News for hackers. Stuff that matters.
Re:Lossless Acceleration of LLM via Adaptive N-Gram Parallel... by SonOfLilit on Sunday April 21, 2024 @03:11PM (#40108308)
Good idea. Generate 8 tokens with a small model, then give them to a large model as one batch (much faster) and if tokens 1..3 are in agreement but 4..8 nut, take only 1..4 and start again. Most tokens are easy to guess so you'll get x3.5 gains.

I do have a feeling of dejavú, like I've seen this before on hn.

Re:Lossless Acceleration of LLM via Adaptive N-Gram Parallel... by johntb86 on Sunday April 21, 2024 @03:48PM (#40108632)
They mention previous work on speculative decoding using similar techniques, but "ANPD dynamically generates draft outputs via an adaptive N-gram module using real-time statistics, after which the drafts are verified by the LLM. This characteristic is exactly the difference between ANPD and the previous speculative decoding methods."
Re:Lossless Acceleration of LLM via Adaptive N-Gram Parallel... by milkey_mouse on Sunday April 21, 2024 @03:43PM (#40108605)
> I do have a feeling of dejavú, like I've seen this before on hn.

You're either thinking of speculative encoding more generally, or Medusa: https://arxiv.org/abs/2401.10774

Re:Lossless Acceleration of LLM via Adaptive N-Gram Parallel... by outofpaper on Sunday April 21, 2024 @03:44PM (#40108607)
I think a number of us are using the method and these folks decided to do a paper and release their findings.
Re:Lossless Acceleration of LLM via Adaptive N-Gram Parallel... by jimmySixDOF on Monday April 22, 2024 @03:28AM (#40112110)
Lots of speedups already on Llama3 like Groq (hardware arch) and Modal [1] getting 200 t/s

From the below Modal link they use: >continuous batching, so multiple generations can take place at the same time on a single container

>PagedAttention, which applies memory paging to the attention mechanism’s key-value cache, increasing throughput

[1]https://modal.com/docs/examples/text_generation_inference

Re:Lossless Acceleration of LLM via Adaptive N-Gram Parallel... by sdenton4 on Sunday April 21, 2024 @03:16PM (#40108369)
Very reminiscent of parallel wavenet, which speed things up by generating multiple audio samples at a time.
Re:Lossless Acceleration of LLM via Adaptive N-Gram Parallel... by contravariant on Sunday April 21, 2024 @06:59PM (#40109959)
Just so I understand, this is essentially faster because reading all the model weights takes more time than calculating the model output?
Re:Lossless Acceleration of LLM via Adaptive N-Gram Parallel... by SonOfLilit on Sunday April 21, 2024 @08:27PM (#40110461)
GPUs are great at doing the same math on every item of a large multidimensional array.

Therefore, unsurprisingly, the cost per item of inference on a batch of items is significantly lower when the batch is e.g. 8 than 1 (in the case of Transformers there are further gains to be made because roughly half of the attention calculations in token k+1 are identical to the calculations of token k and can be easily reused by writing the formulas a certain way, the keyword to look for is causal attention mask).

In any reasonable GPU inference setup the weights would be preloaded.

Re:Lossless Acceleration of LLM via Adaptive N-Gram Parallel... by contravariant on Monday April 22, 2024 @04:08AM (#40112290)
Indeed GPUs are great at doing the same calculation in parallel. But if it was just that there should be enough opportunity to parallelise even without doing the exact same calculation multiple times.

The main reason I can come up with why doing the same calculation 8 times in parallel instead of 8 times sequentially is that you benefit from better locality of reference.

Re:Lossless Acceleration of LLM via Adaptive N-Gram Parallel... by SonOfLilit on Monday April 22, 2024 @12:00PM (#40115613)
As I said, the attention step is O(n^2) per token sequentially and O(n^2) for the entire sequence when calculating the entire sequence in parallel, where n is the length of the sequence.