llama 3.3 70b q4 - coherence on go vs python, go breaks ~1.4k earlier
been testing llama 3.3 70b q4_k_m on code generation and noticing go consistently breaks earlier than python. go loses coherence around 17.9k, python holds to 19.3k imo this is probably the interface definitions and error handling - go has way more explicit interface satisfaction checks and every function returns (result, error) which adds tokens but could be wrong, might just be that my go prompts are structured differently. anyone else seeing this or is it just my setup
what are the exact token counts for identical go vs python code. go has way more type annotation and error handling overhead so comparing equal line counts is meaningless
lol yeah go's error handling is so verbose it probably burns like 30% more tokens than equivalent python with try/except
token overhead alone doesn't explain a 1.4k gap imo. go's error handling creates way more branching paths in the execution trace which probably forces the model to track more potential states simultaneously
what quant? getting way worse coherence on go than you are
q4_k_m on llama.cpp b2701. go breaks at 17.8k, python holds to 19.2k on identical logic. token overhead from type annotations is roughly 1.1k but coherence gap is 1.4k
go's explicit error handling probably creates way more branching paths the model has to track across long context. every `if err != nil { return err }` is another state split python with implicit exceptions collapses all that into the happy path until something actually throws