Feature/peft compatible models

[danbider]

Edits needed to support a combo of composer with hf/peft.

Pipeline is:

1. load a hf model e.g., mpt-7b
2. use hf/peft to add lora modules or adapter modules.
3. convert that peft model (that is loaded into python) into a composer model (use my new function for this)
4. train in composer (required adding the inputs_embeds args to model.forward().

[danbider]

refactored the hf convertor to a single function as suggested by @dakinggg. tested it on my end and ran pre-commit successfully. I want to move forward and push the code updates to the hub.

[samhavens]

Tests are failing with

___________________ ERROR collecting tests/test_training.py ____________________
ImportError while importing test module '/llm-foundry/tests/test_training.py'.
Hint: make sure your test modules/packages have valid Python names.
Traceback:
llmfoundry/__init__.py:8: in <module>
    from llmfoundry.data import (ConcatTokensDataset,
llmfoundry/data/__init__.py:5: in <module>
    from llmfoundry.data.denoising import (MixtureOfDenoisersCollator,
llmfoundry/data/denoising.py:20: in <module>
    from llmfoundry.models import utils
llmfoundry/models/__init__.py:4: in <module>
    from llmfoundry.models.hf import (ComposerHFCausalLM, ComposerHFPrefixLM,
llmfoundry/models/hf/__init__.py:4: in <module>
    from llmfoundry.models.hf.hf_causal_lm import (ComposerHFCausalLM,
llmfoundry/models/hf/hf_causal_lm.py:10: in <module>
    import peft
E   ModuleNotFoundError: No module named 'peft'

[c9o]

Hello @danbider , could you share your yamls for MPT peft/lora training? Thanks.

[mvpatel2000]

Mostly LGTM! Left a few comments. Will approve on next pass / w review from NLP person

[codestar12]

as you have resolved everyone else's concerns I will approve.

[stoperro]

Is there an example on how to fine-tune with this?

[chris-aeviator]

@stoperro according to #416 just use the ordinary peft code (huggingface has ready to go PEFT notebooks) or with llm-foundry add

[palash04]

Hey @chris-aeviator, I noticed that in the repository, LoRA currently only supports MPT models. Can we perform LoRA fine-tuning on other models such as LLAMA?

[dakinggg]

@palash04 this is getting fixed in #435