696,6 +690,7 @@ TWO_GPU_CASES_B = [
             # test ring attn
             ulysses_degree=1,
             ring_degree=2,
+            warmup=True,
         ),
         T2I_sampling_params,
     ),
@@ -706,6 +701,7 @@ TWO_GPU_CASES_B = [
             modality="image",
             num_gpus=2,
             ulysses_degree=2,
+            warmup=True,
         ),
         T2I_sampling_params,
     ),
@@ -715,6 +711,7 @@ TWO_GPU_CASES_B = [
             model_path="black-forest-labs/FLUX.1-dev",
             modality="image",
             num_gpus=2,
+            warmup=True,
         ),
         T2I_sampling_params,
     ),
@@ -725,6 +722,7 @@ TWO_GPU_CASES_B = [
             modality="image",
             num_gpus=2,
             tp_size=2,
+            warmup=True,
         ),
         T2I_sampling_params,
     ),