ptxas version issue

[zohimchandani]

Getting the following warning message but not sure how to install ptxas? Are there instructions on how to do this somewhere?

2022-04-16 11:15:22.281636: W external/org_tensorflow/tensorflow/stream_executor/gpu/asm_compiler.cc:111] *** WARNING *** You are using ptxas 10.0.145, which is older than 11.1. ptxas before 11.1 is known to miscompile XLA code, leading to incorrect results or invalid-address errors.

You may not need to update to CUDA 11.1; cherry-picking the ptxas binary is often sufficient.

[YouJiacheng]

ptxas is a part of CUDA toolkit.
I suggest you to use conda to install CUDA.(Install CUDA by yourself can easily waste 1 day)

conda install cuda -c nvidia

[rtdew1]

I realize this is old, but thanks for this! Saved me a lot of time on a similar error...

[BwandoWando]

conda install cuda -c nvidia

amazing, thank you. It works.

[Yang-Xijie]

Yeah! You save my life! Thanks a lot!
Some supplements:

• If you dont want to install the latest version, you can check https://anaconda.org/nvidia/repo/installers and choose a lable with the desired version. And change the "package" to "cuda".
  • eg. I want to install cuda 11.8 so I use the command conda install --channel "nvidia/label/cuda-11.8.0" cuda

[binpan0423]

Getting the following warning message but not sure how to install ptxas? Are there instructions on how to do this somewhere?

2022-04-16 11:15:22.281636: W external/org_tensorflow/tensorflow/stream_executor/gpu/asm_compiler.cc:111] *** WARNING *** You are using ptxas 10.0.145, which is older than 11.1. ptxas before 11.1 is known to miscompile XLA code, leading to incorrect results or invalid-address errors.

You may not need to update to CUDA 11.1; cherry-picking the ptxas binary is often sufficient.

did you solve this problem？

[hawkinsp]

This means that the version of CUDA in your PATH is too old. Try running which ptxas to determine which copy of CUDA is being used.

JAX requires CUDA 11.1 or newer: https://github.com/google/jax#pip-installation-gpu-cuda

Hope that helps!

[1evergreen]

If there are many cuda toolkits of different version , manually set $CUDA in the .bashrc file and set $CUDA/bin to the $PATH enviroment. then use which ptxas to check if it is the right version.

If you are using conda env, set to the activate config file instead

[conorheins]

sorry for my naivete, but can you explain how to do the following:

manually set $CUDA in the .bashrc file and set $CUDA/bin to the $PATH enviroment.