[Bugfix] Fix default sampling params for /v1/videos on main

[david6666666]

Summary

• initialize /v1/videos and /v1/videos/sync sampling params from the engine stage defaults before request overrides
• treat frame interpolation form fields as optional so omitted request fields no longer overwrite --default-sampling-params
• add regression tests for async and sync video paths covering default propagation and explicit per-request overrides

Validation

• pre-commit run --all-files
• pytest -q tests/entrypoints/openai_api/test_video_server.py -k 'default_sampling_params_apply_to_video_requests or request_params_override_default_video_sampling_params or sync_default_sampling_params_apply_to_video_requests or frame_interpolation_params_pass_to_diffusion_sampling_params or sync_frame_interpolation_params_pass_to_sampling_params or sync_sampling_params_pass_through'
• manual E2E with vllm 0.19.0 on Wan-AI/Wan2.2-I2V-A14B-Diffusers using:
  • CUDA_VISIBLE_DEVICES=5,6 vllm serve ... --omni --port 8099 --enable-diffusion-pipeline-profiler --ulysses-degree 2 --default-sampling-params '{"0":{"enable_frame_interpolation":true,"frame_interpolation_exp":1,"frame_interpolation_scale":1.0,"frame_interpolation_model_path":"/mnt/data1/huggingface/hub/models--elfgum--RIFE-4.22.lite/snapshots/99d6892a9f4c039cb37ff21c9530e79b13f0b30b"}}'
  • python curl_output/curl_w22_normal.py
  • result: request completed successfully, server logged steps=8 guidance=3.5 guidance_2=3.5 seed=42, and output MP4 had average_rate=32.0 for request fps=16, confirming default frame interpolation still applied

Fixes #3017.

[david6666666]

CUDA_VISIBLE_DEVICES=5,6 vllm serve /mnt/data1/huggingface/hub/models--Wan-AI--Wan2.2-I2V-A14B-Diffusers/snapshots/596658fd9ca6b7b71d5057529bbf319ecbc61d74 --omni --port 8099 --enable-diffusion-pipeline-profiler --ulysses-degree 2 --default-sampling-params '{"0":{"enable_frame_interpolation":true,"frame_interpolation_exp":1,"frame_interpolation_scale":1.0,"frame_interpolation_model_path":"/mnt/data1/huggingface/hub/models--elfgum--RIFE-4.22.lite/snapshots/99d6892a9f4c039cb37ff21c9530e79b13f0b30b"}}'

import json, os, pathlib, time, requests, sys
base = 'http://127.0.0.1:8099'
image_path = '/mnt/data4/cwq/tmp/rabbit_real.png'
out_path = '/mnt/data4/cwq/tmp/wan22_bench_out.mp4'
prompt = '一只棕色野兔的正面特写镜头，采用低角度仰拍视角，营造亲密而庄严的视觉冲击。兔子一双圆润漆黑的大眼睛直视镜头深处，眼神中交织着野生动物的警觉与一丝难以言喻的温柔好奇，仿佛在与观者建立跨越物种的静默对话。它毛色呈现层次丰富的棕褐渐变，从浅奶油色腹部过渡到深棕背部，每根毛发纹理清晰可辨，在侧光下泛着丝绸般的光泽。细长洁白的胡须共有三对，随呼吸节奏微微颤动，偶尔因捕捉气流信息而轻轻摇摆。\n两只标志性的长耳完全竖立，耳廓外侧覆盖短密棕毛，内侧则露出粉嫩的血管网络，薄如蝉翼的皮肤下血液流动隐约可见，耳朵以细微幅度不时转动，精准定位声源方向。背景是一片澄澈的蔚蓝天空，形态蓬松的白色积云以缓慢速度横向漂移，云影在兔子头顶交替变化，光线随之明暗流转。晴朗天气的明媚阳光从画面左上方45度角倾泻而下，在兔脸右侧形成柔和的伦勃朗式阴影，强化了面部立体感和皮毛质感。\n兔子湿润的黑鼻子持续进行每秒三至四次的快速抽动，这是它们感知化学信号的本能动作，粉色三瓣嘴随之轻启，露出正在反刍的洁白门齿，下颌以稳定节奏左右研磨。摄影采用大光圈浅景深，焦点牢牢锁定在兔子双眼连线所在的焦平面，背景天空和远景绿色植被虚化成圆润的彩色光斑，前景几根嫩绿草叶闯入画面边缘，以缓慢弧线随风摇曳，暗示着和煦的春日微风。\n整个场景弥漫着宁静致远的田园诗意，色彩温暖饱和，充满生命力。兔子在持续五秒的对视后，以典型 lagomorph 特征完成一次完整的瞬膜眨眼——第三眼睑从内侧横向滑过眼球，继而缓缓歪头向右十五度，这个行为在动物行为学中代表认知加工和好奇心表达，耳朵随之向同一方向倾斜，最终恢复正视姿态，胡须舒展，完成这段短暂而珍贵的自然纪录。'
negative = '色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走'

data = {'prompt': prompt, 'negative_prompt': negative, 'size': '1280x720', 'seconds': '5', 'fps': '16', 'num_inference_steps': '8', 'guidance_scale': '3.5', 'guidance_scale_2': '3.5', 'boundary_ratio': '0.875', 'num_frames': '81', 'flow_shift': '5.0', 'seed': '42'}

with open(image_path, 'rb') as img:
    t0 = time.perf_counter()
    create = requests.post(f'{base}/v1/videos', headers={'Accept': 'application/json'}, data=data, files={'input_reference': (os.path.basename(image_path), img, 'image/png')}, timeout=1200)
    
create.raise_for_status()
video_id = create.json()['id']

while True:
    resp = requests.get(f'{base}/v1/videos/{video_id}', timeout=30)
    resp.raise_for_status()
    payload = resp.json()
    if payload.get('status') in ('completed', 'failed'):
        final_json = payload
        break
    time.sleep(1.0)

download = requests.get(f'{base}/v1/videos/{video_id}/content', timeout=1200)
download.raise_for_status()
pathlib.Path(out_path).write_bytes(download.content)
t3 = time.perf_counter()

result = {
    'artifact_ready_wall_s': round(t3 - t0, 3),
    'server_inference_time_s': final_json.get('inference_time_s'),
    'final_status': final_json.get('status'),
}
print(json.dumps(result, ensure_ascii=False, indent=2))

wan22_bench_out.mp4