http://programmer.ie/tags/torch.compile/ Recent content in Torch.compile on Programmer.ie: Modern AI programming Hugo en-us Sat, 08 Aug 2026 13:56:00 +0100 http://programmer.ie/post/pytorch-zero-to-hero-09/ Sat, 08 Aug 2026 13:56:00 +0100 http://programmer.ie/post/pytorch-zero-to-hero-09/ <h2 id="pytorch-zero-to-hero--step-09">PyTorch: Zero to Hero — Step 09</h2> <p>At this point in the series, the model runs.</p> <p>That does not mean it runs well.</p> <p>A training loop can be correct and still waste most of the machine.</p> <p>A model can fit in memory and still spend half its time waiting on synchronization.</p> <p>A <code>torch.compile</code> call can make code faster, slower, or simply move the bottleneck somewhere else.</p> <p>A CUDA out-of-memory error can be caused by the model, the optimizer, activations, fragmentation, a leaked reference, a larger batch, a longer sequence, or an innocent-looking tensor that was kept alive by Python.</p>