Use the GPU type names Slurm accepts and keep module PyTorch off RTX Pro 6000...

#!/bin/bash

#SBATCH --account=<your-account>

#SBATCH --partition=all

#SBATCH --gres=gpu:1

#SBATCH --gres=gpu:nvidia_l40:1

module purge

module load 2025/gpu

srun python train.py

```

{{% alert title="Module PyTorch does not run on RTX Pro 6000 GPUs" color="warning" %}}

The `py-torch` modules contain GPU code for the A40 and L40 only. On an RTX Pro 6000 node `torch.cuda.is_available()` still returns `True`, but the first GPU operation fails with `CUDA error: no kernel image is available for execution on the device`. Request an A40 (`gpu:nvidia_a40:1`) or L40 (`gpu:nvidia_l40:1`) for jobs that use these modules.

{{% /alert %}}

{{% alert title="Always use module purge" color="info" %}}

Start batch scripts with `module purge` to ensure a clean environment.

{{% /alert %}}

#SBATCH --ntasks=1

#SBATCH --cpus-per-task=4

#SBATCH --mem=8G

#SBATCH --gres=gpu:1

#SBATCH --gres=gpu:nvidia_l40:1

#SBATCH --output=gpu_%j.out

module purge

To request a specific GPU type:

```bash

#SBATCH --gres=gpu:l40:1      # NVIDIA L40

#SBATCH --gres=gpu:a40:1      # NVIDIA A40

#SBATCH --gres=gpu:nvidia_l40:1            # NVIDIA L40

#SBATCH --gres=gpu:nvidia_a40:1            # NVIDIA A40

#SBATCH --gres=gpu:nvidia_rtx_pro_6000:1   # NVIDIA RTX Pro 6000

```

The type must be written exactly as Slurm names it (`sinfo -o "%n %G"` lists it per node); short forms such as `gpu:l40:1` are rejected.

{{% alert title="Module PyTorch does not run on RTX Pro 6000 GPUs" color="warning" %}}

The `py-torch` modules contain GPU code for the A40 and L40 only. On an RTX Pro 6000 node `torch.cuda.is_available()` still returns `True`, but the first GPU operation fails with `CUDA error: no kernel image is available for execution on the device`. Request an A40 or L40 for jobs that use these modules, as `gpu_job.sh` above does.

{{% /alert %}}

## Next steps

- Use [Containers](/tutorials/apptainer/) for custom GPU environments

### Step 1: Create a Python script

First, create a simple script to run. This one just prints some information. Use the editor you are comfortable with: `nano` is the simplest, and the [Vim tutorial](/tutorials/vim/) covers `vim`.

First, create a simple script to run. This one just prints some information. `vim` is the editor installed on DAIC; if you have not used it before, the [Vim tutorial](/tutorials/vim/) explains how to insert text, save, and quit.

```shell-session

$ cd /tudelft.net/staff-umbrella/<project>

$ nano hello.py

$ vim hello.py

```

```python

Now create the Slurm script that will run your Python code:

```shell-session

$ nano hello_job.sh

$ vim hello_job.sh

```

```bash

```bash

#SBATCH --gres=gpu:1    # One GPU (any type)

#SBATCH --gres=gpu:2    # Two GPUs

#SBATCH --gres=gpu:l40:1   # Specifically an L40 GPU

#SBATCH --gres=gpu:a40:2   # Two A40 GPUs

#SBATCH --gres=gpu:nvidia_l40:1            # Specifically an L40 GPU

#SBATCH --gres=gpu:nvidia_a40:2            # Two A40 GPUs

#SBATCH --gres=gpu:nvidia_rtx_pro_6000:1   # An RTX Pro 6000 GPU

```

Available GPU types on DAIC include L40, A40, and RTX Pro 6000. Request specific types only if your code requires it - being flexible gets you through the queue faster.

The type must be written exactly as Slurm names it; a short form such as `gpu:l40:1` is rejected with "Requested node configuration is not available". `sinfo -o "%n %G"` lists the GPU type of every node. Request specific types only if your code requires it - being flexible gets you through the queue faster.

## Running GPU jobs

#SBATCH --ntasks=1

#SBATCH --cpus-per-task=4

#SBATCH --mem=16G

#SBATCH --gres=gpu:1

#SBATCH --gres=gpu:nvidia_l40:1

#SBATCH --output=train_%j.out

# Clean environment and load required modules

The job script above loads all three in one `module load` line. Without `py-torch`, `import torch` fails with `ModuleNotFoundError`. If you manage your own packages instead (for example with `uv`), see [Python environments](/tutorials/python/).

{{% alert title="Module PyTorch does not run on RTX Pro 6000 GPUs" color="warning" %}}

The `py-torch` modules contain GPU code for the A40 and L40 only. On an RTX Pro 6000 node `torch.cuda.is_available()` still returns `True`, but the first GPU operation fails with `CUDA error: no kernel image is available for execution on the device`. This is why the job script requests `gpu:nvidia_l40:1` rather than `gpu:1`. A PyTorch you install yourself from PyPI runs on all DAIC GPU types.

{{% /alert %}}

Why use modules?

- **Version control**: Run `module load python/3.11` today, `python/3.12` tomorrow

Modify the basic job to request a GPU. Add `nvidia-smi` to verify the GPU is available.

{{% alert title="Check your work" color="info" %}}

Your output should include `nvidia-smi` output showing a GPU:

```

+-----------------------------------------------------------------------------+

| NVIDIA-SMI ...    Driver Version: ...    CUDA Version: ...                  |

|-------------------------------+----------------------+----------------------+

| GPU  Name        ...

```

Your output should include the `nvidia-smi` status table, which names the GPU (for example `NVIDIA L40` or `NVIDIA RTX PRO 6000 Blackwell Server Edition`) and shows its memory.

If the command fails or lists no devices, check that you requested a GPU with `--gres=gpu:1`.

{{% /alert %}}

### Request an interactive session

```shell-session

$ salloc --account=<your-account> --partition=all --time=1:00:00 --cpus-per-task=4 --mem=8G --gres=gpu:1

$ salloc --account=<your-account> --partition=all --time=1:00:00 --cpus-per-task=4 --mem=8G --gres=gpu:nvidia_l40:1

salloc: Pending job allocation 12351

salloc: job 12351 queued and waiting for resources

salloc: job 12351 has been allocated resources

$ srun hostname

gpu15.ethernet.tudhpc

$ srun nvidia-smi

+-----------------------------------------------------------------------------+

| NVIDIA-SMI 550.54.15    Driver Version: 550.54.15    CUDA Version: 12.4     |

|-------------------------------+----------------------+----------------------+

| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |

| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |

|===============================+======================+======================|

|   0  NVIDIA L40          On   | 00000000:41:00.0 Off |                    0 |

| N/A   30C    P8    22W / 300W |      0MiB / 46068MiB |      0%      Default |

+-------------------------------+----------------------+----------------------+

$ srun nvidia-smi --query-gpu=name,memory.total,driver_version --format=csv

name, memory.total [MiB], driver_version

NVIDIA L40, 46068 MiB, 610.43.02

```

Plain `srun nvidia-smi` prints the full status table, including current utilisation and memory use.

Modules loaded in this shell are passed on to the commands `srun` starts, so load the software first:

```shell-session

1. Forgot `--gres=gpu:1` in your script

2. Running on login node instead of through `srun`

3. A CPU-only PyTorch build is installed in your environment

4. The PyTorch build needs a newer CUDA version than the node's driver supports (see [CUDA version mismatch](/tutorials/python/#cuda-version-mismatch))

4. The PyTorch build does not support the GPU model or the driver (see [GPU and CUDA compatibility](/tutorials/python/#gpu-and-cuda-compatibility))

## Best practices