metadata

inference: false
language:
  - en
pipeline_tag: text-generation
tags:
  - Yi
  - llama
  - llama-2
license: other
license_name: yi-license
license_link: LICENSE
datasets:
  - jondurbin/airoboros-2.2.1
  - lemonilia/LimaRP

airoboros-2.2.1-limarpv3-y34b-exl2

Exllama v2 quant of Doctor-Shotgun/airoboros-2.2.1-limarpv3-y34b

Branches:

main: measurement.json calculated at 2048 token calibration rows on PIPPA
4.65bpw-h6: 4.65 decoder bits per weight, 6 head bits
- ideal for 24gb GPUs at 8k context (on my 24gb Windows setup with flash attention 2, peak VRAM usage during inference with exllamav2_hf was around 23.4gb with 0.9gb used at baseline)
6.0bpw-h6: 6 decoder bits per weight, 6 head bits
- ideal for large (>24gb) VRAM setups