q4 coherence on code vs prose at same token count
tested llama 3.3 70b q4_k_m yesterday at 18.9k context. prose generation (blog posts, documentation) stays coherent until 21.3k. code generation (python functions with docstrings) breaks at 19.1k. same model, same quant, similar token count. failure modes are totally different too. prose just gets repetitive and drifts off topic. code invents function names and parameters that follow conventions perfectly but don't exist. anyone know why code breaks earlier?
hit this yesterday at 19.3k. q4_k_s breaks way earlier than q4_k_m on same workload.
we measured q4_k_s vs q4_k_m on code generation last month and saw almost identical delta - q4_k_s broke around 17.8k, q4_k_m at 19.2k. what exact failure mode are you seeing - invented function names, syntax errors, or semantic drift?