Discovered during self-quantization: The model is mistakenly registered as Qwen3, but it is actually based on Qwen2.5

[dayuan12138]

I encountered a critical compatibility issue while attempting to quantize this model locally.
I found that the model architecture was accidentally registered as Qwen3 inside the configuration file. However, this multimodal model is essentially built on Qwen2.5.
Furthermore, the Transformers version recommended by this project does not yet support the Qwen3 architecture. This mismatch directly prevents model loading and successful INT8 quantization, and always triggers runtime errors.

[exiawsh]

Hi @dayuan12138 ,

Thanks for catching and reporting this! Could you please point me to the specific file and line of code (e.g., config.json or a specific script) where the model architecture is mistakenly registered as Qwen3?

I'd be happy to take a look and submit a PR to fix this.

[dayuan12138]

Hi @dayuan12138 ,

Thanks for catching and reporting this! Could you please point me to the specific file and line of code (e.g., config.json or a specific script) where the model architecture is mistakenly registered as Qwen3?

I'd be happy to take a look and submit a PR to fix this.

transformers version: 4.44.2
Additional note: I previously got the project’s recommended transformers version wrong and thought it was 4.57.1. This error can be reproduced on version 4.44.2.

Minimal reproduction code：
from transformers import AutoConfig
model_path = "./LocateAnything-3B"
config = AutoConfig.from_pretrained(model_path, trust_remote_code=True)

Error message：
File "C:\Users\Administrator.cache\huggingface\modules\transformers_modules\LocateAnything-3B\configuration_locateanything.py", line 10, in
from transformers.models.qwen3.configuration_qwen3 import Qwen3Config
ModuleNotFoundError: No module named 'transformers.models.qwen3'

[dayuan12138]

Hi @dayuan12138 ,

Thanks for catching and reporting this! Could you please point me to the specific file and line of code (e.g., config.json or a specific script) where the model architecture is mistakenly registered as Qwen3?

I'd be happy to take a look and submit a PR to fix this.

My installed transformers version is outdated and lacks built-in Qwen3 configuration definitions. Meanwhile I have a question: the official technical paper states LocateAnything is built on Qwen2.5-3B-Instruct, so why does configuration_locateanything.py import Qwen3Config? Does this mean the dev team plans to release an updated Qwen3-based version in the future?

[exiawsh]

@dayuan12138 We don't actually have a dependency on Qwen3, so feel free to modify the code according to your needs. We did spend some time previously trying to build a better version using Qwen3VL, but it hasn't worked out yet.

[dayuan12138]

@dayuan12138 We don't actually have a dependency on Qwen3, so feel free to modify the code according to your needs. We did spend some time previously trying to build a better version using Qwen3VL, but it hasn't worked out yet.实际上，我们并不依赖 Qwen3。因此，您可以根据自己的需求随意修改代码。之前，我们确实尝试过使用 Qwen3VL 来打造一个更优秀的版本，但至今仍未成功。

Importing Qwen3Config only causes an error under older transformers versions. Many thanks to the LocateAnything development team for your diligent work on this vision model; the model performs excellently. I hope an official quantized version can be released later for my deployment on the Jetson Orin Nano development board.