[thot_experiment]

I'm curious about the quantization quality claims in the table there. Is this a Gemma 2 specific thing (more subtlety in the weights somehow?). In my testing and testing I've seen elsewhere at least for llama3 8B (and some less rigorous testing with other models) q_8 -> q4_K_M are basically indistinguishable from one another?

[janwas]

Yes, PPL and certain benchmarks do not detect differences from quantization. But recent work gives cause for concern, e.g., https://arxiv.org/pdf/2310.01382, https://arxiv.org/pdf/2405.18137.

[luke-stanley]

The first paper is good to critique the performance of quantised models, it points out that 40-50% 'compression' typically results in only slight loss for RAG tasks relying on in-context learning, but for factual tasks replying on stored knowledge, performance very quickly dropped off. They looked at Vicuna, one of the earlier models, so I wonder how applicable it is to recent models like the Phi 3 range. I don't think deliberate clever adversarial attacks like those of the 2nd paper are a sensible worry for most, but it is fun. Thanks for the links @janwas.