🔬 SFT simplification #2405
Open
🔬 SFT simplification #2405
+377
−412
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
qgallouedec
commented
Dec 1, 2024
Comment on lines
-410
to
-449
| # This should not work because not enough data for one sample | ||
| training_args = SFTConfig( | ||
| output_dir=tmp_dir, | ||
| dataloader_drop_last=True, | ||
| max_steps=2, | ||
| eval_steps=1, | ||
| save_steps=1, | ||
| per_device_train_batch_size=2, | ||
| max_seq_length=1024, # make sure there is NOT at least 1 packed sequence | ||
| packing=True, | ||
| report_to="none", | ||
| ) | ||
| with self.assertRaises(ValueError): | ||
| _ = SFTTrainer( | ||
| model=self.model, | ||
| args=training_args, | ||
| train_dataset=self.dummy_dataset, | ||
| formatting_func=formatting_prompts_func, | ||
| ) | ||
|
|
||
| # This should not work as well | ||
| with self.assertRaises(ValueError): | ||
| training_args = SFTConfig( | ||
| output_dir=tmp_dir, | ||
| dataloader_drop_last=True, | ||
| max_steps=2, | ||
| eval_steps=1, | ||
| save_steps=1, | ||
| per_device_train_batch_size=2, | ||
| packing=False, | ||
| report_to="none", | ||
| ) | ||
| _ = SFTTrainer( | ||
| model=self.model, | ||
| args=training_args, | ||
| train_dataset=self.dummy_dataset, | ||
| formatting_func=formatting_prompts_func, | ||
| ) | ||
|
|
There was a problem hiding this comment.
the new implementation extend the support for these two cases
qgallouedec
commented
Dec 1, 2024
| @@ -807,8 +768,6 @@ def test_sft_trainer_infinite_with_model(self): | |||
| eval_dataset=self.eval_dataset, | |||
| ) | |||
|
|
|||
| self.assertTrue(trainer.train_dataset.infinite) | |||
There was a problem hiding this comment.
the trainer is not "infinite" (subclass of ConstantLenghtDataset) anymore: it's better since we can use num_epochs now
29 tasks
|
FSDP + QLora from https://github.com/huggingface/peft/blob/main/examples/sft/run_peft_qlora_fsdp.sh green: main 
|
* First changes * Other files * Finally * rm comment * fix nashmd * Fix example * Fix example [ci skip]
* fix chat for windows * add some tests back * Revert "add some tests back" This reverts commit 350aef5.
* datast_config_name * Update trl/utils.py [ci skip] * sort import * typo [ci skip] * Trigger CI * Rename `dataset_config_name` to `dataset_config`
* Remove unused deepspeed code * add model prep back * add deepspeed even if it doesn't work * rm old code
* `model_config` -> `model_args` * sort
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
What does this PR do?
Context
This PR is the first iteration of SFT refactoring: tests have deliberately been left unmodified, to ensure continuity of support for the user and, if possible, no breaking changes.
Other PRs will follow to further simplify the SFT (to be discussed on a case-by-case basis). These modifications would require test modifications and breaking changes. For example:
ConstantLengthDatasetin favor of thepackingargument alone.TrainingArgumentsand imposeSFTConfig.What's in this PR
This PR refactor and simplifies SFT.
Add full conversational/standard dataset support
Part of #2071
Refactor packing
Drop
ConstantLengthDatasetin favour of a simpler approach (seepack_examples)Current status
Testing. The goal is to gather code example from anywhere and ensure there're still working.
Before submitting
Pull Request section?
to it if that's the case.
documentation guidelines.
Who can review?
Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.