I've seen lots of headlines like "1-bit quantization" (including this one and https://arxiv.org/abs/2310.16795). What I've found in this space is that the headlines can often be intentionally misleading about what is actually achieved. If you read closer the abstract of this paper, it claims 8.41 perplexity on LLaMA2-70B at 1 bit, which is a HUGE decrease from 3.120 perplexity in FP16 and they will never mention that in the headline. Even LLaMA2-7B at INT8 achieves 5.677 perplexity with half of storage place (better with LESS space and LESS training). Some claim 1.58-bit quantization (each weight is either -1, 0 or 1) but in practice require very small group sizes, which means one or two extra FP16 numbers for every 64 weights and that means another 0.5 bit, so it's actually 2-bit quantization. And every quantization algorithm can claim they make language models smaller, speedier, and use less energy, so there's nothing special about these.
Here are the key metrics that I suggest checking when comparing quantization schemes:
* Perplexity. Note that it also depends on the dataset (either WikiText2 or C4, WikiText2 numbers are usually lower than C4) and context size (1024, 2048 or 4096, higher context sizes usually means less perplexity). Dataset and context size must match to make a meaningful comparison.
* Quantization bits. Many algorithms claiming 2-bit or 1-bit quantization has lots of extra parameters elsewhere, such as grouping. Download the quantized version of the model, check its file size, multiply by 8 and divide by the number of parameters. That gets you the ACTUAL quantization bits.
* Performance. Weights may need to be dequantized during inference which could introduce overhead. Some libraries have custom matmul kernel for dequantization that achieves performance close to FP16, others can be slower. Check its generation speed and inference speed.
Newer architectures such as Ampere contains INT8 cores which may make quantized version even faster than FP16, I haven't tried out yet.
There is also a lot of misleading comparisons in this space. Some methods like GPTQ only provide vector-matrix multiplication kernels, which means only a single token can be generated and batched inference (which is needed for generating initial KV cache, or for serving multiple users) can be much slower. If an algorithm claims a 3x speedup for something, check if they refer to single stream latency or multiple stream throughput. Some of that speedup comes from running a model on 2 cards instead of 5 cards without specifying if the cards have NVLink configured (you shouldn't run inference on multiple cards without NVLink, or you should expecet huge slowdown simply because of using 5 cards).
* Base model. Pick a STRONG base model like Llama-2-7b or Llama-3-8b etc. Not an undertrained model like SwitchTransformer etc which may have lots of redundant parameters in itself.
My personal favourite remains QuIP# (https://github.com/Cornell-RelaxML/quip-sharp). It lacks in the "performance" part as its matrix multiplication performance isn't on par yet but there is room for improvement, and it wins every other metric. And sad news: it's very likely we won't have practical 1-bit LLMs, never ever. We are reaching the end game between 2.5~4 bits. By "practical" I mean it should beat 3-bit LLMs with 3x less parameters or 2-bit LLMs with half as many parameters. There is a Shannon limit to quantization whatever methods you use.