Nous-Capybara-34B and Tess-M-Creative-v1.0 merged, then quantized with exllamav2 on 200 rows (400K tokens) on a long Vicuna format chat, a sci fi story and a fantasy story. This should hopefully yield better chat performance than the default wikitext quantization.

Quantized to 4bpw, enough for ~47K context on a 24GB GPU.

The following merge config was used:

models:
  - model: /home/alpha/Storage/Models/Raw/larryvrh_Yi-34B-200K-Llamafied
    # no parameters necessary for base model
  - model: /home/alpha/Storage/Models/Raw/migtissera_Tess-M-v1.0
    parameters:
      density: 0.6
      weight: 1.0
  - model: /home/alpha/Storage/Models/Raw/Nous-Capybara-34B
    parameters:
      density: 0.6
      weight: 1.0
merge_method: ties
base_model: //home/alpha/Storage/Models/Raw/larryvrh_Yi-34B-200K-Llamafied
parameters:
  normalize: true
  int8_mask: true
dtype: float16

First exllama quantization pass:

python convert.py --in_dir /home/alpha/FastModels/Capybara-Tess-Yi-34B-200K -o /home/alpha/FastModels/Capybara-Tess-Yi-34B-200K-exl2 -om /home/alpha/FastModels/capytessmes.json --cal_dataset /home/alpha/Documents/smol.parquet -l 2048 -r 80 -ml 2048 -mr 40 -gr 40 -ss 4096 -nr -b 3.5 -hb 6

Second exllama quantization pass:

python convert.py --in_dir /home/alpha/FastModels/Capybara-Tess-Yi-34B-200K -o /home/alpha/FastModels/Capybara-Tess-Yi-34B-200K-exl2 -m /home/alpha/FastModels/capytessmes.json --cal_dataset /home/alpha/Documents/medium.parquet -l 2048 -r 200 -ml 2048 -mr 40 -gr 200 -ss 4096 -b 3.1 -hb 6 -cf /home/alpha/FastModels/Capybara-Tess-Yi-34B-200K-exl2-31bpw -nr

Both models have Vicuna syntax, so:

Prompt Format:

SYSTEM: ...
USER: ...
ASSISTANT: ...

Stop token: </s>

Credits:

https://github.com/cg123/mergekit

https://huggingface.co/NousResearch/Nous-Capybara-34B/discussions

https://huggingface.co/migtissera/Tess-M-Creative-v1.0

https://huggingface.co/larryvrh/Yi-34B-200K-Llamafied

https://huggingface.co/01-ai/Yi-34B-200K