Mixed precision insight for Llama3-8B

[ZipXuan]

I wonder whether the strategy use_more_bit for different layers inspired by Llama-v1 can also be the better policy for Llama-v3：

For llama3-8B, my experiment compares three different plans when token embedding is Q4_K and LM head is Q6_K:

1. Tensors in all layers are Q4_K（row0）
2. Choose four consecutive layers for Q6_K separately (32 in total) (row1~8)
3. Quantize four layers in leaps to Q6_K(row9~10)

Observations:
1.The first few layers are important for generation quality
2.The last few layers and the selection of a few layers in a jump may not be as important for generation quality as originally believed in use_more_bit strategy

Maybe we need more different mixed-precision insights for different LLM models rather than only use_more_bit or some toolkits for quantization sensitivity analysis

[ggerganov]

An automated per-model strategy for distributing the bits would be great to have. I am not sure what is the best way to achieve it. At some point I was thinking about a tool that compares the activations per-layer and applies some optimization strategy to improve the distribution of bits (#2783). We now how the tools (e.g. eval_callback), so we should be able to perform the necessary analysis and devise different strategies. We can also implement the strategy used by picoLLM. Suggestions are welcome