Closed hypnopump closed 12 months ago
@hypnopump Could install Trident again? There is no issue on my machine.
bc-user@instance-6285:~/projects/trident$ ipython
Python 3.10.11 (main, Apr 20 2023, 19:02:41) [GCC 11.2.0]
Type 'copyright', 'credits' or 'license' for more information
IPython 8.12.0 -- An enhanced Interactive Python. Type '?' for help.
In [1]: import trident as tr
You can add Trident in requirements.txt
.
trident@git+https://github.com/kakaobrain/trident.git@main
Fails with the same error. Here's my conda env YAML file in case it's useful for reproduction. my_conda_env.yaml.txt
I have also tried a machine with a V100 GPU in case it was the compute architecture or so and it still fails
@ansteve @mejai1206 Could you try reproduce this issue?
Found a solution:
# follow same steps at the start
conda create -n python310 python=3.10 --yes
conda activate python310
pip3 install ipython torch torchvision torchaudio
pip install git+https://github.com/kakaobrain/trident.git
ipython
>>> import trident as td
# this will raise an error
And now, update triton
:
pip install -U --index-url https://aiinfra.pkgs.visualstudio.com/PublicPackages/_packaging/Triton-Nightly/pypi/simple/ triton-nightly
ipython
>>> import trident as td
# this works
Probably it's just a bad commit of triton being installed by default. Could we try to fix the setup.py
and install_package.sh
to point to working versions?
@hypnopump That's very strange. Because the latest of Triton is installed while installing Trident. It means that the installed Triton is more latest version than daily build. I guess somehow the old version of Triton is installed in your conda enviornment. Look at Trident's Actions. Trident is tested with the latest Triton.
đ Describe the bug
Can't import the latest version of triton after following the installation steps, on a linux machine with an NVidia T4 GPU and CUDA version 12.0.
đ§âđŤ Reproduction
To reproduce:
đŻ Expected behavior
Importing it should just work
đźď¸ Screenshots
Here's my stacktrace: And the error i'm getting is:
đť Requirements
đŹ Additional context
If you could share either a conda yaml or a requirements.txt that would make it just work, that would be awesome.