Separate trust_remote_code args

[kylesayrs]

Purpose

• Allow users to trust remote code of datasets
• Fix some of the finetune tests (all except tests/llmcompressor/transformers/finetune/test_oneshot_then_finetune)

Changes

• Renamed ModelArgs.trust_remote_code to ModelArgs.trust_remote_code_model to avoid name conflict during arg parsing
• Added DataTrainingArgs.trust_remote_code_data
• Added trust_remote_code_data arg to tests

Testing

• Removed dataset cache and ran python3 -m pytest tests/llmcompressor/transformers/finetune

[github-actions]

👋 Hi! Thank you for contributing to llm-compressor. Please add the ready label when the PR is ready for review.

[dsikka]

This LGTM and seems to only impact the finetune pathway. One suggestion @kylesayrs is to checkout a branch of https://github.com/neuralmagic/llm-compressor-testing/tree/main/.github/workflows and update such that instead of main, we checkout this branch. Can then kick off the weekly and nightly testing to make sure this doesn't cause any other impacts on args. Let me know if this is unclear. The tests should finish in about an hour each

[kylesayrs]

Initial tests were using a stale cache. In order to clear the HF dataset caches, you must remove two directories

rm -r ~/.cache/huggingface/datasets/*
rm -r ~/.cache/huggingface/modules/datasets_modules/datasets/

[Satrat]

LGTM, but agreed with Dipika lets run the nightly and weekly tests on this change before merging since we're changing argument names

[kylesayrs]

nightly ✅
weekly ✅
e2e ✅