mar*_*lon 5 cuda conda tensorflow2.0
我的当前：
nvidia-smi
Wed Aug 4 01:40:39 2021
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 410.79 Driver Version: 410.79 CUDA Version: 10.0 |
|-------------------------------+----------------------+----------------------+
| GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
|===============================+======================+======================|
| 0 Tesla V100-SXM2... On | 00000000:00:0C.0 Off | 0 |
| N/A 34C P0 37W / 300W | 0MiB / 16130MiB | 0% Default |
+-------------------------------+----------------------+----------------------+
| 1 Tesla V100-SXM2... On | 00000000:00:0D.0 Off | 0 |
| N/A 34C P0 36W / 300W | 0MiB / 16130MiB | 0% Default |
+-------------------------------+----------------------+----------------------+
| 2 Tesla V100-SXM2... On | 00000000:00:0E.0 Off | 0 |
| N/A 33C P0 39W / 300W | 0MiB / 16130MiB | 0% Default |
+-------------------------------+----------------------+----------------------+
| 3 Tesla V100-SXM2... On | 00000000:00:0F.0 Off | 0 |
| N/A 37C P0 41W / 300W | 0MiB / 16130MiB | 0% Default |
+-------------------------------+----------------------+----------------------+
+-----------------------------------------------------------------------------+
| Processes: GPU Memory |
| GPU PID Type Process name Usage |
|=============================================================================|
| No running processes found |
+-----------------------------------------------------------------------------+
我想安装Tensorflow 2.3/2.4，所以我需要在Conda中至少将cuda升级到10.1。我知道如何在 conda 中安装 cudakit：
conda install cudatoolkit=10.1
但这似乎还不够：
Status: CUDA driver version is insufficient for CUDA runtime version
如果我想保留旧版本的cuda 10.0，可以通过Conda将cuda更新到10.1吗？这是行不通的：
conda install cuda=10.1
我正在使用Python 3.8。如果我不能保留cuda 10.0，如何在有或没有conda的情况下直接将cuda升级到10.1？如果能在Conda升级就最好了。
添加：
我安装了cudatoolkit=10.1，但是cuda驱动仍然不好。我的 conda 环境列表显示：
cudatoolkit 10.1.243 h6bb024c_0
tensorflow-gpu 2.3.0 pypi_0 pypi
下面的测试是好的：
import tensorflow as tf
2021-08-04 04:21:31.110443: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library libcudart.so.10.1
In [3]: print("Num GPUs Available: ", len(tf.config.list_physical_devices('GPU')))
2021-08-04 04:21:34.499432: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library libcuda.so.1
2021-08-04 04:21:34.665738: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:982] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-08-04 04:21:34.666369: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1716] Found device 0 with properties:
pciBusID: 0000:00:0c.0 name: Tesla V100-SXM2-16GB computeCapability: 7.0
coreClock: 1.53GHz coreCount: 80 deviceMemorySize: 15.75GiB deviceMemoryBandwidth: 836.37GiB/s
2021-08-04 04:21:34.666459: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:982] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-08-04 04:21:34.667017: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1716] Found device 1 with properties:
pciBusID: 0000:00:0d.0 name: Tesla V100-SXM2-16GB computeCapability: 7.0
coreClock: 1.53GHz coreCount: 80 deviceMemorySize: 15.75GiB deviceMemoryBandwidth: 836.37GiB/s
2021-08-04 04:21:34.667064: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:982] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-08-04 04:21:34.667613: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1716] Found device 2 with properties:
pciBusID: 0000:00:0e.0 name: Tesla V100-SXM2-16GB computeCapability: 7.0
coreClock: 1.53GHz coreCount: 80 deviceMemorySize: 15.75GiB deviceMemoryBandwidth: 836.37GiB/s
2021-08-04 04:21:34.667644: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library libcudart.so.10.1
2021-08-04 04:21:34.670275: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library libcublas.so.10
2021-08-04 04:21:34.672971: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library libcufft.so.10
2021-08-04 04:21:34.673378: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library libcurand.so.10
2021-08-04 04:21:34.676043: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library libcusolver.so.10
2021-08-04 04:21:34.677370: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library libcusparse.so.10
2021-08-04 04:21:34.681850: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library libcudnn.so.7
2021-08-04 04:21:34.681989: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:982] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-08-04 04:21:34.682604: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:982] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-08-04 04:21:34.683196: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:982] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-08-04 04:21:34.683782: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:982] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-08-04 04:21:34.684353: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:982] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-08-04 04:21:34.684961: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:982] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-08-04 04:21:34.685513: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1858] Adding visible gpu devices: 0, 1, 2
Num GPUs Available: 3
但以下测试失败：
import tensorflow as tf
with tf.device('/gpu:0'):
a = tf.constant([1.0, 2.0, 3.0, 4.0, 5.0, 6.0], shape=[2, 3], name='a')
b = tf.constant([1.0, 2.0, 3.0, 4.0, 5.0, 6.0], shape=[3, 2], name='b')
c = tf.matmul(a, b)
with tf.Session() as sess:
print (sess.run(c))
错误信息：
2021-08-04 04:27:30.934969: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1858] Adding visible gpu devices: 0, 1, 2
2021-08-04 04:27:30.935028: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library libcudart.so.10.1
---------------------------------------------------------------------------
InternalError Traceback (most recent call last)
......
InternalError: cudaGetDevice() failed. Status: CUDA driver version is insufficient for CUDA runtime version
如果这个说法属实，为什么我的安装还是不好，因为我已经在Conda中安装了cudatoolkit=10.1：
If you want to install a GPU driver, you could install a newer CUDA toolkit, which will have a newer GPU driver (installer) bundled with it.
cudatoolkit 和 cuda 驱动程序仍然不匹配？
不，您无法通过 conda 更新 GPU 驱动程序，而这正是您的情况需要支持 CUDA 10.1 或更新版本的驱动程序。看这里：
Anaconda 要求用户安装最新的 NVIDIA 驱动程序，且满足下表中的版本要求。
（最新的表格在这里）
如果要安装 GPU 驱动程序，可以安装更新的 CUDA 工具包，该工具包将捆绑更新的 GPU 驱动程序（安装程序）。或者您可以在此处检索驱动程序并安装它。我所说的较新的 CUDA 工具包是指 NVIDIA 提供的 CUDA 工具包安装程序，可以在此处获得，而不是通过 conda。您无法通过 conda 进行驱动程序更新。
我建议您学习CUDA linux 安装指南，因为用于安装上一个驱动程序（运行文件或包管理器）的方法可能就是您想要用于下一个驱动程序的方法。
作为替代方案（例如，如果您没有或无法获得系统的管理员访问权限），您可以研究 CUDA前向兼容性。（这也可能与兼容性有关。）
| 归档时间： |
|
| 查看次数： |
17225 次 |
| 最近记录： |