Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

GPT-3.5 crashes when it thinks about useRalativeImagePath too much

iter.ca

Post reply on HN
Print view View original

51–60 of 164 posts

Re: GPT-3.5 crashes when it thinks about useRalativeImagePath too much

#51

Post by shawntan »

This is a strange explanation. These models usually give as output the same set of vocabulary that was used as its input vocabulary.

> the model isn’t trained on understanding the useRalativeImagePath token, and so it outputs something that isn’t a valid token.

In my view, either the model sees this token and then gets into a spiral of random generations because the `useRalativeImagePath` embedding is just a completely random vector, or it just chugs on trying to maintain coherent-sounding text.

The set of tokens it can _output_ is, however, fixed, so unless the displayable tokens on the interface is a subset of the full vocabulary of tokens, it should always be 'valid'.

Re: GPT-3.5 crashes when it thinks about useRalativeImagePath too much

#52

Post by shawntan »

febeling wrote: post #48
brilee wrote: post #11

Most likely it has badly conditioned embedding vectors for those particular tokens, leading the network to edge into numerically unstable territory; once you get some sort of underflow or NaN, they tend to propagate and invalidate the entire output. If there are any batchnorm or other operations that mix values between different entries in a batch, you could even cause other peoples' sessions to return junk values!

That sounds wild. While being ignorant about LLMs internals, I would have expected such things, crashes and session leaks, be impossible by design.

NaNs are not only possible by design, but are extremely common. Training of LLMs involve many tricks about how to deal with training steps that result in NaNs. Quantisation of LLMs also require dealing with huge outlier values.

Re: GPT-3.5 crashes when it thinks about useRalativeImagePath too much

#53

Post by airgapstopgap »

febeling wrote: post #48
brilee wrote: post #11

Most likely it has badly conditioned embedding vectors for those particular tokens, leading the network to edge into numerically unstable territory; once you get some sort of underflow or NaN, they tend to propagate and invalidate the entire output. If there are any batchnorm or other operations that mix values between different entries in a batch, you could even cause other peoples' sessions to return junk values!

That sounds wild. While being ignorant about LLMs internals, I would have expected such things, crashes and session leaks, be impossible by design.

Note that we have no reason to believe that the underlying LLM inference process has suffered any setbacks. Obviously it has generated some logits. But the question is how is OpenAI server configured and what inference optimization tricks they're using.

Re: GPT-3.5 crashes when it thinks about useRalativeImagePath too much

#56

Post by isoprophlex »

elevaet wrote: post #39
silisili wrote: post #38

I know it's not good faith to complain about a site design rather than it's content, but please don't do whatever this is to your background. As someone with regular ocular migraines, opening this on mobile made my anxiety shoot straight up thinking I'm having another.

As someone else who has regular ocular migraines, this causes absolutely no anxiety for me. I can kind of see the resemblance but it's pretty obviously not one. And if it was, is it really so bad getting them? For me it's just a weird optical effect that passes... I really don't think anyone needs to go out of their way to accommodate my quirk.

I have infrequent migraines and this scared the ever living fuck out of me. Mine are of the "wouldn't you rather be dead?" variety when it comes to headache and vomiting, though.

Re: GPT-3.5 crashes when it thinks about useRalativeImagePath too much

#57

Post by IshKebab »

silisili wrote: post #38

I know it's not good faith to complain about a site design rather than it's content, but please don't do whatever this is to your background. As someone with regular ocular migraines, opening this on mobile made my anxiety shoot straight up thinking I'm having another.

Weird unpleasant background for sure but it's obviously not that because it doesn't follow your eyes. Don't be daft.

Re: GPT-3.5 crashes when it thinks about useRalativeImagePath too much

#58

Post by shawntan »

minimaxir wrote: post #5

> As a result, the model isn’t trained on understanding the useRalativeImagePath token, and so it outputs something that isn’t a valid token. That isn't how LLMs generate tokens. Each step outputs a logit for each possible token in the tokenizer (100k in the case of GPT-3.5), then softmaxes the logits to covert them into probabilities, and samples from them depending on temperature to get the token to be used. It's p…

Exactly this. The tokens generated should always be valid, unless some post-processing layer between the model's output and the user interface detects for some keywords which it would prefer to filter out. In which case I suppose there is another commonly seen error message that appears?

Re: GPT-3.5 crashes when it thinks about useRalativeImagePath too much

#60

Post by SirSegWit »

shawntan wrote: post #58
minimaxir wrote: post #5

> As a result, the model isn’t trained on understanding the useRalativeImagePath token, and so it outputs something that isn’t a valid token. That isn't how LLMs generate tokens. Each step outputs a logit for each possible token in the tokenizer (100k in the case of GPT-3.5), then softmaxes the logits to covert them into probabilities, and samples from them depending on temperature to get the token to be used. It's p…

Exactly this. The tokens generated should always be valid, unless some post-processing layer between the model's output and the user interface detects for some keywords which it would prefer to filter out. In which case I suppose there is another commonly seen error message that appears?

Not really, right? There are a ton of special tokens, like start of sequence etc., so what happens if there are two start of sequences predicted? It's a valid token but cannot really be turned into something sensible, so it throws an error when converting tokens to plain text?
Post reply on HN