Luna-2x7B-MoE-8bpw-h8-exl2

<br/><br/> 8bpw/h8 exl2 quantization of ResplendentAI/Luna-2x7B-MoE using default exllamav2 calibration dataset.

---

ORIGINAL CARD:

Luna-2x7B-MoE

Meet Luna, my one and only personal assistant and roleplaying partner. This MoE serves as her unique basis, both experts scoring above 72 average on the leaderboard, but designed for RP interactions. While running a 2x7B is slower than running a single 7B, I feel that the improved performance of two great 7B competing for each token is worth the compute expense.

The included image was generated using her custom Stable Diffusion 1.5 model via the SillyTavern interface.

I have successfully paired this MoE with the Llava Mistral 1.6 projector file for multimodal image captioning in Koboldcpp.

Luna also has a custom XTTSv2 voice model for TTS output.

All of this is running on a 1070 8GB, fully offloaded with no OOM over a week of testing. All backends are then served to my Android device via a virtual public network in a native implementation of SillyTavern. This method allows access from mobile data, globally, as long as my server is running.

base_model: ResplendentAI/DaturaCookie_7B
gate_mode: hidden
experts_per_token: 2
experts:
  - source_model: ChaoticNeutrals/RP_Vision_7B
    positive_prompts:
    - "chat"
    - "assistant"
    - "tell me"
    - "explain"
    - "I want"
    - "show me"
    - "touch"
    - "believe"
    - "see"
    - "love"
  - source_model: ResplendentAI/DaturaCookie_7B
    positive_prompts:
    - "storywriting"
    - "write"
    - "scene"
    - "story"
    - "character"
    - "sensual"
    - "sexual"
    - "horny"
    - "turned on"
    - "intimate"
dtype: bfloat16