[QEff Finetune]: Made a separate list of dependencies for Finetuning. #636
Conversation
quic-meetkuma
commented
Nov 25, 2025
quic-meetkuma
commented
- With this change the FT dependencies should be explicitly installed via "pip install -e .[ft]".
- Added this so that eager stack can have different pytorch and transformers dependencies than the AOT stack.
Signed-off-by: meetkuma <[email protected]>
quic-meetkuma
requested review from
ochougul,
quic-akuruvil,
quic-hemagnih,
quic-rishinr,
quic-swatia and
tchawada
| quality = ["black", "ruff", "hf_doc_builder@git+https://github.com/huggingface/doc-builder.git"] | ||
|
|
||
| # New FT dependencies | ||
| ft = [ |
There was a problem hiding this comment.
whl file of that would come from Eager team itself just like torch_qaic package.
There was a problem hiding this comment.
That would require explicit installation again by the user right. If it can be included here it would be good. When we do pip install .[ft] all the dependencies required for FT are installed in one go.
There was a problem hiding this comment.
Yes, plan is to include files here and make it install from /opt/ location.
…k to resolve the conflicting deps. Signed-off-by: meetkuma <[email protected]>
| test = ["pytest","pytest-mock"] | ||
| docs = ["Sphinx==7.1.2","sphinx-rtd-theme==2.0.0","myst-parser==3.0.1","sphinx-multiversion"] | ||
| quality = ["black", "ruff", "hf_doc_builder@git+https://github.com/huggingface/doc-builder.git"] | ||
| infer = [ |
There was a problem hiding this comment.
This will be default; we just need to over right when we do pip install -e .[ft]
There was a problem hiding this comment.
No, having torch under dependencies section and different torch version under "ft" optional dependencies -> creates conflicts. Hence, proposing to move different torch versions under different sections. Same will be true for transformers. FT will soon have its own transformer version (either open-source or based on whl file which contains qaic backend changes).
|
In this PR, we should use pip install .[ft] in the Jenkinsfile once #629 is merged. This will test the correct intended environment in the CI. |
| ] | ||
| ft = [ | ||
| "tensorboard ; python_version>='3.10' and python_version<'3.12' and platform_machine=='x86_64'", | ||
| "transformers==4.55.0 ; python_version>='3.10' and python_version<'3.12' and platform_machine=='x86_64'", |
There was a problem hiding this comment.
Dependencies already has the same version of transformers at line number 22 of this file. Why is this needed here?
There was a problem hiding this comment.
Please check the python version comment for this line as well.
There was a problem hiding this comment.
Eager support python 3.10 and 3.12.
Explicitly adding transformers here to simplify future modifications to the FT’s transformer.
There was a problem hiding this comment.
Transformers and torch should be handled uniformly. While both of them are added in the FT section, for inference one of them is kept in dependencies and other in infer section.
There was a problem hiding this comment.
Updated in latest.
| "tensorboard ; python_version>='3.10' and python_version<'3.12' and platform_machine=='x86_64'", | ||
| "transformers==4.55.0 ; python_version>='3.10' and python_version<'3.12' and platform_machine=='x86_64'", | ||
| "torch@https://download.pytorch.org/whl/cpu/torch-2.9.0%2Bcpu-cp310-cp310-manylinux_2_28_x86_64.whl ; python_version=='3.10' and platform_machine=='x86_64'", | ||
| "torch@https://download.pytorch.org/whl/cpu/torch-2.9.0%2Bcpu-cp311-cp311-manylinux_2_28_x86_64.whl ; python_version=='3.11' and platform_machine=='x86_64'", |
There was a problem hiding this comment.
Are we supporting python 3.11? At line number 20, "requires-python = ">=3.8,<3.11" is mentioned. It will not let QEff install for python 3.11.
There was a problem hiding this comment.
Yes, that is global level python. It will soon start to support python=3.12 as well. In that case this line still holds true. No harm in adding this here.
Signed-off-by: meetkuma <[email protected]>
Signed-off-by: meetkuma <[email protected]>
Signed-off-by: meetkuma <[email protected]>
4 participants
