Llama3 tokenizer decode is incorrect for ' ...' with leading space

[Naqu6]

System Info

• transformers version: 4.49.0
• Platform: Linux-5.4.0-187-generic-x86_64-with-glibc2.31
• Python version: 3.12.9
• Huggingface_hub version: 0.29.1
• Safetensors version: 0.5.3
• Accelerate version: not installed
• Accelerate config: not found
• DeepSpeed version: not installed
• PyTorch version (GPU?): 2.6.0+cu124 (True)
• Tensorflow version (GPU?): not installed (NA)
• Flax version (CPU?/GPU?/TPU?): not installed (NA)
• Jax version: not installed
• JaxLib version: not installed
• Using distributed or parallel set-up in script?: no
• Using GPU in script?: yes
• GPU type: NVIDIA L40S

Who can help?

Hi @ArthurZucker @itazap (tagging you per instructions), when I use the Llama3 tokenizers and encode the string ' ...', and then decode the resulting token, I get the string '...' back instead of the string ' ...' (leading space is missing).

I believe that decode should be the inverse of encode in this case, and it's unclear to me why it isn't.

Sorry if I'm misunderstanding something! Thanks for your time :)

Information

• The official example scripts
• My own modified scripts

Tasks

• An officially supported task in the examples folder (such as GLUE/SQuAD, ...)
• My own task or dataset (give details below)

Reproduction

from transformers import AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("meta-llama/Meta-Llama-3-8B-Instruct")

tokenizer.decode(tokenizer.encode(" ...")[1:])

This outputs '...' (no leading space).

Expected behavior

I believe that decode should be the inverse of encode.

E.g

from transformers import AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("meta-llama/Meta-Llama-3-8B-Instruct")

tokenizer.decode(tokenizer.encode(" Hello world")[1:]) #[1:] to remove beginning of sequence token

outputs " Hello world", as expected.

[Naqu6]

See #35938 (comment)