[Bugfix] Fix saving offloaded state dict

[kylesayrs]

Purpose

• Fix bug with saving the entire state dict of an offloaded model, even if it does not have a compressor
• Properly infer sparsity and quantization of offloaded models

Changes

• Load the entire offloaded state dict before checking for a compressor and before inferring global sparsity
• Remove explicit override of save_safetensors kwarg, since the original save_pretrained function already defaults its value to True

Testing

• Added test which throws error when trying to save an offloaded model without compression. This test fails on main but passes on this branch

Main

FAILED tests/llmcompressor/transformers/sparsification/test_compress_tensor_utils.py::test_model_reload[True-torch_dtype3-False-cpu] - NotImplementedError: Cannot copy out of meta tensor; no data!
FAILED tests/llmcompressor/transformers/sparsification/test_compress_tensor_utils.py::test_model_reload[True-torch_dtype5-False-cuda:0] - NotImplementedError: Cannot copy out of meta tensor; no data!
FAILED tests/llmcompressor/transformers/sparsification/test_compress_tensor_utils.py::test_model_reload[True-torch_dtype6-True-cuda:0] - NotImplementedError: Cannot copy out of meta tensor; no data!
FAILED tests/llmcompressor/transformers/sparsification/test_compress_tensor_utils.py::test_model_reload[True-torch_dtype7-True-cuda:0] - NotImplementedError: Cannot copy out of meta tensor; no data!
======================================= 4 failed, 4 passed, 12 deselected in 9.03s =======================================

This branch

============================================ 8 passed, 12 deselected in 8.89s ============================================

[github-actions]

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

[kylesayrs]

I'm pretty sure this should be merged, but I want to fully test checkpoint loading before this is merged in

[dsikka]

LGTM. Just one confirmation question

[dsikka]

@kylesayrs can you resolve the conflict

[dsikka]

@kylesayrs Seems like some of the errors for the sparsification tests are new?

[kylesayrs]

I addressed the cuda errors by separating out GPU-dependent tests
There are other, new failures which are likely caused by upstream CT changes, but they are unrelated and outside of the scope of this PR, which only aims to address the newly added tests test_model_reload and to precursor other test fixes

[rahul-tuli]