Support for quantized models

[Omnyyah]

🚀 Feature

Please extend Opacus to support differentially private training of quantized models.

Motivation

Enabling efficient and private machine learning by combining differential privacy with model quantization.

Additional context

This would facilitate deploying private AI models on resource-constrained devices.

[EnayatUllah]

Thanks for the feature request! Can you add more details on what kind of quantization methods do you want Opacus to support? It may be possible that we can still use Opacus, out of the box, for these? If not, we can distill on what additional properties we need to enable it.

[Omnyyah]

Thank you for your repsonse.
so the thing is that, I am using unsloth to train LLMs faster, and they use "Dynamic 4-bit Quantization, which involves dynamically opting not to quantize certain parameters and this builds on top of BitsandBytes 4-bit. This approach delivers significant accuracy gains while only using <10% more VRAM than BnB 4-bit."

but, unfortunately, applying DP using Opacus is not supported in this case, I must use the standard PyTorch parameters to compute per-sample gradients.
Bitsandbytes 4-bit parameters store weights in custom quantization blocks that break the usual PyTorch grad flow, so Opacus can’t attach its gradient hooks.

[HuanyuZhang]

Thanks to @Omnyyah for the idea. Yeah, quantization is an important feature which is on our radar. However, we do not have a clear timeline to support this function right now. If you could have a proposal/prototyping, we are happy with the consulting and discussion.

[Omnyyah]

Thank you for your reply!
would you please clarify what the proposal/prototyping should contain ??