[Question]: ImportError: cannot import name 'Qwen2_5_VLConfig' from 'transformers'

[chenchen772]

Question

Thank you for providing such an excellent project！
However, I encountered some errors during the operation. I downloaded the relevant environment based on your tutorial and ran the code according to the process:

1. conda activate internnav
   python scripts/eval/start_server.py --config scripts/eval/configs/h1_internvla_n1_async_cfg.py

Then, an error occurred when running the following code:

2. conda activate internutopia
   MESA_GL_VERSION_OVERRIDE=4.6 python scripts/eval/eval.py --config scripts/eval/configs/h1_internvla_n1_

Because the code depends on transformers=4.26.1. Even if I update transformers=4.57.3, other errors will still be reported:

ImportError: cannot import name 'TransformGetItemToIndex' from 'torch._higher_order_ops.flex_attention'

So I updated torch again, but still couldn't solve it. Looking forward to your reply!

[kew6688]

Hi! Thanks for trying the project and for the detailed report.

From your logs, the errors are caused by dependency version mismatch between the old pinned environment (e.g., transformers==4.26.1) and newer model code / newer PyTorch features.

I couldn't replicate your error on my side, please share mroe details about your installations. You can double check with these steps:

1. Please pull the latest code
2. Re-install InternNav with the full extras (model + isaac)

conda activate internutopia   # or the env you use to run eval.py
pip install -U pip setuptools wheel
pip install -e ".[model,isaac]"

This should install a compatible transformers (4.51.0) required by the latest evaluator/model integration.

[chenchen772]

Hello, I updated my environment according to your instructions, but the code still reports an error. What I want to clarify is whether in your tutorial, eval.py runs in the virtual environment of internnav or in the internutopia environment. I had run it on a Windows computer before and could run it successfully by following the tutorial. So this time, I got all kinds of errors on ubuntu, which is quite puzzling. Also, when creating the internutopia environment, I need to cd to the isaac sim folder and conda activate internutopia in order to successfully join the path and run python -m internutopia.demo.h1_locomotion, so python actually requires isaac sim's python.sh environment, right? Finally, what I want to know is which virtual environment the operation of python scripts/eval/eval.py --config scripts/eval/configs/ h1_internvla_n1_async_cff.py ultimately depends on.
I have quite a few questions. I'm really, really looking forward to and thank you for your reply!

Here are my steps：

1. conda create -n navigation python=3.10 libxcb=1.14

2. conda activate navigation
3. pip install internutopia

4. python -m internutopia.setup_conda_pypi

5. conda deactivate && conda activate navigation

6. pip install torch==2.8.0 torchvision==0.23.0 torchaudio==2.8.0 --index-url https://download.pytorch.org/whl/cu129

8. cd /home/xxx/Code/InternNav
9. pip install -e ".[model, isaac]"

10. python scripts/eval/eval.py --config scripts/eval/configs/h1_internvla_n1_async_cfg.py
bug

11. solve

12. python scripts/eval/eval.py --config scripts/eval/configs/h1_internvla_n1_async_cfg.py
Then an error was reported. According to your tutorial, I only need to download my cuda version of torch. But why is cuda11.* still the default here? Does this project have to run under cuda version 11.8 or 12.4, so I need to re-download the cuda version?

[kew6688]

Hi, thanks for providing detailed steps!

The conda env used to start is internutopia, which is installed with flag [isaac]. If you are running the model with the same env, install it with flag [model] as well.

If you find your model depency is conflict with internutopia, you can also use the InternNav model server feature by set use_agent_server=True. Start the server in a seperate env by

# from one process
conda activate <model_env>
python scripts/eval/start_server.py --config scripts/eval/configs/h1_internvla_n1_async_cfg.py

# from another process
conda activate <internutopia>
python scripts/eval/eval.py --config scripts/eval/configs/h1_internvla_n1_async_cfg.py

And yes, it needs isaac sim python to start the simulation app.

For this bug, Root cause

The issue is caused by a CUDA binary mismatch introduced during environment setup.

In step 6, the environment installs PyTorch with CUDA 12.9 (cu12x). However, the FlashAttention wheel used later is prebuilt against CUDA 11.8 (cu118). This creates a hard runtime dependency on libcudart.so.11.*, which is not available in a CUDA-12 PyTorch environment.

I reproduce this error by following your steps, and it is solved by reinstalled torch with cu118
pip install torch==2.5.1 torchvision==0.20.1 torchaudio==2.5.1 --index-url https://download.pytorch.org/whl/cu118

The pre build flash attention wheel I installed is
pip install flash_attn-2.7.2.post1+cu11torch2.5cxx11abiFALSE-cp310-cp310-linux_x86_64.whl