torchtune as an optional dependency: Lazy Import

[Jack-Khuu]

🚀 The feature, motivation and pitch

We would like to make torchtune an optional dependency. The first step towards that is to avoid importing torchtune unless it is actively used.

To make this migration easier, let's move the top level imports into the functions/classes that require them.

We explicitly acknowledge that this initial step isn't a best practice, but will make the following work simpler.

Here's an example where we delay imports here:

torchchat/torchchat/usages/eval.py

Lines 216 to 225 in 1384f7d

	# Having the imports here allow running other evals without installing torchtune
	from torchtune.utils import batch_to_device
	from torchtune.data import (
	format_content_with_images,
	left_pad_sequence,
	Message,
	padded_collate_tiled_images_and_mask,
	)
	from torchtune.generation import generate, sample
	from torchtune.modules.common_utils import local_kv_cache

Task: Update all imports of torchtune in the repo, such that imports are only done when necessary

• Suggestion: Feel free to send out PR's that only update a subset of all use cases

To test your changes, run:

• With torchtune installed: https://github.com/pytorch/torchchat/blob/main/docs/multimodal.md#generation
• With torchtune uninstalled: python torchchat.py generate llama3.2-1B

Alternatives

No response

Additional context

No response

RFC (Optional)

No response

[krammnic]

Hey, I'm from torchtune. Would you mind me to take this issue?

[Jack-Khuu]

Totally, give it a shot