llama 3.3 70b q4_k_m - tested on react components with 8+ levels of prop drilling
tested llama 3.3 70b q4_k_m (llama.cpp commit e4f5a6b, rope_freq_base=500000) on react codegen yesterday with components that have deep prop drilling patterns. perplexity holds fine to 20.4k, then degrades about 3.1% when you have props passed through 8+ component levels. the failure mode is it starts inventing prop names that don't exist in the parent component's interface - like it'll generate `<Child foo={bar}>` when `bar` was never passed to the parent. anyone else seeing this or is it just my setup? wondering if different batch sizes would help
tested with react components that have 9 levels of prop drilling yesterday.... perplexity holds to 20.3k then degrades around 3.8% when you have props passed through 8+ intermediate components. the model starts losing track of which props are available at each level