[menaerus]

> It's usually the case that memory access is also slower than matrix multiplications so this will be faster. Burning flops to save memory movement.

I haven't read this paper (yet) but isn't this the case that mostly applies to training and not so much to inference? A good example would be flash-attention, it trades the higher flops for better memory utilization but it's mostly irrelevant in inference workloads.

[verdverm]

They claim an inference time savings to the kv cache

[menaerus]

I skimmed through the paper real quickly. There's no performance data on inference speedups in the paper. Only the benchmarks relevant for training.

They also, interestingly, don't compare against the flash-attention. Flash-attention outperforms all of the other attention mechanisms mentioned in the paper: MHA, MQA, GQA, and MLA.

[apophis-ren]

Flash attention is an implementation trick; you can implement MHA/GQA, for example, with flash attention.

[verdverm]

They aren't claiming speedups, they are claiming up to an order of magnitude less space needed for the kv cache at runtime. This translates to a smaller GPU or longer sequences in the same GPU

[menaerus]

Under what circumstances can you cut down your LOADS and STORE from and to main memory by an order of magnitude without observing major improvements in algorithm runtime that is memory-bound?

[verdverm]

AI models are compute bound, it's why we use GPUs

[menaerus]

Incorrect. Self-attention is a highly parallel algorithm that makes it a great candidate for being a memory-bound workload once you have enough compute.

Both datacenter grade CPUs and GPUs have enough compute to carry out the self-attention computation but it is only the latter that has enough hi-bandwidth memory to make the algorithm really perform. If this hadn't been the case, the theory behind flash-attention wouldn't materialize, and it does, and reason being that (main) memory is slow.

Deep FFWD networks OTOH are compute-bound.