Skip to content

Checkpoint Conversion to HuggingFace (GPT2) #305

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.

Sign up for GitHub

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

Jump to bottom
Merged
merged 31 commits into from
Mar 19, 2025
Merged

Checkpoint Conversion to HuggingFace (GPT2) #305

merged 31 commits into from
Mar 19, 2025

Conversation

flxst
Copy link
Member

@flxst flxst commented Feb 21, 2025 •
edited by BlueCrescent
Loading

What does this PR do?

This PR implements the true conversion of the modalities gpt2 model into a llama style huggingface model. Note that this does not include the tokenizer.

In addition, the getting started tutorial got refactored and updated. Checkpoint conversion is end-to-end tested based on the checkpoint from the getting started tutorial.

Follow-up issues were created: #308 and #309.

General Changes

  • The configs for the getting started tutorial got updated (no bias, layer norm, swiglu, pytorch_flash attn)
  • Under modalities/conversion/gpt2 code was added for turning a modalities gpt2 checkpoint into something that can be loaded via AutoModelForCausalLM.from_pretrained().

Breaking Changes

  • None

Checklist before submitting final PR

  • My PR is minimal and addresses one issue in isolation
  • I have merged the latest version of the target branch into this feature branch
  • I have reviewed my own code w.r.t. correct implementation, missing type hints, proper documentation, etc.
  • I have run a sample config for model training
  • I have checked that all tests run through (python tests/tests.py)
  • I have updated the internal changelog (CHANGELOG_DEV.md)

@BlueCrescent
feat(huggingface): Added first version of a true conversion of the mo…
492a178 
…dalities gpt2 model into a llama style huggingface model.
@flxst flxst marked this pull request as draft February 21, 2025 07:38
flxst and others added 22 commits February 21, 2025 10:47
@flxst
refactor(conversion): convert_gpt2
84ac2b7
@BlueCrescent
fix(huggingface): Set correct model_type in GPT2Config.
d516469
@BlueCrescent
refactor(huggingface): Shortened some long lines.
49191d1
@BlueCrescent
feat(huggingface): When converting gpt2 now all the necessary model c…
7332916 
…ode gets included in the hf checkpoint.
@BlueCrescent
docs(huggingface): Added docstrings for conversion script.
a390289
@flxst
chore(getting_started): update config (no bias, layer norm, swiglu, p…
0e4310d 
…ytorch_flash attn)
@flxst
test(conversion): apply checkpoint conversion based on getting starte…
d3ab1b5 
…d example
@flxst
test(getting_started): move wandb directory to data (so that it gets …
c01a1fe 
…removed)
@flxst
refactor(getting_started): refactoring (folder structure) and general…
85411be 
… update (README and configs)
@BlueCrescent
fix(huggingface): Added missing import.
262f06c
@BlueCrescent
refactor(huggingface): Minor code improvement.
6564452
@flxst
fix(getting_started): capture subprocess errors properly + bug fix
f7e558c
@flxst
chore: Merge branch 'conversion_modalities_to_huggingface' of https:/…
cb6259e 
…/github.com/Modalities/modalities into conversion_modalities_to_huggingface
@flxst
test(conversion): check checkpoint conversion based on getting starte…
01c75ec 
…d example
@BlueCrescent
feat(huggingface): Added handling of additional config settings to co…
b119726 
…nversion script.
@BlueCrescent
test(huggingface): Added additional tests for conversion functions.
286b37f 
TODO: test_converting_gpt2_does_not_change_outputs currently fails.
@BlueCrescent
fix(checkpointing): Fixed warning.
3948575
@BlueCrescent
test(huggingface): Fixed model output comparison test.
8991797
@BlueCrescent
refactor(huggingface): Split conversion logic into multiple files.
e91633d
@BlueCrescent
test(huggingface): Additional tests and some refactoring.
d43b8d2
@BlueCrescent
revert(checkpointing): For now, forcing weights_only is not supposed …
836e934 
…to be added.
@flxst
fix(conversion): check that criteria for conversion are fulfilled
f4cc962
@flxst flxst mentioned this pull request Feb 25, 2025
Open
@flxst flxst self-assigned this Feb 25, 2025
flxst
flxst commented Feb 25, 2025
View reviewed changes
@BlueCrescent BlueCrescent marked this pull request as ready for review February 25, 2025 15:05
ajude2s
ajude2s reviewed Mar 3, 2025
View reviewed changes
@fromm-m fromm-m self-requested a review March 3, 2025 11:07
@BlueCrescent
fix(huggingface): The config conversion now only allows all norms to …
cf3cba6 
…have the same setting.

Having different settings is currently not supported by our hf model.
Also, some additionally minor fixes and refactorings.
@BlueCrescent
docs(huggingface): small fix
9707027
@ajude2s
Copy link
Collaborator

ajude2s commented Mar 5, 2025

LGTM

fromm-m
fromm-m approved these changes Mar 19, 2025
View reviewed changes
Copy link
Member

@fromm-m fromm-m left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LGTM

@fromm-m fromm-m merged commit adb23a3 into main Mar 19, 2025
3 checks passed
@BlueCrescent BlueCrescent deleted the conversion_modalities_to_huggingface branch June 12, 2025 08:38
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Reviewers

@BlueCrescent BlueCrescent BlueCrescent left review comments

@fromm-m fromm-m fromm-m approved these changes

@ajude2s ajude2s ajude2s approved these changes

Assignees

@BlueCrescent BlueCrescent

@flxst flxst

Labels
None yet
Projects
None yet
Milestone
No milestone
Development

Successfully merging this pull request may close these issues.
4 participants
@flxst @ajude2s @BlueCrescent @fromm-m