Comment by miki123211
14 hours ago
In theory, temperature 0 does make the LLM deterministic.
Well, in theory theory, temperature 0 doesn't really exist. Mathematically, as lim temperature->0, the distribution gets spikier and spikier, the most likely sample goes to almost-but-not-quite infinity and the rest go to almost-but-not-quite 0. In practice, temperature=0 is literally a separate branch of an if statement that just picks the most common sample (using the actual formula that works for non-zero values would cause a zero division).
However, due to things such as batching and even different kinds of floating point imprecisions for different algorithm implementations, the probability distribution itself often differs run-by-run, so what you sample from it also differs.
>in theory theory, temperature 0 doesn't really exist.
It does exist very much, even if you go to pure math. Look at the softmax function and take the limit as T->0. It becomes a dirac-delta function. I.e. in a discrete setting (like for LLMs with a finite set of output tokens), probability P becomes one for argmax and 0 for everything else. Only in coding practice it is easer to implement T=0 as a simple if check that directly chooses argmax instead of calculating the limit of some function that includes 1/T quotients. But setting T to zero is in both, theory and practice, turning the usual probability function into greedy sampling.
> Look at the softmax function and take the limit as T->0. It becomes a dirac-delta function.
In pure math, it does not always do that. It becomes a dirac-delta comb with equal weight on every maximum. There can be more than 1 maximum. Setting the temperature to zero turns into greedy sampling, but greedy sampling is not necessarily deterministic as you can have multiple equally optimal options.
That is not a problem for LLMs, because in practice floating point inaccuracies (in particular after exponentiation) prevent values from being exactly equal. That's why greedy sampling generally produces deterministic output for LLMs. The real gotchas are elsewhere (like with batch inference as we've seen with earlier GPTs). But unlike what the earlier comment says, this is a non-issue mathematically.
6 replies →
> It becomes a dirac-delta function. I.e. in a discrete setting (like for LLMs with a finite set of output tokens), probability P becomes one for argmax and 0 for everything else. Only in coding practice it is easer to implement T=0 as a simple if check that directly chooses argmax instead of calculating the limit of some function that includes 1/T quotients.
I don't understand the distinction you're drawing. A Dirac delta function is a "simple if check".
The point is that the case T=0 doesn't just "exist" as a special code branch - it is still well defined mathematically without any change to the output function. What the above comment refers to with the extra "if" check is just a limitation of computers not liking to divide anything by zero, even if the actual function exists and is well behaved at zero. It is not some weird or special theoretical construction.
7 replies →
Even if it's deterministic that doesn't mean it isn't arbitrary. I can achieve determinism at any temperature by saving the seed. But that wouldn't make rejects feel much better knowing that if a bit was flipped in an arbitrary seed they would be scored differently.
I did large scale tests temp 0 and there was still randomness with the same prompt inputs coming in.
I did this with several model apis.
GPU processing is not going to be the same from what I read but also the AI backend is doing a lot of fancy batching resulting in another layer of randomness.
> Mathematically, as lim temperature->0, the distribution gets spikier and spikier, the most likely sample goes to almost-but-not-quite infinity and the rest go to almost-but-not-quite 0.
That's not how limits work. As the temperature goes to 0, the rest goes to 0. That's it. The "almost-but-not-quite" is part of the "goes to".
Let's say f(x) = 3x+1. It's a continuous function. If we let x go to 10, f(x) goes to 31. Not "almost-but-not-quite 31". No, to 31. (If you don't have a continuous function then it's the same argument, but less intuitive to illustrate.)
> However, due to things such as batching and even different kinds of floating point imprecisions for different algorithm implementations, the probability distribution itself often differs run-by-run, so what you sample from it also differs.
Exactly. While I’m assuming this won’t be news for most here, for those that are still new and/or curious about some more explanation on e.g. the floating-point imprecisions, see this nice article: https://thinkingmachines.ai/blog/defeating-nondeterminism-in...
It is not deterministic because the order of computations in a typical multithreaded system is not deterministic and also because when combined with the devil that is IEEE754, it gets even less deterministic.
As I understood it, the "randomness" affecting what is selected at any temperature still comes from a PRNG or CSPRNG (or whatever RNG you want, maybe a hardware one), and if you where to swap out that with something deterministic you'd get the same results every time (barring non-determinism in other parts of the OS/drivers/maybe even hardware).
But theoretically, the output of every LLM is seed-driven (or could be if you wrote the software to isolate it) just like any computer software. It's just none of the software written (even llama.cpp AFAIK) chooses to support stable-seeding due to the changes in stuff like CPU/Vulkan/CUDA/Metal differences making it difficult to make consistent.
They could though! Hopefully one day someone implements it into the mainstream LLM-engine software and it gets exposed in the APIs serving the models. It'd do a lot to show folks the "internals" of these models.
It's probably due to the fact that it's a cloud service. You have no guarantee that your next request will go to the same machine. So even with an identical seed, and temp 0 you might get different hardware and hence different accuracy/noise in the floating point operations.
How can there be noise in floating point operations? I could buy like completion order for parallized batches i.e. adding a+b+c is different from a+c+b etc.
7 replies →
Stable seeding is not enough. A lot of modern, fast compute kernels are nondeterministic. Floating point multiplication/addition is not strictly associative and e.g. reductions can combine results from different threads in different orders (e.g. through atomic ops). You can write kernels to be deterministic, but it is generally less efficient.
They are only non-deterministic when you’re doing batching and a kernel ends up running across a “random” set of token streams. If you’re only processing one user’s request, they’re very much deterministic.
that's incorrect in the presence of batching. it's tough work making it truly deterministic:
https://x.com/FireworksAI_HQ/status/2069873437217276015
It's not that hard. What is hard is making it truly deterministic and retain high throughput.
PRNG is deterministic.
If you make an exact integer implementation and run with temp=0 it's deterministic.
You don't even need temperature 0, just make a random seed for the sampler part of the input and then its deterministic as a function of the input.
But running autoregressive models at temp=0 tends to expose pathological behavior, because the training process produces a function with a lot of gain so its prone to feedback on its own noise.
> However, due to things such as batching and even different kinds of floating point imprecisions for different algorithm implementations, the probability distribution itself often differs run-by-run
The implementation does not often differ run by run.
> The implementation does not often differ run by run.
If you use a cluster, or even multiple clusters, and they have non-identical hardware, then two consecutive runs could end up being routed to nodes having different GPU models with slightly different floating point behaviour, or even software differences (e.g. newer GPU offers some feature usable to speed up calculations which older model lacked; same code can use the feature when it is available, fall back to slower alternative if it isn’t). The larger your scale, the greater the odds it will happen