ptxas executable missing

[hmaarrfk]

So I'm seeing the same issue as:

tensorflow/tensorflow#40036 (comment)

I'm wondering if we should depend on cudatoolkit-dev. This ensures that the executable ptxas is available to tensorflow.

[jakirkham]

Not that's deprecated.

If you want to look at a more complete set of tooling, would take a look at the packages in issue ( conda-forge/cudatoolkit-feedstock#62 ). Feedback there would also be welcome 🙂

[hmaarrfk]

can you elaborate if ptxas is deprecated or just cudatoolkit-dev

[h-vetinari]

I cannot elaborate on your question (though my belief is that just cudatoolkit-dev is deprecated, not ptxas), but just to note that ptxas also seems to be available in some form with the regular cudatoolkit. E.g. in the faiss feedstock, I constantly get log output of the form

ptxas /tmp/tmpxft_00000d54_00000000-5_PQScanMultiPassPrecomputed.compute_75.ptx, line 9649; warning : ld

but I'm not using anything unusual in terms of build setup.

[hmaarrfk]

cudatoolkit on its own doesn't seem to pull in ptxas for me:

Sample environment

$ mamba create --name cudatoolkit cudatoolkit=11.2 python=3.9 --yes

                  __    __    __    __
                 /  \  /  \  /  \  /  \
                /    \/    \/    \/    \
███████████████/  /██/  /██/  /██/  /████████████████████████
              /  / \   / \   / \   / \  \____
             /  /   \_/   \_/   \_/   \    o \__,
            / _/                       \_____/  `
            |/
        ███╗   ███╗ █████╗ ███╗   ███╗██████╗  █████╗
        ████╗ ████║██╔══██╗████╗ ████║██╔══██╗██╔══██╗
        ██╔████╔██║███████║██╔████╔██║██████╔╝███████║
        ██║╚██╔╝██║██╔══██║██║╚██╔╝██║██╔══██╗██╔══██║
        ██║ ╚═╝ ██║██║  ██║██║ ╚═╝ ██║██████╔╝██║  ██║
        ╚═╝     ╚═╝╚═╝  ╚═╝╚═╝     ╚═╝╚═════╝ ╚═╝  ╚═╝

        mamba (0.17.0) supported by @QuantStack

        GitHub:  https://github.com/mamba-org/mamba
        Twitter: https://twitter.com/QuantStack

█████████████████████████████████████████████████████████████


Looking for: ['cudatoolkit=11.2', 'python=3.9']

ramonaoptics/noarch      [====================] (00m:00s) Done
ramonaoptics/linux-64    [====================] (00m:00s) Done
conda-forge/noarch       [====================] (00m:01s) Done
conda-forge/linux-64     [====================] (00m:07s) Done
Transaction

  Prefix: /home/mark/miniforge3/envs/cudatoolkit

  Updating specs:

   - cudatoolkit=11.2
   - python=3.9


  Package               Version  Build           Channel                     Size
───────────────────────────────────────────────────────────────────────────────────
  Install:
───────────────────────────────────────────────────────────────────────────────────

  + _libgcc_mutex           0.1  conda_forge     conda-forge/linux-64      Cached
  + _openmp_mutex           4.5  1_gnu           conda-forge/linux-64      Cached
  + ca-certificates   2021.10.8  ha878542_0      conda-forge/linux-64      Cached
  + cudatoolkit          11.2.2  he111cf0_9      conda-forge/linux-64      Cached
  + ld_impl_linux-64     2.36.1  hea4e1c9_2      conda-forge/linux-64      Cached
  + libffi                  3.3  h58526e2_2      conda-forge/linux-64      Cached
  + libgcc-ng            11.2.0  h1d223b6_11     conda-forge/linux-64      Cached
  + libgomp              11.2.0  h1d223b6_11     conda-forge/linux-64      Cached
  + libstdcxx-ng         11.2.0  he4da1e4_11     conda-forge/linux-64      Cached
  + libzlib              1.2.11  h36c2ea0_1013   conda-forge/linux-64      Cached
  + ncurses                 6.2  h58526e2_4      conda-forge/linux-64      Cached
  + openssl              1.1.1l  h7f98852_0      conda-forge/linux-64      Cached
  + pip                  21.3.1  pyhd8ed1ab_0    conda-forge/noarch        Cached
  + python                3.9.6  h7c244ba_103    ramonaoptics/linux-64     Cached
  + python_abi              3.9  2_cp39          conda-forge/linux-64      Cached
  + setuptools           59.2.0  py39hf3d152e_0  conda-forge/linux-64        1 MB
  + sqlite               3.36.0  h495f6a3_100    ramonaoptics/linux-64     Cached
  + tk                   8.6.11  h27826a3_1      conda-forge/linux-64      Cached
  + tzdata                2021e  he74cb21_0      conda-forge/noarch        Cached
  + wheel                0.37.0  pyhd8ed1ab_1    conda-forge/noarch        Cached
  + xz                    5.2.5  h516909a_1      conda-forge/linux-64      Cached
  + zlib                 1.2.11  h36c2ea0_1013   conda-forge/linux-64      Cached

  Summary:

  Install: 22 packages

  Total download: 1 MB

───────────────────────────────────────────────────────────────────────────────────

Finished setuptools                           (00m:00s)               1 MB      2 MB/s
Downloading  [=============================================================================] (00m:00s)    2.32 MB/s
Extracting   [=============================================================================] (00m:00s)        1 / 1
Preparing transaction: done
Verifying transaction: done
Executing transaction: \ By downloading and using the CUDA Toolkit conda packages, you accept the terms and conditions of the CUDA End User License Agreement (EULA): https://docs.nvidia.com/cuda/eula/index.html

done
#
# To activate this environment, use
#
#     $ conda activate cudatoolkit
#
# To deactivate an active environment, use
#
#     $ conda deactivate

$ conda activate cudatoolkit
(cudatoolkit) $ ptxas

Command 'ptxas' not found, but can be installed with:

sudo apt install nvidia-cuda-toolkit

[h-vetinari]

@hmaarrfk
Could you raise an issue about this on the cudatoolkit feedstock? As John said, cudatoolkit-dev as a package is not gonna be acceptable anymore, I think.

[hmaarrfk]

will do. thanks