【问题标题】：Anaconda (Win10) Tensorflow not recognizing GPU suddenly (CUDA_ERROR_NO_DEVICE)Anaconda（Win10）Tensorflow突然无法识别GPU（CUDA_ERROR_NO_DEVICE）
【发布时间】：2021-03-10 18:04:30
【问题描述】：

我通过 Anaconda 安装了 tensorflow。它运行良好，并且已经识别了 GPU 一段时间。但是突然之间，从几天前开始，没有任何带有 tensorflow 的环境可以识别我的 GPU。有人知道要检查什么吗？

我尝试过的：

使用 python=3.7 创建了一个全新的环境并安装了 tensorflow-gpu=2.1
重新安装 anaconda
使用 python=3.6 创建了一个全新的环境并安装了 tensorflow-gpu=1.9
安装了 tensorflow-gpu=2.3 并安装了缺少的 cudatoolkit=10.1 和 cudnn=7.6
根据open github issue 安装了具有特定内部版本号的 tensorflow-gpu
我通过python将环境变量CUDA_VISIBLE_DEVICES设置为0（TensorFlow : failed call to cuInit: CUDA_ERROR_NO_DEVICE）
我更新了我的图形驱动程序
删除了修改后的注册表项HKEY_LOCAL_MACHINE\SYSTEM\CurrentControlSet\Control\GraphicsDrivers\TdrDelay

我的测试脚本来检查识别的设备：

import tensorflow as tf
from tensorflow.python.client import device_lib

print(device_lib.list_local_devices())

这是我在每个配置中得到的输出：

> python check.py
2021-03-10 18:48:12.880629: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library cudart64_101.dll
2021-03-10 18:48:14.637784: I tensorflow/core/platform/cpu_feature_guard.cc:142] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX AVX2
2021-03-10 18:48:19.201572: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library nvcuda.dll
2021-03-10 18:48:19.705910: E tensorflow/stream_executor/cuda/cuda_driver.cc:351] failed call to cuInit: CUDA_ERROR_NO_DEVICE: no CUDA-capable device is detected
2021-03-10 18:48:19.715756: I tensorflow/stream_executor/cuda/cuda_diagnostics.cc:169] retrieving CUDA diagnostic information for host: NB-170
2021-03-10 18:48:19.721085: I tensorflow/stream_executor/cuda/cuda_diagnostics.cc:176] hostname: NB-170
[name: "/device:CPU:0"
device_type: "CPU"
memory_limit: 268435456
locality {
}
incarnation: 10539449374211484676
]

系统信息

操作系统：Windows 10 Pro（版本 10.0.18363 Build 18363）
显卡：NVIDIA GeForce GTX 1650
Anaconda 1.10
将注册表：HKEY_LOCAL_MACHINE\SYSTEM\CurrentControlSet\Control\GraphicsDrivers\TdrDelay 更改为 15 以训练 Matterport 的 mask r-cnn 实现
图形驱动程序 - GEFORCE GAME READY 驱动程序 - 版本：461.72 WHQL；发售日期：2021.2.25；操作系统：Windows 10 64 位；语言：英语

我的 nvdia-smi 输出：

+-----------------------------------------------------------------------------+
| NVIDIA-SMI 461.72       Driver Version: 461.72       CUDA Version: 11.2     |
|-------------------------------+----------------------+----------------------+
| GPU  Name            TCC/WDDM | Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|                               |                      |               MIG M. |
|===============================+======================+======================|
|   0  GeForce GTX 1650   WDDM  | 00000000:01:00.0 Off |                  N/A |
| N/A   54C    P8     6W /  N/A |    132MiB /  4096MiB |      0%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+

+-----------------------------------------------------------------------------+
| Processes:                                                                  |
|  GPU   GI   CI        PID   Type   Process name                  GPU Memory |
|        ID   ID                                                   Usage      |
|=============================================================================|
|  No running processes found                                                 |
+-----------------------------------------------------------------------------+

更新 1 (2021-03-14)

我安装了全新的 Anaconda 安装并创建了一个环境（conda create -name tf-gpu tensorflow-gpu=2.1 在我拥有的另一台计算机上。在那台计算机上我的 gpu 被识别，没有任何问题。

2021-03-14 14:21:33.934222: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library cudart64_101.dll
2021-03-14 14:21:37.608844: I tensorflow/core/platform/cpu_feature_guard.cc:142] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX
2021-03-14 14:21:37.612173: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library nvcuda.dll
2021-03-14 14:21:37.658982: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1555] Found device 0 with properties: 
pciBusID: 0000:01:00.0 name: GeForce GTX 970 computeCapability: 5.2
coreClock: 1.253GHz coreCount: 13 deviceMemorySize: 4.00GiB deviceMemoryBandwidth: 208.91GiB/s
2021-03-14 14:21:37.659525: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library cudart64_101.dll
2021-03-14 14:21:38.216002: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library cublas64_10.dll
2021-03-14 14:21:38.625300: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library cufft64_10.dll
2021-03-14 14:21:38.660856: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library curand64_10.dll
2021-03-14 14:21:38.971988: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library cusolver64_10.dll
2021-03-14 14:21:39.247585: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library cusparse64_10.dll
2021-03-14 14:21:39.564512: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library cudnn64_7.dll
2021-03-14 14:21:39.565268: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1697] Adding visible gpu devices: 0
2021-03-14 14:21:41.272007: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1096] Device interconnect StreamExecutor with strength 1 edge matrix:
2021-03-14 14:21:41.272272: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1102]      0 
2021-03-14 14:21:41.272582: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1115] 0:   N
2021-03-14 14:21:41.283835: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1241] Created TensorFlow device (/device:GPU:0 with 2993 MB memory) -> physical GPU (device: 0, name: GeForce GTX 970, pci bus id: 0000:01:00.0, compute capability: 5.2)
[name: "/device:CPU:0"
device_type: "CPU"
memory_limit: 268435456
locality {
}
incarnation: 17009642916451828901
, name: "/device:GPU:0"
device_type: "GPU"
memory_limit: 3139148187
locality {
  bus_id: 1
  links {
  }
}
incarnation: 5677250807137925801
physical_device_desc: "device: 0, name: GeForce GTX 970, pci bus id: 0000:01:00.0, compute capability: 5.2"
]

【问题讨论】：

你安装的是什么版本的nvidia驱动？
GEFORCE GAME READY 驱动程序：版本：461.72 WHQL；发售日期：2021.2.25；操作系统：Windows 10 64 位；语言：英语
对于 Cuda 11.2 你需要 cuDNN 8.0，你可以在这里找到列表developer.nvidia.com/rdp/cudnn-archive
不要忘记将它们包含在您的 PATH 中
我正在通过 Anaconda 安装 Cuda 和 cuDNN。如果我正确阅读了我的输出，则 cuda ist 加载正常。所以这不应该是我的路径错误。 nvidia-smi 输出中显示的 Cuda 版本不是已安装的版本，而是图形驱动程序支持的最高版本。所以我的图形驱动程序支持高达 11.2 版的 Cuda。（请参阅此问题的第一条评论：stackoverflow.com/q/53422407/4295853）

标签： tensorflow anaconda

【解决方案1】：

就我而言，我遇到了同样的错误：failed call to cuinit: CUDA_ERROR_NO_DEVICE。但是 nvidia-smi.exe 正在检测 gpu。我的系统（Windows 10）中安装了 CUDA 9.0。然后我意识到我的应用程序路径中不小心有一个 CUDA 10.0 版本的 dll nvcuda.dll。从我的应用程序路径中删除这个 dll 解决了这个问题。

【讨论】：