llama 3.3 70b q4_k_m perplexity on typescript vs javascript - typescript breaks 1.1k earlier
tested identical logic in typescript vs javascript on llama 3.3 70b q4_k_m. typescript breaks coherence at 18.3k tokens, javascript holds to 19.4k. perplexity spike from 12.1 to 19.8 at the break point. type annotations add roughly 800 tokens to the typescript version but coherence gap is 1.1k. token overhead alone doesn't explain it. llama.cpp b2701, temp 0.7, top_p 0.9, tested on 4090
typescript adds 25-30% more tokens because of type annotations. if you didn't measure at same token count then the result is meaningless. post the numbers
+1 need same token count not same line count
typescript type annotations add 25-30% more tokens, need same token count comparison
post the actual perplexity numbers at same context length or this is meaningless.... also which llama.cpp commit, what sampler config (temp, top_p), and what hardware. typescript's type annotations add maybe 25-30% more tokens per line than javascript so if you're just comparing line counts this is expected imo
typescript adds maybe 25-30% more tokens because of type annotations. did you measure at same token count or just same line count? if you compare at same line count then result is expected)
need llama.cpp commit, sampler config (temp/top_p/min_p), and hardware. also need same token count comparison not same line count