[WIP] Proper tool calling support in the torchtune

[krammnic]

Context

What is the purpose of this PR? Is it to

• add a new feature
• fix a bug
• update tests and/or documentation
• other (please add here)

Please link to any issues this PR addresses.

Changelog

What are the changes made in this PR?

• tool calling support + modifications for the HFBase/HFModel tokenizers

Test plan

Please make sure to do each of the following if applicable to your PR. If you're unsure about any one of these just ask and we will happily help. We also have a contributing page for some guidance on contributing.

• run pre-commit hooks and linters (make sure you've first installed via pre-commit install)
• add unit tests for any new functionality
• update docstrings for any new or updated methods or classes
• run unit tests via pytest tests
• run recipe tests via pytest tests -m integration_test
• manually run any new or modified recipes with sufficient proof of correctness
• include relevant commands and any other artifacts in this summary (pastes of loss curves, eval results, etc.)

UX

If your function changed a public API, please add a dummy example of what the user experience will look like when calling it.
Here is a docstring example
and a tutorial example

• I did not change any public API
• I have added an example to docs or docstrings

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/torchtune/2794

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[kamalojasv181]

Hey @krammnic, does this support tool calls for all formats (like openai, sharegpt etc)?

[krammnic]

It still WIP, but yes, it will

[ebsmothers]

Thanks for working on this! I left a few small comments. We should also add a test to ensure that this actually works and generates the expected outputs on a tool-calling dataset

[ebsmothers]

Is this related to tool-calling? Or a separate issue?

[krammnic]

It is caused by separate issue in HuggingfaceBaseTokenizer.

[ebsmothers]

In this case I wonder whether we should just modify _get_token_from_config to directly return None (possibly logging a warning) rather than use this try/except

[ebsmothers]

Should we also update the role "ipython" to "tool" to match what's done by Hugging Face?

[krammnic]

Yes, good catch, that argument seemed to me weird.

[nathan-az]

Was this an intentional removal?

[krammnic]

Nope :/ I don't like to do it on Mac (in another case it will not install) and then remove the comment. Will open separate PR to address this

[nathan-az]

While I agree with this change, it currently breaks existing tokenizers. Repro:

from torchtune.datasets import alpaca_cleaned_dataset
from torchtune.models.qwen2_5 import qwen2_5_tokenizer

vocab_path = "/tmp/Qwen2.5-14B-Instruct/vocab.json"
merges_path = "/tmp/Qwen2.5-14B-Instruct/merges.txt"
tokenizer_json_path = "/tmp/Qwen2.5-14B-Instruct/tokenizer.json"
tokenizer_config_path = "/tmp/Qwen2.5-14B-Instruct/tokenizer_config.json"

tokenizer_qwen = qwen2_5_tokenizer(
    path=vocab_path,
    merges_file=merges_path,
    max_seq_len=512
)
dataset_qwen = alpaca_cleaned_dataset(tokenizer=tokenizer_qwen, packed=False)

[krammnic]

Hmm, yep. We need to remove ipython everywhere.

[nathan-az]

@krammnic This is great progress. It looks like a bit of work needs to be done on BC with existing tokenizers (unless the plan is to fully deprecate them). In addition I'm seeing some issues with the jinja rendering - I think it may require explicitly passing tools to the renderer (hf ref).

Repro:

from torchtune.datasets import alpaca_cleaned_dataset
from torchtune.modules.transforms.tokenizers import HuggingFaceModelTokenizer


tokenizer_json_path = "/tmp/Qwen2.5-14B-Instruct/tokenizer.json"
tokenizer_config_path = "/tmp/Qwen2.5-14B-Instruct/tokenizer_config.json"

tokenizer_hf = HuggingFaceModelTokenizer(
    tokenizer_json_path=tokenizer_json_path,
    tokenizer_config_json_path=tokenizer_config_path,
    max_seq_len=512,
)
dataset_hf = alpaca_cleaned_dataset(tokenizer=tokenizer_hf, packed=True)

Basically optionally propagating tools might solve this (basic example). I haven't tested beyond seeing that the above repro runs but haven't checked correctness.