The canonical PyTorch training loop annotated line by line, with train and eval modes, no_grad, optimiser and scheduler order, clipping, autocast, checkpoints, seeds and DataLoader traps. · verified 13 Nov 2026
Broadcasting rules with worked shape tables, einsum patterns for attention, reshape versus permute, batch dimensions, and the shape bugs that hide in attention masks, convolutions and loss functions. · verified 13 Nov 2026
Provider-neutral prompt patterns that hold up in production: roles, ordering, delimiters, few-shot, when chain-of-thought helps, JSON with a schema and validation, tool calling, retries and injection hygiene. · verified 13 Nov 2026
Classification, regression, ranking and generation metrics in four tables with the formula, when to use each and what it hides, plus runnable scikit-learn blocks and bootstrap confidence intervals. · verified 13 Nov 2026
Fifteen ML frameworks and inference engines on one page, with the version at the time of writing, licence, hardware support, what each is good at and what to watch for. · verified 13 Nov 2026
FP32 to INT4, NF4, GGUF k-quants, AWQ and GPTQ on one page, with bytes per parameter, memory for 7 B and 70 B models, reported accuracy impact and hardware support. · verified 13 Nov 2026
A build and review checklist for retrieval-augmented generation covering ingestion, chunking, embeddings, index type, hybrid search with BM25, reranking, prompt assembly with citations, evaluation and monitoring. · verified 13 Nov 2026