Pad the examples for QLoRa finetuning test

[ckvermaAI]

1. Pad the examples up to max_seq_len of 1024
2. Increase the max_steps (from 5 to 50) and eval_steps (from 3 to 10). Set throughput-related arguments (adjust_throughput, throughput_warmup_steps)
3. Update the fine-tuning test name
4. Also, adjust the reference "eval loss" value for fine-tuning test and "output" for inference test.

Additional updates
6. Enable the eager mode for the test (disable the torch.compile mode for now).
7. Add new requirement for installing the bitsandbytes (from https://github.com/bitsandbytes-foundation/bitsandbytes/tree/multi-backend-refactor)

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[vivekgoe]

LGTM. @regisss can we add these tests to slow_tests you run for G2/G3? This will help us avoid regressions. We can publish QLoRA support as experimental/beta future on G2/G3 (works for limited configurations and performance is not great). Note this PR is dependent on Synapse 1.21.0 release (not backward compatible).

[ckvermaAI]

Support for NF4 quantization/dequantization using Intel Gaudi hardware: bitsandbytes-foundation/bitsandbytes#1592

[regisss]

LGTM

[uartie]

We're starting to see OSError: /usr/local/lib/python3.10/dist-packages/bitsandbytes/libbitsandbytes_cpu.so: cannot open shared object file: No such file or directory when running some of the tests (e.g. test_diffusers.py)

We've run python -m pip install .[tests] and it installed bitsandbytes 1.0.0... so not sure why we're seeing this error.

[uartie]

...

pip install -r examples/stable-diffusion/requirements.txt
pip install -r examples/stable-diffusion/training/requirements.txt

... appears to cause the error to show up during test run.

[uartie]

Possibly related to old peft version in examples/stable-diffusion/training/requirements.txt... please fix the requirements file.

[ckvermaAI]

This error is coming from bitsandbytes, it should not be related to peft. I've run other tests locally (test_bnb_qlora.py, test_bnb_inference.py), and I didn't face this issue.

Let me check it again.

[ckvermaAI]

In case of HPU, bitsandbytes loads the CPU binaries (https://github.com/bitsandbytes-foundation/bitsandbytes/blob/multi-backend-refactor/bitsandbytes/cextension.py#L73), which are not required.
And if there is no issue in test/topology, you'll not see OSError due to missing *.so file, but if failure occurs due to any other reason, you'll also see the OSError due to missing *.so file.

In short,

1. I'll move the bitsandbytes installation from setup.py to bitsandbytes tests (test_bnb_qlora.py, test_bnb_inference.py). This should resolve any issue you're seeing.
2. Later, we'll try to upstream the fix to the bitsandbytes repo.

[ckvermaAI]

Fix for the above issue:
Move bitsandbytes requirements from setup.py to bnb tests
#1946