Very interesting trick, using a dictionary of basis vectors which are quickly computed from a seed without storage. But the result is the same 3 or 4 bit quantization, with only a slight improvement. Their tiles are small, just 8 or 12 weights, it's why compression doesn't go too far. It would have been great if this trick lowered quantization <1 bit/weight, that would require longer tiles. Wondering what are the limits if we use a larger reservoir of cheap entropy as part of neural net architecture, even in training.
Congrats to Apple and Meta, makes sense they did the research, this will go towards efficient serving of LLMs on phones. And it's very easy to implement.
I was about to post something similar. While the research is interesting, it doesn’t offer any advantages over 3- or 4-bit quantization. I also have to assume they explored using longer tiles but found it to be ineffective — which would make sense to me from an information theory perspective.
> it doesn’t offer any advantages over 3- or 4-bit quantization.
"zero-shot accuracy retention at 4- and 3-bit compression to be on par with or better than state-of-the-art methods, while maintaining performance comparable to FP16 baselines."
My reading of that says FP16 accuracy at Q3 or Q4 size / memory bandwidth. Which is a huge advantage.
I think the main advantage is that you can compute the extra parameters (the PRNG seeds) from the network weights alone, whereas most other quantization methods require simulating the quantization procedure at training time (Quantization-Aware Training) or setting them from a calibration dataset (Post-Training Quantization)
This technique has three significant advantages over popular low bit quantization: 1) it retains more accuracy, 2) it does not require calibration data, 3) it's easier to implement in hardware.
It sounds like they basically find part of a pseudo-random sequence that is closest to the desired data, then store the random seed and corrections (which are small so take less space).
Pretty fascinating from an information theory point of view. Surprising that it works at all. Is this, like, the JPEG of uniformly distributed, uncorrelated data?
We don't know. They basically look for sequences that approximate NN weights well, in the same way sinusoidal functions work well with "natural" images, but not with graphics with hard edges.
> What makes this technique particular to LLM weights
This is my understanding as a non-expert.
LLM activations tend to be relatively sparse with large outliers. With linear quantization, this means you either have to clip off the outliers or you have to stretch your range to include the outliers, which wastes precious bits. Neither of these works well, so essentially all LLM quantization research is using various heuristics to get around these outliers. For example, you can do linear quantization but split the activations up into smaller blocks to make it less likely that any given block contains an outlier.
Another trick people have discovered (predates LLMs) is applying a random rotation/projection to the embeddings. This has the effect of making sure no one dimension in the vector dominates the others (which again hurts quantization). This works because in order for a single dimension to dominate, all the others have to "conspire" to be near zero. When you have 10,000+ dimensions, that's very unlikely.
This paper applies the latter trick. Instead of pre-generating the random projection matrices, they generate them on the fly on the accelerator from a seed that is fixed for each block. The seed is chosen from an offline brute-force search that needs only the weights of the network. This separates it from a lot of other quantization methods that either require calibration data or have to be simulated at training time so the network learns the quantization parameters itself.
You might think this is wasteful/might hurt performance, but it turns out that LLM inference is heavily memory-bound as it involves streaming a very large neural network into the accelerator (GPU/TPU/NPU/whatever) to operate on a relatively small amount of data, so there are lots of "free cycles" to generate these random numbers. Of course, if you care about power usage that might not be a great idea...
Weights in neural networks don't always need to be precise. Not all weights are equally useful to the network.
There seems to be a lot of redundancy that can be replaced with approximations.
This technique seems a bit similar to lossy image compression that replaces exact pixels with a combination of pre-defined patterns (DCT in JPEG), but here the patterns aren't from cosine function, but from a pseudo-random one.
It may also be beating simple quantization from just adding noise that acts as dithering, and breaks up the bands created by combinations of quantized numbers.
A variant I have been thinking of: each parameter matrix (or block) is the sum of a random matrix (generated from a seed) and a low rank matrix (a LoRA). I'd like to experiment training from scratch in that setting.
I suspect that you are literally the only person on this planet who would find this to be funny enough for Apple to waste the time of a dozen AI Researchers, Meta, arXiv and Apple Legal who vet everything.
I submit that there is no "compression into PRNG seeds" going on here. This is just a quantisation method that happens to leverage PRNs, which might have some specific advantages and disadvantages. What I am sure it does not do, is what it's title seems to claim, if taken literally. I suspect they're having a good laugh, getting away with what they must know is borderline trolling. I'm impressed!
Readit News
Congrats to Apple and Meta, makes sense they did the research, this will go towards efficient serving of LLMs on phones. And it's very easy to implement.
"zero-shot accuracy retention at 4- and 3-bit compression to be on par with or better than state-of-the-art methods, while maintaining performance comparable to FP16 baselines."
My reading of that says FP16 accuracy at Q3 or Q4 size / memory bandwidth. Which is a huge advantage.
1: https://libraryofbabel.info/
2: https://news.ycombinator.com/item?id=9480949
[1] https://arxiv.org/abs/2410.10714
This is my understanding as a non-expert.
LLM activations tend to be relatively sparse with large outliers. With linear quantization, this means you either have to clip off the outliers or you have to stretch your range to include the outliers, which wastes precious bits. Neither of these works well, so essentially all LLM quantization research is using various heuristics to get around these outliers. For example, you can do linear quantization but split the activations up into smaller blocks to make it less likely that any given block contains an outlier.
Another trick people have discovered (predates LLMs) is applying a random rotation/projection to the embeddings. This has the effect of making sure no one dimension in the vector dominates the others (which again hurts quantization). This works because in order for a single dimension to dominate, all the others have to "conspire" to be near zero. When you have 10,000+ dimensions, that's very unlikely.
This paper applies the latter trick. Instead of pre-generating the random projection matrices, they generate them on the fly on the accelerator from a seed that is fixed for each block. The seed is chosen from an offline brute-force search that needs only the weights of the network. This separates it from a lot of other quantization methods that either require calibration data or have to be simulated at training time so the network learns the quantization parameters itself.
You might think this is wasteful/might hurt performance, but it turns out that LLM inference is heavily memory-bound as it involves streaming a very large neural network into the accelerator (GPU/TPU/NPU/whatever) to operate on a relatively small amount of data, so there are lots of "free cycles" to generate these random numbers. Of course, if you care about power usage that might not be a great idea...
This technique seems a bit similar to lossy image compression that replaces exact pixels with a combination of pre-defined patterns (DCT in JPEG), but here the patterns aren't from cosine function, but from a pseudo-random one.
It may also be beating simple quantization from just adding noise that acts as dithering, and breaks up the bands created by combinations of quantized numbers.
It covers some experiments on weight tying, one of which is actually LoRA and random weights.
Does he mean they did pretraining but not fine tuning?
In general, compression using PRNGs is not a thing. There might be a special exception for this case, but I somewhat doubt it. =)