Exllama v2 Quantizations of internlm2-math-20b-llama
Using turboderp's ExLlamaV2 v0.0.12 for quantization.
The "main" branch only contains the measurement.json, download one of the other branches for the model (see below)
Each branch contains an individual bits per weight, with the main one containing only the meaurement.json for further conversions.
Original model: https://huggingface.co./internlm/internlm2-math-20b
Branch | Bits | lm_head bits | Size | Description |
---|---|---|---|---|
6_5 | 6.5 | 8.0 | 21.0 GB | Near unquantized performance at vastly reduced size, recommended. |
4_25 | 4.25 | 6.0 | 15.2 GB | GPTQ equivalent bits per weight, slightly higher quality. |
3_5 | 3.5 | 6.0 | 13.8 GB | Lower quality, only use if you have to. |
3_0 | 3.0 | 6.0 | 12.5 GB | Very low quality. Usable on 12GB if you reduce context or use 8 bit cache. |
All VRAM requirements estimated from 16k context. For 32k context add ~2 GB.
Download instructions
With git:
git clone --single-branch --branch 6_5 https://huggingface.co./bartowski/internlm2-math-20b-llama-exl2 internlm2-math-20b-llama-exl2-6_5
With huggingface hub (credit to TheBloke for instructions):
pip3 install huggingface-hub
To download the main
(only useful if you only care about measurement.json) branch to a folder called internlm2-math-20b-llama-exl2
:
mkdir internlm2-math-20b-llama-exl2
huggingface-cli download bartowski/internlm2-math-20b-llama-exl2 --local-dir internlm2-math-20b-llama-exl2 --local-dir-use-symlinks False
To download from a different branch, add the --revision
parameter:
Linux:
mkdir internlm2-math-20b-llama-exl2-6_5
huggingface-cli download bartowski/internlm2-math-20b-llama-exl2 --revision 6_5 --local-dir internlm2-math-20b-llama-exl2-6_5 --local-dir-use-symlinks False
Windows (which apparently doesn't like _ in folders sometimes?):
mkdir internlm2-math-20b-llama-exl2-6.5
huggingface-cli download bartowski/internlm2-math-20b-llama-exl2 --revision 6_5 --local-dir internlm2-math-20b-llama-exl2-6.5 --local-dir-use-symlinks False