Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

don't do max versions as much #1826

Open
wants to merge 1 commit into
base: main
Choose a base branch
from
Open

don't do max versions as much #1826

wants to merge 1 commit into from

Conversation

t-vi
Copy link
Contributor

@t-vi t-vi commented Nov 6, 2024

See #1825

This also blocks bumping the litgpt version in Thunder.

@t-vi
Copy link
Contributor Author

t-vi commented Nov 6, 2024

As expected, this is tricky in the bnb tests. I'll look at making them conditional or some other fix.

@rasbt
Copy link
Collaborator

rasbt commented Nov 6, 2024

Thanks for the PR! My concern here is that if we remove the max pin for PyTorch, then quantization will be broken out of the box for regular users who use the pip package.

Question: If you install LitGPT first and then upgrade PyTorch manually to the latest version, wouldn't this help with supporting LitGPT in Thunder?

In any case, what we really need is a new Lightning release. That would fix everything.

@t-vi
Copy link
Contributor Author

t-vi commented Nov 6, 2024

Yeah, with Lightning-AI/pytorch-lightning#20351 it looks like we get a new lightning soon, too.
The trouble in the thunder CI is that the PyTorch comes from docker, so we would need to special-case LitGPT to be installed with no-deps.
But also when I consider my own computing needs, I am quite annoyed when packages mess with my PyTorch installation, and downgrading seems particularly bad.

@rasbt
Copy link
Collaborator

rasbt commented Nov 6, 2024

Just had a brief chat with Luca and I think the new Lightning version will be released very soon. (He said he's quite close).

The trouble in the thunder CI is that the PyTorch comes from docker, so we would need to special-case LitGPT to be installed with no-deps.

Hm yeah, so is it possible in the meantime to install LitGPT without dependencies, i.e.,

pip install litgpt --no-deps

and then following up with installing the other ones manually:

pip install \
    "lightning==2.4.0" \
    "jsonargparse[signatures]>=4.30.1,<=4.32.1" \
    "huggingface_hub>=0.23.5" \
    "safetensors>=0.4.3" \
    "tokenizers>=0.15.2" \
    "tqdm>=4.66.0" \
    "bitsandbytes==0.42.0" \
    "sentencepiece>=0.2.0" \
    "requests>=2.31.0" \
    "litdata==0.2.17" \
    "litserve>=0.1.5" \
    "zstandard>=0.22.0" \
    "pandas>=1.9.0" \
    "pyarrow>=15.0.2" \
    "tensorboard>=2.14.0" \
    "torchmetrics>=1.3.1" \
    "datasets>=2.18.0" \
    "transformers>=4.38.0" \
    "lm-eval>=0.4.2" \
    "huggingface_hub[hf_transfer]>=0.21.0" \
    "uvloop>=0.2.0; sys_platform != 'win32'"

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

2 participants