Build And Dependency

Build and Dependency

Two core principles: build and develop inside containers, and always use uv.

Why Containers

Megatron Bridge depends on CUDA, NCCL, PyTorch with GPU support, Transformer Engine, and optional components like TRT-LLM, vLLM, and DeepEP. Installing these on a bare host is fragile and hard to reproduce. The project ships production-quality Dockerfiles that pin every dependency.

Use the container as your development environment. This guarantees:

• Identical CUDA / NCCL / cuDNN versions across developers and CI.
• uv.lock resolves the same way locally and in CI (the lockfile is Linux-only; it cannot be regenerated on macOS).
• GPU-dependent operations work out of the box.

Container Options

Option 1: NeMo Framework Container (fastest)

Find available tags at https://catalog.ngc.nvidia.com/orgs/nvidia/containers/nemo/tags

skopeo list-tags docker://nvcr.io/nvidia/nemo \
  | python3 -c "import sys,json,re; tags=json.load(sys.stdin)['Tags']; [print(t) for t in sorted((t for t in tags if re.match(r'^\d{2}\.\d{2}', t)), reverse=True)]"

docker run --rm -it --gpus all --shm-size=24g \
  nvcr.io/nvidia/nemo:<tag> \
  bash

Option 2: Build the Megatron Bridge Container

See @docker/README.md for build commands, build arguments, and the full NeMo-FW image stack.

Running the Container

docker run --rm -it -w /opt/Megatron-Bridge \
  -v $(pwd):/opt/Megatron-Bridge \
  -v $HOME/.cache/uv:/root/.cache/uv \
  --gpus all \
  --shm-size=24g \
  --ulimit memlock=-1 \
  --ulimit stack=67108864 \
  megatron-bridge:latest \
  bash

Mounting $HOME/.cache/uv avoids re-downloading wheels on every run.

Containers on Slurm

On Slurm clusters with Enroot/Pyxis, pass containers directly to srun:

srun --mpi=pmix \
  --container-image="$CONTAINER_IMAGE" \
  --container-mounts="$CONTAINER_MOUNTS" \
  --no-container-mount-home \
  bash -c "cd /opt/Megatron-Bridge && uv run --no-sync python ..."

If you bind-mount a custom source tree into the container, only rank 0 should sync while others wait:

if [ "$SLURM_LOCALID" -eq 0 ]; then uv sync; else sleep 10; fi

Note: --no-container-mount-home is an srun flag, not an #SBATCH directive. Set UV_CACHE_DIR to shared storage to avoid filling /root/.cache/.

Always Use uv

Never use pip install, conda, or bare python — always go through uv. All uv commands must be run inside a container. Never install or upgrade dependencies outside the CI container.

Essential Commands

Adding Dependencies

Submit dependency changes as a separate PR before the feature PR:

# Optional dependency (preferred)
uv add --optional --extra <group> <package>

# Required dependency (needs strong justification — affects all downstream)
uv add <package>

Commit both modified files:

git add pyproject.toml uv.lock
git commit -s -m "build: add <package>"

Regenerating uv.lock

The lockfile is Linux-only (resolves CUDA wheels). Run inside Docker:

docker run --gpus all --rm \
  -v $(pwd):/opt/Megatron-Bridge \
  megatron-bridge:latest \
  bash -c 'cd /opt/Megatron-Bridge && uv lock'

Switching MCore Branches

# Switch to dev branch
./scripts/switch_mcore.sh dev
uv sync              # without --locked

# Switch back to main
./scripts/switch_mcore.sh main
uv sync --locked     # lockfile matches again

Quick Start

# 1. Clone and init submodules
git clone https://github.com/NVIDIA-NeMo/Megatron-Bridge megatron-bridge
cd megatron-bridge
git submodule update --init 3rdparty/Megatron-LM

# 2. Build the container
docker build -f docker/Dockerfile.ci --target megatron_bridge -t megatron-bridge:latest .

# 3. Start a dev shell
docker run --rm -it -v $(pwd):/opt/Megatron-Bridge --gpus all --shm-size=24g megatron-bridge:latest bash

# 4. Install pre-commit hooks (inside container)
uv run --group dev pre-commit install

# 5. Sanity check
uv run python -m torch.distributed.run --nproc_per_node=1 \
  scripts/training/run_recipe.py \
  --recipe vanilla_gpt_pretrain_config \
  train.train_iters=5 train.global_batch_size=8 train.micro_batch_size=4 \
  scheduler.lr_warmup_iters=1 scheduler.lr_decay_iters=5 \
  logger.log_interval=1

Common Pitfalls