[Feature]: BitsandBytes quantization with TP>1

[jvlinsta]

🚀 The feature, motivation and pitch

Any QLoRA adapters trained on large checkpoints (e.g., 70B) are unusable as we cannot use TP>1 to shard the model over multiple GPUs. Therefore, resolving this would enable models that were trained with quantization, rather than having to rely on GPTQ and AWQ, which are applied post-hoc after training.

Alternatives

No response

Additional context

No response

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.

[copasseron]

feature added in v0.6.2 #8434.