llama 3.3 70b q4_k_m - does schema validation order affect coherence past 20k
been testing llama 3.3 70b q4_k_m on json schema validation tasks and seeing weird behavior. when i validate schemas in document order (top to bottom) coherence breaks around 21.3k tokens. when i validate in dependency order (referenced schemas first) it holds to 24.1k. same schemas, same json, just different validation traversal order. has anyone else seen ordering affect coherence like this? trying to figure out if this is real or if im measuring noise. running llama.cpp main branch from 3 days ago, temp 0.7, top_p 0.9, no min_p. hardware is single 4090. specific question: does the order that model sees structured tokens actually matter for long context, or is this just variance in my setup?
schema validation order shouldn't affect model coherence at all unless your validation logic is doing something completely insane like mutating the context or recursively traversing the entire schema tree for every token. this sounds like either a measurement artifact or you're conflating validation latency with coherence degradation
probly not, unless your schema validation is doing something weird like recursive deep traversal that blows up token attention patterns. schema validation order shouldnt affect model coherence, it affects parse success. post the actual schemas if you want real feedback