Hacker News Reader

Comment by chaos_emergent | Hacker News Reader

The findings are that the best fine-tune performance comes from fine-tuning all weights, followed my MLPs, followed by attention heads, using LoRA. Authors assert that the performance difference is based on the target module of the NN.

Isn’t an equally valid argument that MLPs tend to constitute a greater number of weights in transformer networks than attention heads, and the performance difference can be traced to a greater number of weights having freedom to change? I’d be curious to know if randomly choosing a subset of matrices to train, regardless of where they are in the network, would provide analogous performance to LoRA on a specific module with comparable learnable weights.

View on HN

I think the QLoRA paper https://arxiv.org/pdf/2305.14314 paper also showed LoRA on all MLP + Attention layers > all MLP layers > just Attention layers.

Other papers show finetuning a select few layers can also work well.

Any real world performance comparison between QLoRa and LoRa?

The QLoRA paper itself provided some cool benchmarks across many many experiments - QLoRA is near equivalent to LoRA, with it sometimes exceeding or losing 1-2% accuracy (it depends on the use case)

as a follow up curiosity, has anyone tried using LoRA on the entire model for pretraining to compare regular training model performance to LoRA?

This paper [1] does atempt that and reports similar performance compared to conventional pre-training. However, they do start off by doing a normal full-rank training and claim that it is needed to 'warm start' the training process.

[1] https://arxiv.org/abs/2307.05695

Oh yes this paper! The main issue is the scaling of the A and B LoRA matrices. Some papers show scaling the B matrix with larger learning rates (LoRA+) could be beneficial. DoRA for eg learns an auto scaling vector of numbers which tries to alleviate these issues.

Galore might be more equivalent to full pretraining with the gradients being low rank.

Do you mean leaving most of the model in its initial, randomised state and only training a LoRA?

I’ve tested specifically this (on my personal time) :) It will train but I found the loss is proportional to the number of trainable parameters. So roughly to hit the performance of a standard 70m param model, you need to train ~70m lora params anyway.

It's worse than that, because lora requires two matrices per layer. At full rank, you have an additional NxN parameters to learn versus full finetuning, where N is min(input_features, output_features).

For example, tuning a layer of 128 in x 256 out is 32k params. Learning a full-rank lora for that layer would be two matrices of 128x128 and 128x256 = 48k params.

Yeah, exactly. Though the 48k param lora might be as good as a 48k param layer of higher rank, I haven't looked into that case really.

Yes, I’ve tested this out. It does train, but the scaling doesn’t seem to pan out. It’ll perform slightly better than the number of trainable parameters, but never improves as you scale, so for now there’s no benefit.

i would be shocked if this worked well

Reply on news.ycombinator.com