llama 3.3 70b q4_k_m - perplexity on terraform configs with lots of nested modules
1. tested on production terraform with 8+ levels of module nesting yesterday 2. perplexity holds fine to 20.2k 3. spikes hard around 20.9k when you have complex variable interpolation in nested modules 4. failure mode is the model starts hallucinating resource references that don't exist in the dependency graph the exact pattern that breaks is something like `module.foo.module.bar.module.baz.output.value` - past 20.9k it invents module paths. are other people using llama 3.3 for terraform generation seeing similar issues with nested module references?
we hit this with terraform configs that have 6+ levels of module nesting.... perplexity holds fine until around 20.8k then spikes when you have recursive module calls with lots of variable interpolation. the model starts mixing up which variables belong to which module scope
post perplexity curves or it's vibes
post perplexity curves with exact context sizes or this is just vibes. everyone says "it breaks around 20k" but without actual numbers we can't tell if this is rope config, quantization artifacts, or model issue
what nesting depth breaks it? we use llama 3.3 70b q4_k_m for terraform generation and modules nest 5-7 levels deep in our setup