llama 3.3 70b q4_k_m - tested on python with many async/await chains

i tested llama 3.3 70b q4_k_m yesterday on python code with async function chains (like 6-8 awaits chained together in single expression). perplexity was stable until around 20.3k context, then degrade maybe 3.4% when you have deeply nested async comprehensions with multiple await points. model starts to confuse execution order and sometimes generates await in wrong position. anyone else seeing this or is my rope config wrong?