Using the GPU

When you create a workspace you choose how it accesses the GPU. The available modes depend on what your admin has enabled on the server.

GPU modes

None (CPU only)

The workspace gets no GPU access. Suitable for data preparation, CPU-only inference, or when all GPUs are in use and you want to queue work.

Whole GPU

Your workspace gets exclusive use of one or more entire physical GPUs - all of their memory and compute. No other workspace shares those GPUs while yours is running.

Set the GPU count with the number input. The form shows how many whole GPUs are free right now and your policy cap. If the count you request exceeds what’s free, you can still create the workspace; LabPod will grant the GPUs when you click Start and they are available.

Shared GPU (fractional)

One physical GPU is shared among multiple workspaces simultaneously. Your workspace is given a hard VRAM limit and a soft compute time share. Work-conserving: if the GPU is otherwise idle, your workspace can use more compute than its share; the limit is only enforced under contention.

When configuring, you set:

• GPU memory limit - hard cap in MB on how much VRAM your processes can allocate.
• Compute limit (%) - your share of the GPU’s compute time, up to your policy maximum.

Only available if your admin has enabled this mode on the server. The workspace detail page shows a VRAM bar so you can see how close you are to the limit.

Not the same as a whole GPU

With a shared GPU your workspace will get a CUDA OOM error if a model exceeds the memory cap, even if other processes aren’t using the GPU. Plan your batch sizes and model sizes accordingly.

GPU slice (MIG)

A hardware-partitioned slice of an A100 or H100 GPU with its own dedicated memory and compute engines - physically isolated, not just a software limit. Your workspace gets exclusive use of that slice.

Slices are pre-configured by your admin. Each slice has a profile name like 1g.10gb (1 GPU instance, 10 GB memory). The form shows which slice sizes have free slots. This mode is only available if the server has MIG-capable GPUs and your admin has provisioned slices.

Choosing a mode: quick guide

The workspace creation form offers a Light / Medium / Heavy picker that selects the best available mode automatically:

Use case	What it picks
Light - notebook, small model, quick experiments	Smallest available GPU slice or smallest shared GPU fraction
Medium - fine-tuning, mid-size inference	Larger shared GPU fraction or a whole GPU if no sharing is available
Heavy - training a large model, long-running jobs	A whole GPU

For precise control, expand Advanced: pick exact GPU mode inside the form.

Checking your GPU inside the workspace

nvidia-smi          # see the GPU, VRAM, running processes
python -c "import torch; print(torch.cuda.is_available())"

With a shared GPU, nvidia-smi reports the full card but your processes are subject to the VRAM cap set at workspace creation. You can view current VRAM usage on the workspace detail page.

/dev/shm sizing for DataLoader and DDP

LabPod sets the container’s /dev/shm to half the workspace memory limit (rather than Podman’s default of 64 MB). This means:

• torch.utils.data.DataLoader(num_workers > 0) works without Bus error.
• Single-container multi-GPU DDP training over NCCL’s shared-memory transport works without extra flags.

If you started a container yourself

If you use podman run outside LabPod and hit a Bus error (core dumped) in a DataLoader worker, add --shm-size=4g (or an appropriate fraction of your memory) to the run command. LabPod handles this automatically for workspaces it manages.