max_position_embeddings and tokenizer max discrepancies

#22
by ghpu - opened

It seems that tokenizer_config has a small value for "model_max_length": 16384
while in config.json, "max_position_embeddings": 131072.

In Qwen/Qwen2.5-1.5B, we had : "model_max_length": 131072

Same issue for some of the responses think tag is missing which likely due to a truncated output

Sign up or log in to comment