Paper flags 19 batch-invariance violations in 83 released GPU kernel artifacts

[arXiv] According to an arXiv preprint, LLM serving systems dynamically batch user requests for GPU execution, so the same request may run alongside different requests across runs; if GPU kernels lack batch invariance (BI), this changing batch context can alter numerical output for an unchanged input and undermine reproducibility. The authors say they confirmed 19 BI violations among 83 batch-applicable released artifacts. The paper proposes a tool that combines contract-guided static dependency analysis with dynamic validation of bitwise consistency across batch contexts, designed as a plug-in for general kernel agents and used as successive acceptance gates in the optimization loop, rejecting structurally risky or dynamically non-invariant candidates before performance benchmarking and returning feedback to the agent. The authors report that across three agent frameworks on real-world LLM serving tasks, the tool raised the BI pass rate from 67.6% to 91.9% while retaining 91.8% of the baseline geometric mean speedup. The work is a preprint and has not been peer-reviewed.