alt Hacker NewsNice to see a provider being transparent about KV cache quantisation. I've been suspecting that some providers do this silently whilst heavily promoting their unquantised weights, even though KV quantisation can degrade quality more than weight quantisation.
However, I wish their testing were more detailed. Firstly, some model families are more sensitive to KV quantisation than others (only Kimi K2.6 was tested). Secondly, the evaluation suite they use to claim that FP8 KV quantisation is indistinguishable is noticeably lacking coding benchmarks; in long-running tasks, minor tool call errors compound over time.
Replies
vLLM's study also concluded that "FP8 can deliver meaningful latency and capacity gains with small or negligible accuracy loss". Their benchmarks include LiveCodeBench 6.
They made an extremely strong claim:
> None of this would matter if it changed the model's answers
If they want to assert that the answers don’t change, then perhaps they should calculate the statistical distance between the token probability outputs or something to that effect. I doubt the results would indicate that the answers don’t change by any reasonable interpretation.
Maybe the results are still good enough.