Bluesky Thread

“LLMs are deterministic!”

View original thread
“LLMs are deterministic!”

no they’re not

“well, if you set temperature = 0”

still no

“and remove floating point calculations?”

keep going

“and only serve one query at a time?”

you’re so close
47 2
Thinking Machines wrote a superb blog post on determinism in LLMs and it is fascinating (if you like computer architecture otherwise you’re going to tear your hair out)

spoiler: did you know GPUs have instructions that aren’t deterministic?

thinkingmachines.ai/blog/defeati...
thinkingmachines.ai
Defeating Nondeterminism in LLM Inference
Reproducibility is a bedrock of scientific progress. However, it’s remarkably difficult to get reproducible results out of large language models. For example, you might observe that asking ChatGPT the...
67 15
while you’re at it, Groq (not Grok) does fast inference by removing all non-determinism, ALL THE WAY DOWN

it’s a cool trade-off. they basically turn hardware into a compiler problem

blog.codingconfessions.com/p/groq-lpu-d...
blog.codingconfessions.com
The Architecture of Groq's LPU
What powers the ground breaking performance of Groq's Langauge Processing Unit?
19
4 hours later
strong “buckle up!” vibes
Batch Invariance and "Determinism"
To explain batch invariance, let's simplify the system and look solely at matmuls.
You can assume that all matmul implementations are "run-to-run deterministic."* However, they are not
4 This is not totally true, but most common matmul implementations do have this property.
"batch-invariant." In other words, when the batch size changes, each element in the batch can get different results.
3
47 likes 2 reposts

More like this

×