Information

GPT4-X-Alpaca 30B 4-bit working with GPTQ versions used in Oobabooga's Text Generation Webui and KoboldAI.

This was made using Chansung's GPT4-Alpaca Lora

Update 05.26.2023

Updated the ggml quantizations to be compatible with the latest version of llamacpp (again).

What's included

GPTQ: 2 quantized versions. One quantized --true-sequential and act-order optimizations, and the other was quantized using --true-sequential --groupsize 128 optimizations

GGML: 3 quantized versions. One quantized using q4_1, another one was quantized using q5_0, and the last one was quantized using q5_1.

GPU/GPTQ Usage

To use with your GPU using GPTQ pick one of the .safetensors along with all of the .jsons and .model files.

Oobabooga: If you require further instruction, see here and here

KoboldAI: If you require further instruction, see here

CPU/GGML Usage

To use your CPU using GGML(Llamacpp) you only need the single .bin ggml file.

Oobabooga: If you require further instruction, see here

KoboldAI: If you require further instruction, see here

Training Parameters

num_epochs=10
cutoff_len=512
group_by_length
lora_target_modules='[q_proj,k_proj,v_proj,o_proj]'
lora_r=16
micro_batch_size=8

Benchmarks

--true-sequential --act-order

Wikitext2: 4.481280326843262 Ptb-New: 8.539161682128906 C4-New: 6.451964855194092 Note: This version does not use --groupsize 128, therefore evaluations are minimally higher. However, this version allows fitting the whole model at full context using only 24GB VRAM.

--true-sequential --groupsize 128

Wikitext2: 4.285132884979248 Ptb-New: 8.34856128692627 C4-New: 6.292652130126953 Note: This version uses --groupsize 128, resulting in better evaluations. However, it consumes more VRAM.