FLA's KDA safe_gate fast path leaks future gates into prefix outputs; fix coming
YouJiacheng · x · 2026-09-22
A causal-interference bug was reported in flash-linear-attention (5.8k stars): on the current main branch, the CUDA/Triton KDA chunk path with usegateinkernel=True and safegate=True breaks prefix causal non-interference — inputs identical through prefix position i produce different prefix outputs, prefix-only loss, and gradients when only a later raw gate changes, even without altering the prefix. Runs are bitwise identical, ruling out ordinary nondeterminism.
DrissP wrote up an analysis of the bug, and maintainers confirmed the fix will land in the next FLA release.
More from Research
- HWREBench: AI researcher hacks Amazon smart devices daily to benchmark hardware reverse engineering — johnowhitaker · 2026-09-22
- TMLR submissions quadruple on AI-generated influx; ICLR 2027 caps single authors — petitegeek · 2026-09-22
- Stanford's VirtualBiotech puts tens of thousands of AI scientist agents in Science, NYT reports — StanfordAILab · 2026-09-22
- Irit Dinur wins Gödel Prize for her landmark 2005 proof of the PCP theorem — willcb · 2026-09-22
- Researcher proposes using RL to teach AI when to give up — sqcai · 2026-09-22
- Nature publishes Delphy: scalable near-real-time Bayesian phylogenetics for outbreak tracking — burny_tech · 2026-09-22