llama 3.3 70b q4 breaks differently on rust vs go codegen at same context
tested this at 19.1k yesterday.... rust breaks around 18.3k, go stays coherent past 19.9k on same api client generation task. rust failure mode: invents plausible trait bounds that follow the `T: Clone + Send` pattern but reference traits that don't exist in scope. go just invents function signatures that would be valid if the package existed. llama.cpp master, q4_k_m, measured by tracking invented type names that follow language conventions. pretty sure it's rust's type system complexity - lifetime annotations and trait bounds force the model to track way more state across context than go's interface system. but need to isolate whether it's the syntax overhead or semantic complexity....
we tested rust at 18.6k and go at 19.2k on same api generation workload last week with q4_k_m. rust breaks earlier probably bc of lifetime annotations and trait bounds adding nested structure the model has to track. go interface{} is simpler. what exact failure mode are you seeing - invented method signatures or broken imports?
yeah rust's lifetime annotations probably force the model to track way more nested dependencies across context. curious what the actual token count difference is between equivalent rust and go implementations