ValueError: trust_remote_code=True required
transformers.modeling_utils.RepositoryNotFoundError / ValueError (trust_remote_code parameter)
Stack trace
Traceback (most recent call last):
File "app.py", line 42, in <module>
model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2-VL-7B-Instruct")
File "/usr/local/lib/python3.10/site-packages/transformers/models/auto/auto_factory.py", line 555, in from_pretrained
return model_class.from_pretrained(
File "/usr/local/lib/python3.10/site-packages/transformers/modeling_utils.py", line 3114, in from_pretrained
raise ValueError(
ValueError: Custom modeling code detected in repo Qwen/Qwen2-VL-7B-Instruct. Set `trust_remote_code=True` to use this model. For more information, see https://huggingface.co/docs/transformers/custom_code Why it happens
Qwen2-VL uses custom vision encoding logic (for image patch tokenization) and image processing functions that aren't part of the official transformers library distribution. The transformers library blocks loading custom modeling code by default as a security measure: arbitrary Python code from HuggingFace repos could contain malware. You must explicitly opt-in with trust_remote_code=True to acknowledge you've reviewed the repo and accept the risk.
Detection
Run your model loading code and catch ValueError for 'trust_remote_code' in the message. Enable debug logging via transformers.logging.set_verbosity_debug() to see when custom code paths are invoked during model initialization.
Causes & fixes
AutoModelForCausalLM.from_pretrained() called without trust_remote_code parameter
Add trust_remote_code=True to from_pretrained() call: AutoModelForCausalLM.from_pretrained('Qwen/Qwen2-VL-7B-Instruct', trust_remote_code=True)
AutoProcessor.from_pretrained() also needs trust_remote_code because Qwen2-VL has custom image processing
Add trust_remote_code=True to processor loading: processor = AutoProcessor.from_pretrained('Qwen/Qwen2-VL-7B-Instruct', trust_remote_code=True)
Using quantized versions (GPTQ, AWQ) which may have additional custom kernels
For quantized models, enable both trust_remote_code=True and device_map='auto': model = AutoModelForCausalLM.from_pretrained('Qwen/Qwen2-VL-7B-Instruct-GPTQ', trust_remote_code=True, device_map='auto')
Loading model config separately without trust_remote_code blocks downstream vision processing
Pass trust_remote_code=True when loading config: config = AutoConfig.from_pretrained('Qwen/Qwen2-VL-7B-Instruct', trust_remote_code=True)
Code: broken vs fixed
from transformers import AutoModelForCausalLM, AutoProcessor
import os
model_id = "Qwen/Qwen2-VL-7B-Instruct"
# This line will fail with ValueError about trust_remote_code
model = AutoModelForCausalLM.from_pretrained(
model_id,
device_map="auto",
torch_dtype="auto"
)
processor = AutoProcessor.from_pretrained(model_id)
print("Model loaded successfully")from transformers import AutoModelForCausalLM, AutoProcessor
import os
from PIL import Image
import requests
model_id = "Qwen/Qwen2-VL-7B-Instruct"
# FIX: Add trust_remote_code=True to both model and processor loading
model = AutoModelForCausalLM.from_pretrained(
model_id,
device_map="auto",
torch_dtype="auto",
trust_remote_code=True # Enable custom vision encoding logic
)
processor = AutoProcessor.from_pretrained(
model_id,
trust_remote_code=True # Enable custom image processing
)
print("Model loaded successfully")
# Verify with a simple inference
image_url = "https://upload.wikimedia.org/wikipedia/commons/thumb/3/3a/Cat03.jpg/1200px-Cat03.jpg"
image = Image.open(requests.get(image_url, stream=True).raw)
messages = [
{
"role": "user",
"content": [
{"type": "image"},
{"type": "text", "text": "What is in this image?"}
]
}
]
text = processor.apply_chat_template(
messages, tokenize=False, add_generation_prompt=True
)
inputs = processor(text=[text], images=[image], padding=True, return_tensors="pt")
inputs = inputs.to("cuda")
with torch.no_grad():
output_ids = model.generate(**inputs, max_new_tokens=128)
response = processor.decode(output_ids[0], skip_special_tokens=True)
print(f"Model response: {response}")Workaround
If you cannot update transformers and absolutely must load Qwen2-VL locally, manually download the modeling files from the HuggingFace repo (modeling_qwen2_vl.py, image_processing_qwen2_vl.py), save them in your local directory, and use AutoModelForCausalLM.from_pretrained('./local_qwen_path', local_files_only=True): but this is fragile and not recommended for production.
Prevention
Always audit custom modeling code before enabling trust_remote_code=True: visit the HuggingFace repo, review modeling_*.py files, and confirm they don't contain suspicious code. For production deployments, use official quantized versions (GPTQ, AWQ) or switch to fully open-source multimodal models with standard transformers implementations like LLaVA-1.5 which don't require custom code.