26：【CUDA Tile / Blackwell】新架构报错 no kernel image / compute capability不匹配

# no kernel image错误示例
RuntimeError: CUDA error: no kernel image is available for execution on the device
CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect.
For debugging consider passing CUDA_LAUNCH_BLOCKING=1.

# compute capability不匹配错误示例
UserWarning: PyTorch was compiled with CUDA capability sm_80, sm_86, sm_89, sm_90, but your current GPU has CUDA capability sm_95. You might experience unexpected behavior or slower performance.

# 升级CUDA Toolkit（推荐使用CUDA 13.1+）
# 从NVIDIA官网下载并安装最新的CUDA Toolkit

# 升级PyTorch到支持Blackwell架构的版本
pip install torch --upgrade --index-url https://download.pytorch.org/whl/cu131

# 升级其他依赖库
pip install --upgrade torchvision torchaudio transformers diffusers

# 设置CUDA架构兼容性环境变量
export TORCH_CUDA_ARCH_LIST="8.0;8.6;8.9;9.0;9.5"

# 或在Windows系统中设置
# set TORCH_CUDA_ARCH_LIST=8.0;8.6;8.9;9.0;9.5

# 设置CUDA启动阻塞模式（用于调试）
export CUDA_LAUNCH_BLOCKING=1

# 克隆PyTorch源码
git clone --recursive https://github.com/pytorch/pytorch
cd pytorch

# 配置编译选项，指定目标架构
python setup.py build --cmake-args="-DCUDA_ARCH_LIST=9.5;9.0;8.9;8.6;8.0"

# 安装编译后的版本
pip install -e .

#!/usr/bin/env python3
"""
CUDA架构兼容性检测工具
"""
import torch
import subprocess
import json

def get_gpu_info():
    """获取GPU信息"""
    try:
        result = subprocess.run(
            ["nvidia-smi", "--query-gpu=name,compute_capability", "--format=json"],
            capture_output=True,
            text=True,
            check=True
        )
        return json.loads(result.stdout)['gpu']
    except Exception as e:
        print(f"获取GPU信息失败: {e}")
        return []

def get_pytorch_cuda_info():
    """获取PyTorch CUDA信息"""
    try:
        import torch
        return {
            "pytorch_version": torch.__version__,
            "cuda_version": torch.version.cuda,
            "cuda_available": torch.cuda.is_available(),
            "device_count": torch.cuda.device_count(),
            "current_device": torch.cuda.current_device() if torch.cuda.is_available() else None,
            "device_name": torch.cuda.get_device_name(0) if torch.cuda.is_available() else None
        }
    except Exception as e:
        print(f"获取PyTorch CUDA信息失败: {e}")
        return {}

def check_compatibility(gpu_info, pytorch_info):
    """检查兼容性"""
    compatibility_report = []
    
    for gpu in gpu_info:
        gpu_name = gpu['name']
        compute_capability = gpu['compute_capability']
        
        # 解析compute capability
        major, minor = map(int, compute_capability.split('.'))
        cc_value = major * 10 + minor
        
        # 检查PyTorch支持情况
        if pytorch_info.get('cuda_available', False):
            # 简单的兼容性检查逻辑
            is_compatible = cc_value <= 95  # 假设最高支持sm_95
            
            report = {
                "gpu_name": gpu_name,
                "compute_capability": compute_capability,
                "is_compatible": is_compatible,
                "recommendation": ""
            }
            
            if not is_compatible:
                report["recommendation"] = "您的GPU架构较新，建议使用最新版本的PyTorch和CUDA Toolkit"
            elif cc_value >= 90:  # Blackwell架构
                report["recommendation"] = "您的GPU支持CUDA Tile技术，建议使用CUDA 13.1+以获得最佳性能"
            elif cc_value >= 80:  # Ampere架构
                report["recommendation"] = "您的GPU支持CUDA 11.0+，建议使用PyTorch 2.0+以获得最佳性能"
            else:
                report["recommendation"] = "您的GPU架构较老，可能无法充分利用最新的CUDA特性"
            
            compatibility_report.append(report)
    
    return compatibility_report

def main():
    """主函数"""
    print("=== CUDA架构兼容性检测工具 ===")
    
    # 获取GPU信息
    print("\n1. 获取GPU信息...")
    gpu_info = get_gpu_info()
    for gpu in gpu_info:
        print(f"  - {gpu['name']} (Compute Capability: {gpu['compute_capability']})")
    
    # 获取PyTorch信息
    print("\n2. 获取PyTorch CUDA信息...")
    pytorch_info = get_pytorch_cuda_info()
    for key, value in pytorch_info.items():
        print(f"  - {key}: {value}")
    
    # 检查兼容性
    print("\n3. 兼容性检查结果...")
    compatibility_report = check_compatibility(gpu_info, pytorch_info)
    for report in compatibility_report:
        print(f"\n  GPU: {report['gpu_name']}")
        print(f"  Compute Capability: {report['compute_capability']}")
        print(f"  兼容性: {'兼容' if report['is_compatible'] else '不兼容'}")
        print(f"  建议: {report['recommendation']}")
    
    print("\n4. 总结...")
    if not gpu_info:
        print("  未检测到GPU，请检查NVIDIA驱动是否正确安装")
    elif not pytorch_info.get('cuda_available', False):
        print("  PyTorch CUDA不可用，请检查CUDA安装和PyTorch版本")
    else:
        all_compatible = all(report['is_compatible'] for report in compatibility_report)
        if all_compatible:
            print("  所有GPU均兼容当前PyTorch版本")
        else:
            print("  部分GPU可能存在兼容性问题，请参考上述建议")

if __name__ == "__main__":
    main()

# CUDA Tile技术应用示例
import torch

# 检查是否支持CUDA Tile
def check_cuda_tile_support():
    if not torch.cuda.is_available():
        return False
    
    # 获取GPU架构
    device = torch.device("cuda")
    arch = torch.cuda.get_device_capability(device)
    
    # Blackwell架构(9.5+)支持CUDA Tile
    return arch[0] >= 9 and arch[1] >= 5

# 使用CUDA Tile优化的矩阵乘法
def optimized_matmul():
    if not check_cuda_tile_support():
        print("CUDA Tile not supported on this GPU")
        return
    
    # 创建大矩阵
    a = torch.randn(10240, 10240, device="cuda")
    b = torch.randn(10240, 10240, device="cuda")
    
    # 预热
    torch.matmul(a, b)
    torch.cuda.synchronize()
    
    # 测试性能
    import time
    start = time.time()
    c = torch.matmul(a, b)
    torch.cuda.synchronize()
    end = time.time()
    
    print(f"Matrix multiplication time: {end - start:.4f} seconds")
    print(f"Performance: {((10240**3) * 2 / (end - start)) / 1e12:.2f} TFLOPS")

if __name__ == "__main__":
    print(f"CUDA Tile support: {check_cuda_tile_support()}")
    optimized_matmul()

#!/bin/bash

# 检查当前环境
echo "=== 检查当前环境 ==="
nvidia-smi

# 检查PyTorch版本
python -c "import torch; print('PyTorch版本:', torch.__version__); print('CUDA版本:', torch.version.cuda); print('CUDA可用:', torch.cuda.is_available())"

# 设置环境变量
echo "\n=== 设置环境变量 ==="
export TORCH_CUDA_ARCH_LIST="8.0;8.6;8.9;9.0;9.5"
export CUDA_LAUNCH_BLOCKING=1
echo "TORCH_CUDA_ARCH_LIST=$TORCH_CUDA_ARCH_LIST"
echo "CUDA_LAUNCH_BLOCKING=$CUDA_LAUNCH_BLOCKING"

# 测试CUDA功能
echo "\n=== 测试CUDA功能 ==="
python -c "
import torch
if torch.cuda.is_available():
    # 测试基本功能
    x = torch.randn(1000, 1000, device='cuda')
    y = torch.randn(1000, 1000, device='cuda')
    z = torch.matmul(x, y)
    print('基本CUDA功能测试通过')
    
    # 测试CUDA Tile支持
    arch = torch.cuda.get_device_capability()
    print(f'GPU架构: {arch}')
    if arch[0] >= 9 and arch[1] >= 5:
        print('CUDA Tile技术支持: 是')
    else:
        print('CUDA Tile技术支持: 否')
else:
    print('CUDA不可用')
"

echo "\n=== 配置完成 ==="