System Info

• transformers version: 4.57.3
• huggingface_hub version: 0.36.2
• Python: 3.12
• OS: Linux (Ubuntu 24.04, inside NVIDIA container)

Who can help?

@ArthurZucker @itazap

Regression introduced in

PR #42389 ([Mistral Tokenizers] Fix tokenizer detection), included in v4.57.2 → v4.57.3.

Information

• The official example scripts
• My own modified scripts

Tasks

• An officially supported task in the examples folder
• My own task or dataset (give details below)

Reproduction

import os
from huggingface_hub import snapshot_download

# Step 1: Pre-download a non-Mistral model
snapshot_download("Qwen/Qwen3-0.6B")

# Step 2: Enable offline mode
os.environ["HF_HUB_OFFLINE"] = "1"

# Step 3: Load tokenizer — crashes even though model is fully cached
from transformers import AutoTokenizer
tok = AutoTokenizer.from_pretrained("Qwen/Qwen3-0.6B")

This raises:

huggingface_hub.errors.OfflineModeIsEnabled: Cannot reach
https://huggingface.co/api/models/Qwen/Qwen3-0.6B: offline mode is enabled.

Full traceback:

transformers/models/auto/tokenization_auto.py:1156, in from_pretrained
transformers/tokenization_utils_base.py:2113, in from_pretrained
transformers/tokenization_utils_base.py:2395, in _from_pretrained
transformers/tokenization_utils_base.py:2438, in _patch_mistral_regex
transformers/tokenization_utils_base.py:2432, in is_base_mistral
huggingface_hub/hf_api.py:2660, in model_info
→ OfflineModeIsEnabled

Expected behavior

AutoTokenizer.from_pretrained() should work in offline mode (HF_HUB_OFFLINE=1) when the model is already cached locally. This worked in transformers 4.57.1.

Root cause

In tokenization_utils_base.py, _patch_mistral_regex defines is_base_mistral() which calls huggingface_hub.model_info() — an API call — for every model, not just Mistral models. The call happens before any cache check or local_files_only guard:

# tokenization_utils_base.py, _patch_mistral_regex
def is_base_mistral(model_id: str) -> bool:
    model = model_info(model_id)  # <-- unconditional API call
    ...

if _is_local or is_base_mistral(pretrained_model_name_or_path):  # <-- called for ALL non-local models

The local_files_only parameter is passed to _patch_mistral_regex but is never used to guard the is_base_mistral() call.

Suggested fix

The is_base_mistral() call should either:

1. Be wrapped in a try/except that catches OfflineModeIsEnabled and returns False (safe default — if we can't reach the API, assume it's not a Mistral model), or
2. Be skipped when local_files_only=True or HF_HUB_OFFLINE=1 is set

Impact

This breaks any CI/CD pipeline or air-gapped environment that:

1. Pre-downloads models with snapshot_download()
2. Sets HF_HUB_OFFLINE=1 to prevent network access during test execution
3. Loads tokenizers via AutoTokenizer.from_pretrained()

This pattern is common in ML CI pipelines (we hit this in NVIDIA Dynamo's CI with TensorRT-LLM).