GPU support in containers in MLOps - Step-by-Step Execution

Process Flow - GPU support in containers

Start Container

↓

Check GPU Availability

Yes↓

Use NVIDIA Container Toolkit

↓

Mount GPU Drivers & Libraries

↓

Run Container with GPU Access

↓

Container Uses GPU for Tasks

↓

Stop Container

This flow shows how a container is started with GPU support by checking GPU availability, using NVIDIA toolkit to mount drivers, and enabling GPU access inside the container.

Execution Sample

MLOps

docker run --gpus all nvidia/cuda:12.0-base nvidia-smi

Runs a container with full GPU access and executes 'nvidia-smi' to show GPU status inside the container.

Process Table

Step	Action	Command/Check	Result/Output
1	Start container with GPU flag	docker run --gpus all nvidia/cuda:12.0-base nvidia-smi	Container starts with GPU access enabled
2	Container checks for GPU devices	nvidia-smi	Lists GPU devices and their status
3	Container runs GPU-enabled task	Any CUDA program inside container	Task uses GPU for computation
4	Stop container	docker stop <container_id>	Container stops, GPU resources freed

💡 Container stops after GPU tasks complete or user stops it manually

Status Tracker

Variable	Start	After Step 1	After Step 2	After Step 3	Final
Container State	Not running	Running with GPU access	Running with GPU detected	Running with GPU task executing	Stopped
GPU Access	Unavailable	Enabled via --gpus flag	Confirmed by nvidia-smi output	Used by CUDA tasks	Released

Key Moments - 3 Insights

Why do we need the '--gpus all' flag when running the container?

The '--gpus all' flag tells Docker to provide the container access to all GPUs on the host. Without it, the container cannot see or use the GPU devices, as shown in step 1 of the execution table.

What role does 'nvidia-smi' play inside the container?

'nvidia-smi' is a tool that lists GPU devices and their status. Running it inside the container (step 2) confirms that the container has access to the GPUs.

How does the container use GPU drivers and libraries from the host?

The NVIDIA Container Toolkit mounts the necessary GPU drivers and libraries from the host into the container automatically when using the '--gpus' flag, enabling GPU tasks inside the container as seen in step 3.

Visual Quiz - 3 Questions

Look at the execution table, what command confirms GPU availability inside the container?

A nvidia-smi

B docker stop

C docker run --gpus all

D Any CUDA program

💡 Hint

Check step 2 in the execution table where GPU devices are listed.

At which step does the container start using the GPU for computation?

A Step 1

B Step 3

C Step 2

D Step 4

💡 Hint

Refer to step 3 where GPU-enabled tasks run inside the container.

If the '--gpus all' flag is omitted, what will happen according to the variable tracker?

A Container State becomes 'Running with GPU access'

B GPU Access becomes 'Used by CUDA tasks'

C GPU Access remains 'Unavailable'

D Container stops immediately

💡 Hint

Look at 'GPU Access' variable changes in the variable tracker after step 1.

Concept Snapshot

GPU support in containers:
- Use 'docker run --gpus all' to enable GPU access
- NVIDIA Container Toolkit mounts drivers inside container
- Run 'nvidia-smi' inside container to verify GPU
- GPU tasks run inside container using host GPU
- Stop container to release GPU resources

Full Transcript

To use GPU inside containers, start by running the container with the '--gpus all' flag. This flag enables GPU access by mounting necessary drivers and libraries using the NVIDIA Container Toolkit. Inside the container, running 'nvidia-smi' confirms GPU availability by listing GPU devices. Then, GPU-enabled programs can run inside the container using the host's GPU. Finally, stopping the container frees GPU resources. This process ensures containers can leverage GPUs for machine learning or other compute tasks.