Question: Are there officially supported prebuilt Docker images for VeOmni GPU training?

[Eisenhower]

Hi VeOmni team,

I am trying to set up a VeOmni training environment, but building the Docker image locally is taking a very long time.

I found that the GitHub GPU CI workflow uses this image:

verl-ci-cn-beijing.cr.volces.com/veomni/open_veomni-gpu:pytorch_25_03_py3_cuda12_9

I also found public Ascend/NPU images such as:

ascendai/veomni:veomni-8.3.rc2-910b-ubuntu22.04-py3.11-v0.1.9a2
quay.io/ascend/veomni:veomni-8.3.rc2-910b-ubuntu22.04-py3.11-v0.1.9a2
ascendai/veomni:veomni-8.3.rc2-a3-ubuntu22.04-py3.11-latest

Could you please confirm:

1. Is verl-ci-cn-beijing.cr.volces.com/veomni/open_veomni-gpu:pytorch_25_03_py3_cuda12_9 intended to be publicly usable for GPU training, or is it only for internal/CI usage?
2. Is there an officially recommended prebuilt GPU Docker image for users who do not want to build from source?
3. Which image should be used for current main branch training on NVIDIA GPUs?
4. Does the recommended image already include compatible CUDA, PyTorch, flash-attn, uv, and VeOmni dependencies?
5. For transformers v5 development, should users still run this after entering the container?

uv sync --frozen --no-group transformers-stable --extra transformers5-exp --extra gpu --extra audio --group dev

It would be very helpful if the README or Docker documentation could list the recommended prebuilt images and their supported hardware/runtime versions.

Thanks!

[Luosuu]

1. It is just used for CI but I think it would be fine to use it for training since veomni does not really have many complicated dependencies.
2. Sorry but we do not have a CI to build nightly images.
3. please check the cu129 docker file
4. In the dockerfile we install everything
5. Yes because v4 is still the default