|
|
--- |
|
|
inference: false |
|
|
base_model: TheDrummer/Behemoth-R1-123B-v2 |
|
|
base_model_relation: quantized |
|
|
tags: |
|
|
- exl2 |
|
|
library_name: exllamav2 |
|
|
pipeline_tag: text-generation |
|
|
--- |
|
|
|
|
|
exllamav2 quantizations of TheDrummer's [Behemoth-R1-123B-v2](https://huggingface.co/TheDrummer/Behemoth-R1-123B-v2) |
|
|
|
|
|
[2.25bpw h6](https://huggingface.co/MikeRoz/Behemoth-R1-123B-v2-exl2/tree/2.25bpw_H6) (32.964 GiB) |
|
|
[4.25bpw h6](https://huggingface.co/MikeRoz/Behemoth-R1-123B-v2-exl2/tree/4.25bpw_H6) (61.324 GiB) |
|
|
[5.00bpw h6](https://huggingface.co/MikeRoz/Behemoth-R1-123B-v2-exl2/tree/5.00bpw_H6) (71.959 GiB) |
|
|
[8.00bpw h8](https://huggingface.co/MikeRoz/Behemoth-R1-123B-v2-exl2/tree/8.00bpw_H8) (114.559 GiB) |
|
|
[measurement.json](https://huggingface.co/MikeRoz/Behemoth-R1-123B-v2-exl2/resolve/main/measurement.json?download=true) |
|
|
|
|
|
The 2.25bpw quant will load with 28k fp16 context on 2 24 GB GPUs, or 89k fp16 context on 3 24 GB GPUs. |
|
|
The 4.25bpw quant will squeeze into 3 24GB GPUs with 16k fp16 context, but can load with 73k of fp16 context in 4 24GB GPUs. |
|
|
The 8.00bpw quant requires 6 24 GB GPUs (or equivalent) |
|
|
|
