• 容器中pytorch的cpu速度很慢，原因找到了

# import numpy as np
# import time
# import torch
# import os

# os.environ['OMP_NUM_THREADS'] = '8'  # 增加線程數
# torch.set_num_threads(8)  # 設置PyTorch線程數

# # 創建測試矩陣
# device = 'cuda' if torch.cuda.is_available() else 'cpu'
# print(f"使用設備: {device}")
# B = torch.randn(2000, 2000).to(device)  # 使用GPU

# # 預熱GPU
# if device == 'cuda':
#     _ = torch.svd(torch.randn(100,100).to(device))

# # 計時SVD
# start = time.time()
# with torch.no_grad():  # 禁用梯度計算
#     x = torch.svd(B)
# end = time.time()

# print(f"SVD耗時: {end - start:.2f}秒")

import time
import numpy as np
import torch
import os

def diagnose_torch_svd():
    print("=== PyTorch SVD性能診斷 ===")
    
    # 1. 檢查系統配置
    print("\n1. 系統配置:")
    print(f"PyTorch線程數: {torch.get_num_threads()}")
    print(f"PyTorch interop線程數: {torch.get_num_interop_threads()}")
    print(f"OMP_NUM_THREADS: {os.environ.get('OMP_NUM_THREADS', '未設置')}")
    print(f"MKL_NUM_THREADS: {os.environ.get('MKL_NUM_THREADS', '未設置')}")
    
    # 2. 創建測試數據
    print("\n2. 性能測試:")
    size = 1000
    numpy_array = np.random.randn(size, size).astype(np.float64)
    torch_tensor = torch.from_numpy(numpy_array.copy())
    
    # 3. NumPy基準
    start = time.time()
    U_np, s_np, Vt_np = np.linalg.svd(numpy_array, full_matrices=False)
    numpy_time = time.time() - start
    
    # 4. PyTorch測試
    start = time.time()
    U_pt, s_pt, Vt_pt = torch.svd(torch_tensor, some=True)
    torch_time = time.time() - start
    
    print(f"NumPy SVD: {numpy_time:.4f}s")
    print(f"PyTorch SVD: {torch_time:.4f}s")
    print(f"速度比: {torch_time/numpy_time:.2f}x")
    
    # 5. 優化建議
    print("\n3. 優化建議:")
    if torch_time > 2 * numpy_time:
        print("??  PyTorch SVD明顯慢于NumPy，建議:")
        print("   - 設置 torch.set_num_threads(4)")
        print("   - 檢查環境變量 OMP_NUM_THREADS 和 MKL_NUM_THREADS")
        print("   - 考慮對大型矩陣使用NumPy后端")
    else:
        print("? PyTorch SVD性能正常")

# 運行診斷
# diagnose_torch_svd()

# 優化后重新測試
print("\n" + "="*50)
print("應用優化后:")

# 應用優化
# os.environ['OMP_NUM_THREADS'] = '4'
# os.environ['MKL_NUM_THREADS'] = '4'
torch.set_num_threads(8)

diagnose_torch_svd()