API requests are made despite setting local_files_only=True.

[haok1402]

System Info

API requests are made despite setting local_files_only=True when loading a pretrained tokenizers.

Here is the relevant part of the code that triggers the requests.

transformers/src/transformers/tokenization_utils_tokenizers.py

Lines 1199 to 1211 in 9495ae2

	def is_base_mistral(model_id: str) -> bool:
	model = model_info(model_id)
	if model.tags is not None:
	if re.search("base_model:.*mistralai", "".join(model.tags)):
	return True
	return False
	if is_offline_mode():
	is_local = True
	if pretrained_model_name_or_path is not None and (
	is_local or (not is_local and is_base_mistral(pretrained_model_name_or_path))
	):

model_info inside is_base_mistral triggers this.

Who can help?

No response

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

AutoTokenizer.from_pretrained("Qwen/Qwen3-30B-A3B", local_files_only=True)

Expected behavior

There should be no network requests being made with local_files_only=True.

[vaibhav-research]

checking this, will let you know the outcome and approach to fix.

[vaibhav-research]

I was able to consistently reproduce Hub API requests being made despite local_files_only=True, even when all required tokenizer/config files are already present in the local cache.

Environment
• transformers: local checkout (current main)
• huggingface_hub: recent version (uses httpx)
• Python 3.11
• macOS

Step 1 — Pre-cache tokenizer + config files
I first ensured that no file downloads are required at load time by explicitly caching only the tokenizer/config artifacts:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="Qwen/Qwen3-30B-A3B",
    allow_patterns=[
        "config.json",
        "tokenizer.json",
        "tokenizer.model",
        "tokenizer_config.json",
        "special_tokens_map.json",
        "added_tokens.json",
        "merges.txt",
        "vocab.json",
        "*.tiktoken",
    ],
)

This completes successfully and places all required files in the local HF cache.

Step 2 — Load tokenizer with local_files_only=True
I then loaded the tokenizer while explicitly disabling network access:

from transformers import AutoTokenizer

AutoTokenizer.from_pretrained(
    "Qwen/Qwen3-30B-A3B",
    local_files_only=True,
)

To observe any network activity, I enabled httpx / httpcore debug logging.

Observed behavior:
Despite local_files_only=True and all files being available locally, the following Hub API calls are made:

HTTP Request: GET https://huggingface.co/api/models/Qwen/Qwen3-30B-A3B/revision/main
HTTP Request: GET https://huggingface.co/api/models/Qwen/Qwen3-30B-A3B

These requests are visible in httpx/httpcore debug logs and return 200 OK.

This indicates that model metadata is being fetched from the Hub, even though no downloads are needed and local_files_only=True is set.

Comparison with another model

I repeated the exact same procedure with:

"mistralai/Mistral-7B-v0.1"

After pre-caching tokenizer/config files in the same way:
• AutoTokenizer.from_pretrained(..., local_files_only=True) completes successfully
• No HTTP requests are made
• No calls to /api/models/... appear in the logs

So the behavior appears model-specific, not a general tokenizer loading issue.

Likely root cause
Based on the endpoints being hit (/api/models/...), this looks like a model metadata lookup that is not respecting local_files_only and is bypassing offline guards.

digging more to find the exact solution for this and will open a PR.

cc @zucchini-nlp

[vaibhav-research]

also, I tested this with the following models

MODELS = [
    "Qwen/Qwen3-30B-A3B",
    "Qwen/Qwen2.5-7B-Instruct",
    "Qwen/Qwen2-7B-Instruct",
    "mistralai/Mistral-7B-v0.1",
    "TinyLlama/TinyLlama-1.1B-Chat-v1.0",
    "microsoft/phi-3-mini-4k-instruct",
    "gpt2",
]

and the issue seems to be only related to Qwen models

Step 1: prepared cache:

python prepare_cache.py
Fetching 5 files: 100%|█████████████████████████████████████████████████████████████████| 5/5 [00:00<00:00, 17742.40it/s]
Download complete: : 0.00B [00:00, ?B/s]              ✅ cached minimal: Qwen/Qwen3-30B-A3B        | 0/5 [00:00<?, ?it/s]
Fetching 5 files: 100%|█████████████████████████████████████████████████████████████████| 5/5 [00:00<00:00, 20580.49it/s]
Downloading (incomplete total...): 0.00B [00:00, ?B/s]✅ cached minimal: Qwen/Qwen2.5-7B-Instruct
Fetching 5 files: 100%|█████████████████████████████████████████████████████████████████| 5/5 [00:00<00:00, 15627.06it/s]
                                                      ✅ cached minimal: Qwen/Qwen2-7B-Instruct
Download complete: : 0.00B [00:00, ?B/s] [00:00, ?B/s]
Download complete: : 0.00B [00:00, ?B/s]                                                           | 0/5 [00:00<?, ?it/s]
Download complete: : 0.00B [00:00, ?B/s]
Fetching 5 files: 100%|█████████████████████████████████████████████████████████████████| 5/5 [00:00<00:00, 13017.70it/s]
                                                      ✅ cached minimal: mistralai/Mistral-7B-v0.1
Fetching 5 files: 100%|████████████████████████████████████████████████████████████████████| 5/5 [00:00<00:00,  8.06it/s]
Download complete: : 2.34MB [00:00, 8.23MB/s]              ✅ cached minimal: TinyLlama/TinyLlama-1.1B-Chat-v1.07.69it/s]
Fetching 6 files: 100%|████████████████████████████████████████████████████████████████████| 6/6 [00:00<00:00, 15.76it/s]
Downloading (incomplete total...): 504kB [00:00, 1.63MB/s]✅ cached minimal: microsoft/phi-3-mini-4k-instruct            
Fetching 5 files: 100%|████████████████████████████████████████████████████████████████████| 5/5 [00:00<00:00, 29.44it/s]
✅ cached minimal: gpt2 total...):  37%|████████████████                            | 1.04M/2.85M [00:00<00:00, 6.70MB/s]
Download complete: : 0.00B [00:01, ?B/s]████████████████████████████████████████████| 2.85M/2.85M [00:00<00:00, 6.70MB/s]
Download complete: : 2.34MB [00:01, 1.71MB/s]█████▏                                        | 2/5 [00:00<00:00, 12.35it/s]
Download complete: : 2.44MB [00:00, 3.73MB/s]
Download complete: 100%|████████████████████████████████████████████████████████████| 2.85M/2.85M [00:00<00:00, 15.3MB/s]

Step 2:

turned off the wifi and then tried to load the model.

python repro.py        

=== Qwen/Qwen3-30B-A3B ===
❌ failed: HTTPX NETWORK CALL DETECTED: GET https://huggingface.co/api/models/Qwen/Qwen3-30B-A3B

=== Qwen/Qwen2.5-7B-Instruct ===
❌ failed: HTTPX NETWORK CALL DETECTED: GET https://huggingface.co/api/models/Qwen/Qwen2.5-7B-Instruct

=== Qwen/Qwen2-7B-Instruct ===
❌ failed: HTTPX NETWORK CALL DETECTED: GET https://huggingface.co/api/models/Qwen/Qwen2-7B-Instruct

=== mistralai/Mistral-7B-v0.1 ===
✅ loaded: <class 'transformers.tokenization_utils_tokenizers.TokenizersBackend'>

=== TinyLlama/TinyLlama-1.1B-Chat-v1.0 ===
✅ loaded: <class 'transformers.tokenization_utils_tokenizers.TokenizersBackend'>

=== microsoft/phi-3-mini-4k-instruct ===
✅ loaded: <class 'transformers.tokenization_utils_tokenizers.TokenizersBackend'>

=== gpt2 ===
✅ loaded: <class 'transformers.models.gpt2.tokenization_gpt2.GPT2Tokenizer'>

Hence the issue seems to be only related to Owen models.

cc @zucchini-nlp

[haok1402]

@vaibhav-research Thanks for the reproduction!! This is a very clean, and elegant manner to demonstrate the issue.

[haok1402]

I was running tokenization across a very large corpus, which gets distributed into multiple nodes, and forked among many processes; even with local_files_only=True, I hit my API requests limit immediately. So I started to suspect something is off.

[vaibhav-research]

@haok1402 Thanks for sharing that context. that makes a lot of sense.

I’ve traced this to a Qwen-specific tokenizer path where a Mistral regex patch check calls into model_info() without respecting offline mode. I’m working on a fix to short-circuit that lookup when running local-only so tokenizer initialization stays fully offline. I’ll follow up once the PR is ready.