• llama_in_sunglasses@alien.topB
    link
    fedilink
    English
    arrow-up
    1
    ·
    11 months ago

    To create quants of new models, one has to create Hessians for it and it uses several GB of RedPajama to calibrate these. Generating Hessians for Mistral is taking 17 minutes per LAYER on my 3090. I’ll see if it can even finish later. Much later. That’s over 16 hours just to quantize a 7B model, yikes.

    The paper for this is one of the worst for me in years, full on “I know some of these words.” I didn’t think 8-dimensional sphere packing was going to be in my attempted light reading for the night.

    P…S.: Rollback to transformers 4.34.0 or edit the code in hessian_offline_llama.py and change all instances of

    attention_mask = model.model._prepare_decoder_attention_mask(
    

    to

    attention_mask = _prepare_4d_causal_attention_mask(
    

    and add an import to the top of the same file.

    from transformers.modeling_attn_mask_utils import _prepare_4d_causal_attention_mask
    
    • llama_in_sunglasses@alien.topB
      link
      fedilink
      English
      arrow-up
      1
      ·
      11 months ago

      With Llama-2-70b-chat-E8P-2Bit from their zoo, quip# seems fairly promising. I’d have to try l2-70b-chat in exl2 at 2.4 bpw to compare but this model does not really feel like a 2 bit model so far, I’m impressed.

      • a_beautiful_rhind@alien.topB
        link
        fedilink
        English
        arrow-up
        1
        ·
        11 months ago

        From the issue about this in the exllamav2 repo, quip was using more memory and slower than exl. How much context can you fit?