build-and-dependency

To set up a reproducible Megatron-LM environment, use the CI container image that already includes the correct CUDA toolkit, PyTorch build, and pre-compiled native extensions, ensuring consistent GPU-ready dependency resolution across machines.

To resolve uv.lock merge conflicts safely, avoid manual editing and use the container-based workflow to run `uv lock` after changes to pyproject.toml, regenerating the locked resolution without causing local and CI environment mismatches.

Host installation of GPU-dependent CUDA, NCCL, and PyTorch packages is fragile and varies across machines, so you must run all uv operations inside the container’s /opt/venv environment using the pinned CI container build stages to prevent setup failures.

The best way to add a dependency is updating pyproject.toml and running `uv lock` inside the container to regenerate uv.lock, followed by `uv sync` with the correct dependency groups to maintain a safe dependency change workflow.

Yes, uv-driven dependency management supports installing and updating dependencies via `uv sync` with locked resolution across multiple dependency groups, ensuring consistent environments inside the container's /opt/venv.

To fix common container or uv errors during installation and module import, use the pinned CI container build stages and execute all uv operations inside the container’s /opt/venv environment to ensure native extensions match the expected CUDA toolkit.