简单的一个测试脚本，用来测硬件做深度学习训练时的性能。

# train_simple.py
import torch
import time
from diffusers import UNet2DModel
from pathlib import Path
import platform
import psutil

# -----------------------------
# 1. 系统信息打印
# -----------------------------
def print_system_info():
    print("? 系统信息检测中...")

    cpu_name = platform.processor()
    if not cpu_name:
        try:
            with open("/proc/cpuinfo", "r") as f:
                for line in f:
                    if "model name" in line:
                        cpu_name = line.split(":")[1].strip().split(" @ ")[0]
                        break
        except:
            cpu_name = "Unknown CPU"
    cpu_cores = psutil.cpu_count(logical=False)
    cpu_threads = psutil.cpu_count(logical=True)
    print(f"?  CPU: {cpu_name} ({cpu_cores} 核 / {cpu_threads} 线程)")

    memory = psutil.virtual_memory()
    total_ram = memory.total / (1024**3)
    available_ram = memory.available / (1024**3)
    print(f"? 内存: {total_ram:.2f} GB 总计, {available_ram:.2f} GB 可用")

    if torch.cuda.is_available():
        gpu_count = torch.cuda.device_count()
        print(f"? GPU: {gpu_count} 张可用 (CUDA)")
        for i in range(gpu_count):
            gpu_name = torch.cuda.get_device_name(i)
            print(f"   ? [{i}] {gpu_name}")
    else:
        print("? GPU: 未检测到 (CUDA 不可用)")

    print("-" * 50)

print_system_info()

print("? 开始执行 UNet2DModel 分阶段训练测试...")

# -----------------------------
# 2. 全局计时开始
# -----------------------------
total_start_time = time.time()

# 创建模型（不指定设备）
model = UNet2DModel(
    sample_size=32,
    in_channels=3,
    out_channels=3,
    layers_per_block=1,
    block_out_channels=(32, 64),
    down_block_types=("DownBlock2D", "AttnDownBlock2D"),
    up_block_types=("UpBlock2D", "AttnUpBlock2D"),
)

# 模拟数据
batch_size = 2
x_cpu = torch.randn(batch_size, 3, 32, 32)
y_cpu = torch.randn(batch_size, 3, 32, 32)
timesteps_cpu = torch.zeros(batch_size, dtype=torch.long)

# 显示模型参数量
def count_params(m):
    return sum(p.numel() for p in m.parameters() if p.requires_grad)

print(f"? 模型参数量: {count_params(model):,}")

# 优化器（后面会重建）
optimizer = None

# -----------------------------
# 3. 阶段1：CPU 训练
# -----------------------------
print("? 开始 CPU 训练...")
cpu_start_time = time.time()

# 使用 CPU
model_cpu = UNet2DModel(
    sample_size=32,
    in_channels=3,
    out_channels=3,
    layers_per_block=1,
    block_out_channels=(32, 64),
    down_block_types=("DownBlock2D", "AttnDownBlock2D"),
    up_block_types=("UpBlock2D", "AttnUpBlock2D"),
)
model_cpu.train()
model_cpu = model_cpu.to("cpu")

optimizer = torch.optim.Adam(model_cpu.parameters(), lr=1e-3)
n_steps = 10

for step in range(n_steps):
    optimizer.zero_grad()
    output = model_cpu(x_cpu, timesteps_cpu).sample
    loss = torch.nn.functional.mse_loss(output, y_cpu)
    loss.backward()
    optimizer.step()
    print(f"Step {step+1:2d} | Device: CPU | Loss: {loss.item():.6f}")

cpu_end_time = time.time()
cpu_elapsed = cpu_end_time - cpu_start_time
print(f"⏱️  CPU 训练耗时: {cpu_elapsed:.2f} 秒")

# -----------------------------
# 4. 阶段2：GPU 训练（如果可用）
# -----------------------------
gpu_elapsed = 0.0
if torch.cuda.is_available():
    print("? 开始 GPU 训练...")
    gpu_start_time = time.time()

    model_gpu = UNet2DModel(
        sample_size=32,
        in_channels=3,
        out_channels=3,
        layers_per_block=1,
        block_out_channels=(32, 64),
        down_block_types=("DownBlock2D", "AttnDownBlock2D"),
        up_block_types=("UpBlock2D", "AttnUpBlock2D"),
    )
    model_gpu.train()
    device = torch.device("cuda")
    model_gpu = model_gpu.to(device)

    x_gpu = x_cpu.to(device)
    y_gpu = y_cpu.to(device)
    timesteps_gpu = timesteps_cpu.to(device)
    optimizer = torch.optim.Adam(model_gpu.parameters(), lr=1e-3)

    for step in range(n_steps):
        optimizer.zero_grad()
        output = model_gpu(x_gpu, timesteps_gpu).sample
        loss = torch.nn.functional.mse_loss(output, y_gpu)
        loss.backward()
        optimizer.step()
        print(f"Step {step+1:2d} | Device: GPU | Loss: {loss.item():.6f}")

    gpu_end_time = time.time()
    gpu_elapsed = gpu_end_time - gpu_start_time
    print(f"⏱️  GPU 训练耗时: {gpu_elapsed:.2f} 秒")
else:
    print("⚠️  GPU 不可用，跳过 GPU 训练阶段")

# -----------------------------
# 5. 保存模型（仅保存最后一次训练的模型）
# -----------------------------
save_dir = Path("output") / "test-unet-final"
save_dir.mkdir(parents=True, exist_ok=True)

# 保存 GPU 模型优先，否则保存 CPU
final_model = model_gpu if torch.cuda.is_available() else model_cpu
final_model.eval()
final_model.save_pretrained(save_dir)
print(f"? 模型已保存至: {save_dir.resolve()}")

# 清理命令
if platform.system() == "Windows":
    delete_cmd = f'rd /s /q "{save_dir}"'
else:
    delete_cmd = f'rm -rf "{save_dir}"'
print(f"? 清理命令（如需删除）: {delete_cmd}")

# -----------------------------
# 6. 总耗时统计
# -----------------------------
total_end_time = time.time()
total_elapsed = total_end_time - total_start_time
print(f"✅ 总训练耗时: {total_elapsed:.2f} 秒")
print(f"? 详细耗时: CPU={cpu_elapsed:.2f}s, GPU={gpu_elapsed:.2f}s")
print("? 所有训练阶段完成！环境一切正常，可开始真实训练。")

20250903...