Google Colab not using GPU? Fix slow training

Colab says “GPU” is selected, but your training is still slow or your CPU is at 100% and GPU at 0%. This usually means your code never moved the model or tensors to the GPU. Let’s fix that.

1. Make sure GPU is actually enabled

1. Runtime → Change runtime type → Hardware accelerator: GPU → Save.
2. In a cell, run:

import torch
print(torch.cuda.is_available())

You should see True. For TensorFlow:

import tensorflow as tf
print(tf.config.list_physical_devices("GPU"))

2. Move model and data to GPU (PyTorch)

device = "cuda" if torch.cuda.is_available() else "cpu"

model = MyModel().to(device)

for inputs, labels in train_loader:
    inputs = inputs.to(device)
    labels = labels.to(device)
    outputs = model(inputs)
    loss   = criterion(outputs, labels)
    # backward, step, etc.

If you forget .to(device), your training will silently run on CPU.

3. Move model and data to GPU (TensorFlow / Keras)

import tensorflow as tf

print(tf.config.list_logical_devices("GPU"))

with tf.device("/GPU:0"):
    model = create_model()
    model.compile(optimizer="adam", loss="categorical_crossentropy")

model.fit(train_dataset, epochs=10)

In many simple cases, Keras will automatically use the GPU when available, but putting the model under with tf.device makes intent explicit.

4. Check whether data pipeline is the real bottleneck

• If you stream thousands of tiny files from Drive, I/O can dominate.
• Zip your dataset and extract to /content/ (fast local disk) before training.
• Use prefetching / parallel loading in data pipelines.

5. Snapshot a working GPU setup with NoteCapsule

Once your notebook definitely uses GPU and runs at a reasonable speed, capture a Capsule so you can reuse the same setup (and show your professor/teammates).

from notebookcapsule import create_capsule

create_capsule(
    name="gpu-training-ok",
    notebook_path="notebooks/train_on_gpu.ipynb",
    data_dirs=["./data"],
    base_dir=".",  # project root
)