• Upload the starter notebook and the .py files from its folder to Colab, then choose Runtime → Change runtime type → T4 GPU. A full run takes about as long as on an M1 Max.
• First cell: !pip install d2l==1.0.3 --no-deps gensim (torch, transformers and tokenizers are preinstalled). Data downloads automatically, as on a laptop.
• Sessions reset after a while, and the files with them. Mount Google Drive and work in a folder there, so the saved runs (hw6_runs/) and HW6's results_hw6.json / hw6_seq2seq.pt survive.

• You need an up-to-date NVIDIA driver (nvidia-smi shows it). You do not need to install CUDA separately; PyTorch includes it.
• Linux: the plain pip install torch torchvision above already has GPU support.
• Windows: plain pip install torch is CPU-only. Instead, use the command from the PyTorch install page (choose Pip / Windows / CUDA). It looks like pip install torch torchvision --index-url https://download.pytorch.org/whl/cu128.
• Check: python -c "import torch; print(torch.cuda.is_available())" prints True.

• The Mac GPU supports float32 only. When you turn numpy arrays (float64 by default) into tensors, use torch.tensor(X, dtype=torch.float32). Otherwise you get Cannot convert a MPS Tensor to float64.
• If an operation reports not currently implemented for the MPS device, start Jupyter with the environment variable PYTORCH_ENABLE_MPS_FALLBACK=1 set, so that operation runs on the CPU.
• Very small models can be faster on the CPU. That's normal; the starters choose accordingly.
• The d2l book's own GPU helpers only know NVIDIA GPUs. The starters' first cell patches them, so d2l.Trainer(..., num_gpus=1) uses the Mac GPU.

The starters use d2l's Module / Trainer classes, which match the book. You may use a plain PyTorch training loop instead (the most common style in practice), as long as you evaluate exactly as the starter does. This is what d2l.Trainer does for you:

model = model.to(device)
opt = torch.optim.SGD(model.parameters(), lr=0.1)
loss_fn = torch.nn.CrossEntropyLoss()
for epoch in range(num_epochs):
    model.train()                                  # dropout / batch norm in training mode
    for X, y in train_loader:
        X, y = X.to(device), y.to(device)
        loss = loss_fn(model(X), y)
        opt.zero_grad(); loss.backward(); opt.step()
    model.eval()                                   # evaluation mode
    with torch.no_grad():
        val_acc = sum((model(X.to(device)).argmax(1) == y.to(device)).float().mean().item()
                      for X, y in val_loader) / len(val_loader)
    print(f'epoch {epoch}: val acc {val_acc:.3f}')

Apart from HW7's Hugging Face packages (used as the reference to compare against), no other frameworks (Lightning, Keras, JAX) are needed.

• conda works too: conda create -n cs6140 python=3.13, conda activate cs6140, then the same pip install lines. Install torch with pip, not conda.
• If Jupyter can't find torch although you installed it, it's using a different Python. Start jupyter lab from the activated environment, or register it once as a kernel: pip install ipykernel then python -m ipykernel install --user --name cs6140, and pick "cs6140" in Jupyter.