add Kerberos docs and fix multi-GPU examples for daic-new

On first login, symlinks are created in your home directory pointing to TU Delft network storage:

### Kerberos authentication

TU Delft network storage requires a valid Kerberos ticket. Without it, you will get "Permission denied" or "Stale file handle" errors when accessing `linuxhome`, `windowshome`, project, group, or bulk storage.

**When logging in via SSH with a password**, a Kerberos ticket is created automatically.

**When logging in via SSH with a public key** or through OpenOndemand, you must manually obtain a ticket:

```bash

kinit

```

Enter your NetID password when prompted.

Check your current ticket status:

```bash

klist

```

Example output with a valid ticket:

```

Ticket cache: KCM:656519

Default principal: <NetID>@TUDELFT.NET

Valid starting     Expires            Service principal

03/23/26 11:05:12  03/23/26 21:05:12  krbtgt/TUDELFT.NET@TUDELFT.NET

        renew until 03/30/26 12:05:03

```

{{% alert title="First access delay" color="info" %}}

Network storage locations may take up to 30 seconds to mount on first access. If you get a "Stale file handle" error, wait a moment and try again.

{{% /alert %}}

- `~/linuxhome` - Your Linux home on TU Delft storage

- `~/windowshome` - Your Windows home on TU Delft storage

You should connect without entering a password.

## Kerberos for network storage

{{% alert title="Important" color="warning" %}}

SSH key login does not create a Kerberos ticket. Without a ticket, you cannot access network storage (`~/linuxhome`, project storage, etc.).

{{% /alert %}}

After connecting with SSH keys, run:

```shell-session

$ kinit

Password for <YourNetID>@TUDELFT.NET:

```

Verify your ticket:

```shell-session

$ klist

Ticket cache: KCM:656519

Default principal: <YourNetID>@TUDELFT.NET

Valid starting     Expires            Service principal

03/23/26 11:05:12  03/23/26 21:05:12  krbtgt/TUDELFT.NET@TUDELFT.NET

```

{{% alert title="First access delay" color="info" %}}

Network storage may take up to 30 seconds to mount on first access. If you see "Stale file handle", wait and retry.

{{% /alert %}}

## SSH config shortcut

Add to `~/.ssh/config` on your local machine:

┌─────────────────────────────────────────────────────────────────────────┐

│                           STORAGE                                        │

│    /home/<netid>              - 5 MB, config only                       │

│    ~/linuxhome                - 8 GB, personal files                    │

│    ~/linuxhome                - ~25 GB, personal files                  │

│    /tudelft.net/staff-umbrella/<project>  - Project data                │

│    /tudelft.net/staff-bulk/<project>      - Large datasets              │

└─────────────────────────────────────────────────────────────────────────┘

**I just got access to DAIC**

→ Start with [Bash Basics](/tutorials/bash/), then [Slurm Basics](/tutorials/slurm/)

{{% alert title="Using SSH keys?" color="info" %}}

If you log in with SSH keys instead of a password, run `kinit` after connecting to access network storage (`linuxhome`, project storage). See [Storage](/docs/storage/) for details.

{{% /alert %}}

**I know Linux but not clusters**

→ Start with [Slurm Basics](/tutorials/slurm/)

#!/bin/bash

#SBATCH --job-name=accelerate-multi-gpu

#SBATCH --account=<your-account>

# SBATCH --account=<your-account>  # Uncomment and set if required

#SBATCH --partition=all

#SBATCH --time=00:30:00

#SBATCH --nodes=1

module purge

module load 2025/gpu cuda/12.9

# Navigate to script directory

cd "$(dirname "$0")"

# Set up UV paths (adjust if UV is installed elsewhere)

export PATH="$HOME/linuxhome/.local/bin:$PATH"

export UV_CACHE_DIR="$HOME/linuxhome/.cache/uv"

export UV_PYTHON_INSTALL_DIR="$HOME/linuxhome/.local/share/uv/python"

# Navigate to script directory (SLURM_SUBMIT_DIR is where sbatch was called)

cd "$SLURM_SUBMIT_DIR"

# Print job info

echo "========================================"

    uv sync

fi

# Run training with accelerate

srun uv run accelerate launch \

# Pre-download data to avoid race conditions

uv run python -c "from torchvision import datasets; datasets.MNIST('${TMPDIR:-/tmp}/mnist', train=True, download=True); datasets.MNIST('${TMPDIR:-/tmp}/mnist', train=False, download=True)"

# Run training with accelerate (handles multi-GPU spawning)

uv run accelerate launch \

    --num_processes=$NUM_GPUS \

    --mixed_precision=fp16 \

    train.py \

name = "accelerate-multi-gpu"

version = "0.1.0"

description = "Multi-GPU training example with Hugging Face Accelerate"

requires-python = ">=3.10"

requires-python = ">=3.10,<3.13"

dependencies = [

    "torch>=2.0.0",

    "torchvision>=0.15.0",

]

[tool.uv]

index-url = "https://download.pytorch.org/whl/cu124"

extra-index-url = ["https://download.pytorch.org/whl/cu124"]