Reference
Companion experiment: run-to-run variance
A separate, smaller experiment that tests one assumption the main comparison relies on: that running the same inputs again returns the same answers.
Qwen 2.5 3B, vLLM 0.12.0, one dedicated Baseten deployment on a single NVIDIA L4.
20 gold questions, replayed at concurrency 1, 16 and 32. Temperature 0.
Does batching under load change what the model produces?
Concurrency is the variable that matters here: it is what changes how requests are batched together, and batch composition is the most commonly cited reason that identical requests can return different text at temperature 0. If batching were going to perturb outputs, a concurrency sweep is where it would show.
Result: it did not. At every level, 15 of 20 raw intents matched the expected answer, and zero intents differed from the sequential (concurrency-1) baseline — the outputs were bit-identical across the sweep. The companion throughput measurement over the same session was likewise highly reproducible: across three repeated 150-request sweeps, throughput varied by under 1% at each level.
The wider point is the reason this page exists at all: repeatability is something you measure and scope, not something a setting guarantees. Tokens, context and generation works through why that is, and what to measure instead.