The collection time of 'Mjlab-Velocity-Rough-Unitree-G1' task is 20X slower than ' Mjlab-Velocity-Flat-Unitree-G1' in Nvidia 4090？

[VincentFEI]

Hello, dear developer,
I found the collection time of 'Mjlab-Velocity-Rough-Unitree-G1' task is 20X slower than ' Mjlab-Velocity-Flat-Unitree-G1' in Nvidia 4090.
The commands exeucted are as follows:
uv run train Mjlab-Velocity-Rough-Unitree-G1 --env.scene.num-envs 8192 --agent.logger tensorboard --agent.max-iterations 15000
uv run train Mjlab-Velocity-Flat-Unitree-G1 --env.scene.num-envs 8192 --agent.logger tensorboard --agent.max-iterations 15000
The collection time of 'Mjlab-Velocity-Rough-Unitree-G1' is 63.476s, and that of ' Mjlab-Velocity-Flat-Unitree-G1' is 2.19s.
Is there something wrong with the command or environment I'm running?

[VincentFEI]

For your information, I am using Ubuntu 22.04 and CUDA 12.2.
This is the Dockerfile of the imgae I used:

# Refer to uv-docker-example:
# https://github.com/astral-sh/uv-docker-example/blob/main/standalone.Dockerfile
# Note that we use uv to launch, so we omit the second half of the example (non-UV final image)

# Use nvcr.io to avoid Docker Hub rate limit (429) when unauthenticated
# 12.2.0 无 ubuntu24.04 官方 tag，改用 ubuntu22.04；若需 24.04 可用 12.8.0-runtime-ubuntu24.04
FROM nvcr.io/nvidia/cuda:12.2.0-runtime-ubuntu22.04
COPY --from=ghcr.io/astral-sh/uv:latest /uv /uvx /bin/

ENV DEBIAN_FRONTEND=noninteractive
RUN apt-get update && apt-get install -y \
    git \
    curl \
    libegl-dev \
    && rm -rf /var/lib/apt/lists/*

ENV UV_COMPILE_BYTECODE=1
ENV UV_LINK_MODE=copy
ENV UV_PYTHON_PREFERENCE=only-managed

RUN uv python install 3.13

WORKDIR /app

RUN --mount=type=cache,target=/root/.cache/uv \
    --mount=type=bind,source=uv.lock,target=uv.lock \
    --mount=type=bind,source=pyproject.toml,target=pyproject.toml \
    uv sync --locked --no-install-project --no-editable --no-dev

ADD . /app

RUN --mount=type=cache,target=/root/.cache/uv \
    uv sync --locked --no-editable --no-dev

RUN uv pip install pip

ENV MUJOCO_GL=egl
EXPOSE 8080

CMD ["uv", "run", "python", "tests/smoke_test.py"]