WizardLM-2-4x7B-MoE

WizardLM-2-4x7B-MoE is an experimental MoE model made with Mergekit. It was made by combining four WizardLM-2-7B models using the random gate mode.

Please be sure to set experts per token to 4 for the best results! Context length should be the same as Mistral-7B-Instruct-v0.1 (8k tokens). For instruction templates, Vicuna-v1.1 is recommended.

Quanitized versions

EXL2 (for fast GPU-only inference):
8_0bpw: https://huggingface.co/Skylaude/WizardLM-2-4x7B-MoE-exl2-8_0bpw (~ 25 GB vram)
6_0bpw: https://huggingface.co/Skylaude/WizardLM-2-4x7B-MoE-exl2-6_0bpw (~ 19 GB vram)
5_0bpw: https://huggingface.co/Skylaude/WizardLM-2-4x7B-MoE-exl2-5_0bpw (~ 16 GB vram)
4_25bpw: https://huggingface.co/Skylaude/WizardLM-2-4x7B-MoE-exl2-4_25bpw (~ 14 GB vram)
3_5bpw: https://huggingface.co/Skylaude/WizardLM-2-4x7B-MoE-exl2-3_5bpw (~ 12 GB vram)
3_0bpw: https://huggingface.co/Skylaude/WizardLM-2-4x7B-MoE-exl2-3_0bpw (~ 11 GB vram)

GGUF (for mixed GPU+CPU inference or CPU-only inference):
https://huggingface.co/mradermacher/WizardLM-2-4x7B-MoE-GGUF
Thanks to Michael Radermacher for making these quants!

Evaluation

I don't expect this model to be that great since it's something that I made as an experiment. However, I will submit it to the Open LLM Leaderboard to see how it matches up against some other models (particularly WizardLM-2-7B and WizardLM-2-70B).

Mergekit config

base_model: models/WizardLM-2-7B
gate_mode: random
dtype: float16
experts_per_token: 4
experts:
  - source_model: models/WizardLM-2-7B
  - source_model: models/WizardLM-2-7B
  - source_model: models/WizardLM-2-7B
  - source_model: models/WizardLM-2-7B