Tensor Almanachttps://tensoralmanac.comUnderstand the machine.When NumPy broadcasting gives you the wrong losshttps://tensoralmanac.com/guides/numpy-broadcasting-shape-bugs/https://tensoralmanac.com/guides/numpy-broadcasting-shape-bugs/A shape-compatible operation can still compare the wrong values. Diagnose a silent pairwise-error bug before it reaches your metrics.Split first, fit second: a preprocessing leakage checkhttps://tensoralmanac.com/guides/prevent-preprocessing-leakage/https://tensoralmanac.com/guides/prevent-preprocessing-leakage/Use a small scaler diagnostic to see whether held-out data is influencing training, then put the boundary inside a scikit-learn pipeline.Estimate KV-cache memory before choosing a context lengthhttps://tensoralmanac.com/guides/estimate-kv-cache-memory/https://tensoralmanac.com/guides/estimate-kv-cache-memory/Calculate the dense cache payload from layers, KV heads, token count and dtype, and see why the answer is not a GPU-fit guarantee.