search

OpenGVLab / OmniQuant

[ICLR2024 spotlight] OmniQuant is a simple and powerful quantization technique for LLMs.
MIT License
626 stars 49 forks source link

Questions about quantization #81

Open mxjmtxrm opened 1 month ago

mxjmtxrm commented 1 month ago

Hi, great work! I met some problems during 4bit weight-only quantization(--lwc).

  1. Is there any problem if the norm is nan?
  2. what's the best lwc hyper-parameter of LLama2 with different scales? like lwc-lr and epochs?
  3. Does more calib data bring better results?

I quantized a llama model using different lwc hyper-parameters and received different results.

  1. nsamples=1000, batch_size=1, epoch=2, the ppl is correct.
  2. nsamples=2000, batch_size=8, epoch=10, the ppl is super large (40000+). What's the problem?