Inconsistency between fast and slow codellama tokenizers

[UniverseFly]

System Info

• transformers version: 4.33.2
• Platform: Linux-5.15.0-82-generic-x86_64-with-glibc2.35
• Python version: 3.10.12
• Huggingface_hub version: 0.16.4
• Safetensors version: 0.3.3
• Accelerate version: 0.22.0
• Accelerate config: not found
• PyTorch version (GPU?): 2.0.1+cu117 (True)
• Tensorflow version (GPU?): not installed (NA)
• Flax version (CPU?/GPU?/TPU?): not installed (NA)
• Jax version: not installed
• JaxLib version: not installed
• Using GPU in script?: no
• Using distributed or parallel set-up in script?: no

Who can help?

@ArthurZucker

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

A simple reproduce:

from transformers import AutoTokenizer

t_fast = AutoTokenizer.from_pretrained("codellama/codellama-7B-Instruct-Hf", use_fast=True)
t_slow = AutoTokenizer.from_pretrained("codellama/codellama-7B-Instruct-Hf", use_fast=False)

ids_fast = t_fast.encode("<s>[INST]", add_special_tokens=False)
ids_slow = t_slow.encode("<s>[INST]", add_special_tokens=False)

assert ids_fast == ids_slow, f"Fast: {ids_fast}, Slow: {ids_slow}"
# AssertionError: Fast: [1, 518, 25580, 29962], Slow: [1, 25580, 29962]

Expected behavior

I'm not sure which one is correct. Actually decoding the fast tokenizer outputs will get '<s> [INST]', while the slow tokenizer '<s>INST]', both not same as the original string.

[ArthurZucker]

Hey! This was already reported and is a duplicate of #25881 and related to #26318 as well. Will be fixed in tokenizers soon!

[ArthurZucker]

The PR is looking good, had a few delays but was able to reproduce 1-1 with fast but also with sentencepiece actual token addition.