scripts/batch_inference.py

#!/usr/bin/env python3
"""
批量推理: 11 张图片 × 随机 WASD 方向变换
每 ~30 latents (~120 帧) 变换一个方向
视频长度 497 帧 (~125 latents, ~20s @24fps)
"""
import random
import subprocess
import os
import time

random.seed(42)

IMAGES = [
    "1.png", "2.png", "3.png", "4.png", "5.png",
    "6.jpeg", "7.png", "8.png", "9.png", "10.png", "test.png"
]

DIRECTIONS = ["w", "a", "s", "d"]
# 加入转向动作
ALL_MOVES = ["w", "a", "s", "d", "left", "right", "up", "down"]

IMAGE_DIR = "/root/HY-WorldPlay/assets/img"
OUTPUT_BASE = "/root/test_results/batch"

MODEL_PATH = subprocess.check_output(
    "find /root/models -maxdepth 3 -name 'HunyuanVideo*' -type d | grep -v temp | head -1",
    shell=True
).decode().strip()

WP_PATH = subprocess.check_output(
    "find /root/models -maxdepth 3 -name 'HY-WorldPlay' -type d | grep -v temp | head -1",
    shell=True
).decode().strip()

ACTION_CKPT = f"{WP_PATH}/ar_distilled_action_model/diffusion_pytorch_model.safetensors"


def generate_random_pose(total_latents=31, segment_latents=8):
    """生成随机 WASD 方向序列，每 segment_latents 变换一次"""
    segments = []
    remaining = total_latents
    while remaining > 0:
        direction = random.choice(DIRECTIONS)
        # 随机混入转向
        if random.random() < 0.3:
            turn = random.choice(["left", "right"])
            turn_len = min(random.randint(2, 4), remaining)
            segments.append(f"{turn}-{turn_len}")
            remaining -= turn_len
            if remaining <= 0:
                break
        seg_len = min(segment_latents + random.randint(-2, 2), remaining)
        if seg_len <= 0:
            break
        segments.append(f"{direction}-{seg_len}")
        remaining -= seg_len
    return ",".join(segments)


def run_inference(image_name, pose, output_dir):
    """运行单次推理"""
    image_path = os.path.join(IMAGE_DIR, image_name)
    
    cmd = [
        "python3", "/root/scripts/run_fp8_turbo3_gpu.py",
        "--model_path", MODEL_PATH,
        "--action_ckpt", ACTION_CKPT,
        "--prompt", "Explore a vivid 3D world with smooth camera movement.",
        "--image_path", image_path,
        "--resolution", "480p",
        "--aspect_ratio", "16:9",
        "--video_length", "125",
        "--seed", str(random.randint(0, 99999)),
        "--rewrite", "false",
        "--sr", "false",
        "--pose", pose,
        "--output_path", output_dir,
        "--few_step", "true",
        "--num_inference_steps", "4",
        "--model_type", "ar",
        "--use_vae_parallel", "false",
        "--use_sageattn", "true",
        "--use_fp8_gemm", "false",
        "--transformer_resident_ar_rollout", "true",
        "--width", "832",
        "--height", "480",
    ]
    
    env = os.environ.copy()
    env["PYTHONPATH"] = "/root/HY-WorldPlay:" + env.get("PYTHONPATH", "")
    env["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"
    
    result = subprocess.run(cmd, env=env, cwd="/root/HY-WorldPlay",
                          capture_output=True, text=True)
    return result.returncode, result.stdout, result.stderr


if __name__ == "__main__":
    os.makedirs(OUTPUT_BASE, exist_ok=True)
    
    print("=" * 60)
    print("批量推理: 11 张图片 × 随机 WASD")
    print("=" * 60)
    
    results = []
    
    for i, img in enumerate(IMAGES):
        pose = generate_random_pose(total_latents=31, segment_latents=8)
        output_dir = os.path.join(OUTPUT_BASE, img.split(".")[0])
        os.makedirs(output_dir, exist_ok=True)
        
        print(f"\n[{i+1}/11] {img}")
        print(f"  pose: {pose}")
        print(f"  output: {output_dir}")
        
        t0 = time.time()
        code, stdout, stderr = run_inference(img, pose, output_dir)
        elapsed = time.time() - t0
        
        # 提取关键信息
        success = code == 0
        peak_mem = ""
        total_time = ""
        for line in (stdout + stderr).split("\n"):
            if "峰值显存" in line:
                peak_mem = line.strip()
            if "总耗时" in line:
                total_time = line.strip()
        
        status = "✅" if success else "❌"
        print(f"  {status} {elapsed:.0f}s | {peak_mem} | {total_time}")
        
        if not success:
            # 打印最后 10 行错误
            err_lines = stderr.strip().split("\n")[-10:]
            for l in err_lines:
                print(f"  ERR: {l}")
        
        results.append({
            "image": img,
            "pose": pose,
            "success": success,
            "time": elapsed,
        })
    
    # 汇总
    print("\n" + "=" * 60)
    print("汇总")
    print("=" * 60)
    ok = sum(1 for r in results if r["success"])
    print(f"成功: {ok}/11")
    print(f"总耗时: {sum(r['time'] for r in results):.0f}s")
    for r in results:
        s = "✅" if r["success"] else "❌"
        print(f"  {s} {r['image']:12s} {r['time']:.0f}s  pose={r['pose']}")
    
    print(f"\n视频输出目录: {OUTPUT_BASE}/")
    print("完成!")