| Update (4/1): Added ggml for Cuda model | |
| Dataset is here (instruct): https://github.com/teknium1/GPTeacher | |
| Okay... Two different models now. One generated in the Triton branch, one generated in Cuda. Use the Cuda one for now unless the Triton branch becomes widely used. | |
| Cuda info (use this one): | |
| Command: | |
| CUDA_VISIBLE_DEVICES=0 python llama.py ./models/chavinlo-gpt4-x-alpaca --wbits 4 --true-sequential --groupsize 128 --save gpt-x-alpaca-13b-native-4bit-128g-cuda.pt | |
| Prev. info | |
| Quantized on GPTQ-for-LLaMa commit 5955e9c67d9bfe8a8144ffbe853c2769f1e87cdd | |
| GPTQ 4bit quantization of: https://huggingface.co./chavinlo/gpt4-x-alpaca | |
| Note: This was quantized with this branch of GPTQ-for-LLaMA: https://github.com/qwopqwop200/GPTQ-for-LLaMa/tree/triton | |
| Because of this, it appears to be incompatible with Oobabooga at the moment. Stay tuned? | |
| Command: | |
| CUDA_VISIBLE_DEVICES=0 python llama.py ./models/chavinlo-gpt4-x-alpaca --wbits 4 --true-sequential --act-order --groupsize 128 --save gpt-x-alpaca-13b-native-4bit-128g.pt | |