如何在Windows 10上用NVIDIA Quadro K1200 GPU运行TensorFlow？

设备配置 #

• Windows 10 22H2
• Windows原生环境安装Python 3.11和3.10，WSL中安装Python 3.8
• TensorFlow 2.12
• Intel i7-6700 CPU
• 32GB DDR4内存
• NVIDIA Quadro K1200 GPU（显卡驱动版本528.89）

我使用的代码在一个Jupyter Notebook中。

已尝试的操作及问题 #

1. Windows原生环境安装tensorflow-directml-plugin后，运行代码出现UnimplementedError: Graph execution error.（对应Notebook第20单元格）
2. WSL环境按文档安装CUDA Toolkit和TensorFlow，执行验证命令时输出如下：

(condaenv) root@DESKTOP-CHV5AM:/home/admin/Python# python3 -c "import tensorflow as tf; print(tf.config.list_physical_devices('GPU'))"
2023-06-15 12:52:11.886160: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
To enable the following instructions: AVX2 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
2023-06-15 12:52:12.493026: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Could not find TensorRT
2023-06-15 12:52:13.151393: I tensorflow/compiler/xla/stream_executor/cuda/cuda_gpu_executor.cc:982] could not open file to read NUMA node: /sys/bus/pci/devices/0000:01:00.0/numa_node
Your kernel may have been built without NUMA support.
2023-06-15 12:52:13.174058: I tensorflow/compiler/xla/stream_executor/cuda/cuda_gpu_executor.cc:982] could not open file to read NUMA node: /sys/bus/pci/devices/0000:01:00.0/numa_node
Your kernel may have been built without NUMA support.
2023-06-15 12:52:13.174904: I tensorflow/compiler/xla/stream_executor/cuda/cuda_gpu_executor.cc:982] could not open file to read NUMA node: /sys/bus/pci/devices/0000:01:00.0/numa_node
Your kernel may have been built without NUMA support.
[PhysicalDevice(name='/physical_device:GPU:0', device_type='GPU')]

解决方案 #

针对WSL环境的提示 #

你的WSL环境已经成功识别GPU（输出最后一行的[PhysicalDevice(name='/physical_device:GPU:0', device_type='GPU')]），其余输出均为非致命提示：

• NUMA相关信息只是说明WSL内核未启用NUMA支持，完全不影响GPU训练功能
• TF-TRT警告是因为未安装TensorRT插件，仅影响TensorRT加速，不影响基础GPU训练

直接按以下步骤运行训练即可：

1. 激活你的conda环境，安装Jupyter：

pip install jupyter

2. 启动Jupyter Notebook，加载你的代码文件：

jupyter notebook

3. 运行代码前，可添加以下代码确认GPU是否在被使用：

import tensorflow as tf
print("GPU可用状态:", tf.test.is_gpu_available())
with tf.device('/GPU:0'):
    # 测试GPU计算
    a = tf.constant([1.0, 2.0, 3.0])
    b = tf.constant([4.0, 5.0, 6.0])
    print("GPU计算结果:", a + b)

针对Windows原生环境的UnimplementedError #

这个错误是因为tensorflow-directml-plugin对部分TensorFlow算子支持不完善，建议优先使用WSL环境。如果坚持用Windows原生：

• 检查代码中是否有自定义算子、小众层或新版本特有的操作，替换为directml兼容的实现
• 降级TensorFlow到2.10版本（该版本与directml插件兼容性更好），重新安装插件后测试

内容的提问来源于stack exchange，提问作者SOuser