Update eos_token_id / bos_token_id in config.json
#19
by nickhugs - opened
The current values don't appear to even be in the vocab. This makes it match the recent updates done to the tokenizer config.
Can confirm that this may cause bugs. To reproduce, take the example from the README and pass min_new_tokens to generate:
from transformers import AutoModelForCausalLM, AutoTokenizer
checkpoint = "bigcode/santacoder"
device = "cuda"
tokenizer = AutoTokenizer.from_pretrained(checkpoint)
inputs = tokenizer.encode("def print_hello_world():", return_tensors="pt").to(device)
outputs = model.generate(inputs, min_new_tokens=5) # <== change
# raises IndexError: index 50256 is out of bounds for dimension 1 with size 49280
Passing min_new_tokens results in the logits_processor containing a MinNewTokensLengthLogitsProcessor with the incorrect eos_token_id, which causes the IndexError. Other additional arguments may have similar effects, I haven't tested.
thanks for the fix!
loubnabnl changed pull request status to merged