Hacker News Reader

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!

View on HN

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.

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.

The operation of this server is very uniform, in my imagination. Just emitting chunks of string. That this can be disrupted and an edge case occur, by the content of the strings - I find it puzzling.

Reply on news.ycombinator.com