magnum-12b-v2.5-kto-iMat-GGUF

Important Note: Inferencing in llama.cpp has now been merged in PR #8604. Please ensure you are on release b3438 or newer. Text-generation-web-ui (Ooba) is also working as of 7/23. Kobold.cpp working as of v1.71.

Quantized from magnum-12b-v2.5-kto fp16

• Weighted quantizations were creating using fp16 GGUF and groups_merged.txt (special thanks to Kalomaze) in 92 chunks and n_ctx=512
• Static fp16 will also be included in repo
• For a brief rundown of iMatrix quant performance please see this PR
• All quants are verified working prior to uploading to repo for your safety and convenience

KL-Divergence Reference Chart (Click on image to view in full size)

Quant-specific Tips:

• If you are getting a cudaMalloc failed: out of memory error, try passing an argument for lower context in llama.cpp, e.g. for 8k: -c 8192
• If you have all ampere generation or newer cards, you can use flash attention like so: -fa
• Provided Flash Attention is enabled you can also use quantized cache to save on VRAM e.g. for 8-bit: -ctk q8_0 -ctv q8_0

Original model card can be found here