Runtime Error: torchvision::nms Operator Doesn't Exist

[prestonzen]

Runtime Error: torchvision::nms Operator Doesn't Exist

System Configuration

• Platform: Runpod Container (Ubuntu Jammy 22.04)
• Python Version: 3.10.9
• CUDA Version: 12.4
• GPU: 2 x RTX A6000
• Installation Command Used:

  conda create -n HunyuanVideo python==3.10.9
  conda activate HunyuanVideo
  conda install pytorch==2.4.0 torchvision==0.19.0 torchaudio==2.4.0 pytorch-cuda=12.4 -c pytorch -c nvidia

Error Message

After installing torchvision, I get the following error when trying to run sample_video.py:

RuntimeError: Didn't find operator for torchvision::nms. This could be because the operator doesn't exist or was omitted when the model was exported.

Steps to Reproduce

1. Followed all steps from the [HunyuanVideo Installation Guide for Linux](https://github.com/Tencent/HunyuanVideo?tab=readme-ov-file#installation-guide-for-linux).
2. Installed all models listed in the [HunyuanVideo Model Checkpoints](https://github.com/Tencent/HunyuanVideo/blob/main/ckpts/README.md).
3. Tried running the following command:

   python3 sample_video.py \
       --video-size 720 1280 \
       --video-length 129 \
       --infer-steps 50 \
       --prompt "A cat walks on the grass, realistic style." \
       --flow-reverse \
       --use-cpu-offload \
       --save-path ./results

4. Encountered the runtime error.

Debugging Attempts

• Verified torchvision installation with:

  python -c "import torchvision; print(torchvision.__version__)"

  Output: 0.19.0
• Checked available torchvision.ops:

  import torchvision.ops
  print(dir(torchvision.ops))  # `nms` is missing from the list

• Reinstalled torchvision:

  pip uninstall torchvision
  pip install torchvision --no-cache-dir --force-reinstall

• Ensured CUDA is available:

  import torch
  print(torch.cuda.is_available())  # Returns: True

• Tried installing PyTorch and torchvision with:

  pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121

Expected Behavior

The torchvision.ops.nms function should work correctly without missing operator errors, allowing sample_video.py to run successfully.

Questions

1. Is torchvision==0.19.0 fully compatible with pytorch==2.4.0 on CUDA 12.4?
2. Should torchvision be compiled from source to fix this?
3. Is this a known issue with nms being omitted in recent builds?
4. Does HunyuanVideo require a specific torchvision version for proper functionality?

Any insights or solutions would be greatly appreciated. Thanks!

[prestonzen]

It looks like the requirements.txt file didn't specify a CUDA version, leading to the wrong CUDA version being installed. By default, CUDA 12.8 was installed, but the correct version for compatibility with the installed PyTorch version is CUDA 11.8.

[prestonzen]

Seems like a pip package dependency overrode the pytorch version

[qihanghou726]

Same problem

[mengsiwei]

After "pip install --upgrade torchvision", problem solved.