Llama.cpp 30B runs with only 6GB of RAM now

Has anyone done any comprehensive analysis on exactly how much quantization affects the quality of model output? I haven't seen any more than people running it and being impressed (or not) by a few sample outputs.

I would be very curious about some contrastive benchmarks between a quantized and non-quantized version of the same model.

I've done some experiments here with Llama 13B, in my subjective experience the original fp16 model is significantly better (particularly on coding tasks). There are a bunch of synthetic benchmarks such a wikitext2 PPL and all the whiz bang quantization schemes seem to score well but subjectively something is missing.

I've been able to compare 4 bit GPTQ, naive int8, LLM.int8, fp16, and fp32. LLM.int8 does impressively well but inference is 4-5x slower than native fp16.

Oddly I recently ran a fork of the model on the ONNX runtime, I'm convinced that the model performed better than pytorch/transformers, perhaps subtle differences in floating point behavior etc between kernels on different hardware significantly influence performance.

The most promising next step in the quantization space IMO has to be fp8, there's a lot of hardware vendors adding support, and there's a lot of reasons to believe fp8 will outperform most current quantization schemes [1][2]. Particularly when combined with quantization aware training / fine tuning (I think OpenAI did something similar for GPT3.5 "turbo").

If anybody is interested I'm currently working on an open source fp8 emulation library for pytorch, hoping to build something equivalent to bitsandbytes. If you are interested in collaborating my email is in my profile.

1. https://arxiv.org/abs/2208.09225 2. https://arxiv.org/abs/2209.05433

Isn't that related to architecture? The most recent GPUs and tensor procs have native support for 4-bit(partially) and 8-bit int whereas older GPUs take noticeable performance hits for 8-bit vs fp16/32.