Tensorflow 2.4.1 - Couldn't invoke ptxas.exe

Question

I try to run Tensorflow with GPU support (GTX 1660 SUPER).

I created an enviroment using anaconda, than installed cudatoolkit (version 11.0.221) and tensorflow-gpu (version 2.4.1). Afterwards, I downloaded cuDNN (version 8.0.4), and copied all files from cuDNN's bin folder to my environment's bin folder at anaconda3\envs\<env name>\Library\bin. In my script, I've set the memory limit to my GPU's memory using tf.config.experimental.set_memory_growth.

When I run the script (which uses convolutional algorithms), I get a warning that says Couldn't invoke ptxas.exe --version which comes after an Call to CreateProcess failed. Error code: 2 error. After the launch failure, I get: Relying on driver to perform ptx compilation. Modify $PATH to customize ptxas location.

I've already tried switching to cuDNN version 8.1.1.

How I fix this?

Answer 1

I got a new fix for this.

First I tried using tensorflow=2.3, cudnn=7.6.5 and cudatoolkit=10.1 as mentioned in previous answers. However, every time I put a model to train, the process was going stale and the training seemed to be stuck in epoch 1.

I then managed to include ptxas in my conda environment by running conda install -c nvidia cuda-nvcc The packages I am using are:

tensorflow=2.9, cudnn=8.1.0, cudatoolkit=11.2.2, cuda-nvcc=11.7.99 and python=3.9

I am running everything on windows 10 flawlessly now.

Answer 2

This always fixes it for me on Windows x64:

conda install -c nvidia cuda-nvcc

To check, run this:

(python39) C:\>ptxas.exe --version

And expect something like this:

ptxas: NVIDIA (R) Ptx optimizing assembler
Copyright (c) 2005-2023 NVIDIA Corporation
Built on Fri_Jan__6_19:00:08_Pacific_Standard_Time_2023
Cuda compilation tools, release 12.0, V12.0.140
Build cuda_12.0.r12.0/compiler.32267302_0