Radiologists labelling 3D CT scans with MONAILabel wait minutes for the model's suggestion when it runs on a CPU.

A small port to AMD Instinct GPUs that changes only three files and runs everything through PyTorch, so existing models work unchanged and future updates keep working.

In PyTorch, one operation can turn into many small GPU tasks, and the first training steps include one-off compile costs that skew measurements.

A step-by-step profiling guide on a small protein-folding model, comparing PyTorch fusion, torch.compile and custom Triton kernels.

People talk about 'the AI product manager' as one job, but the work differs for research, models and user-facing features.

An account from three PMs on how quality, evaluation and shipping mean different things at each layer, and how ideas move from research to product.