keyfan/Mixtral-8x7B-Instruct-2bit

This is 2-bit quantization of mistralai/Mixtral-8x7B-Instruct-v0.1 using QuIP#

Model loading

Please follow the instruction of QuIP-for-all for usage.

As an alternative, you can use vLLM branch for faster inference. QuIP has to launch like 5 kernels for each linear layer, so it's very helpful for vLLM to use cuda-graph to reduce launching overhead. BTW, If you have problem installing fast-hadamard-transform from pip, you can also install it from source

Perplexity

Measured at Wikitext with 4096 context length

fp16	2-bit
3.8825	5.2799

Speed

Measured with examples/benchmark_latency.py script at vLLM repo. At batch size = 1, it generates at 16.3 tokens/s with single 3090.