Element-wise absolute value: |x|
Element-wise clamp: min(max(x, minVal), maxVal)
Element-wise exponential: e^x
Element-wise exp(x) - 1, numerically stable for small x.
Gather rows from input along dim using integer indices. Primary use: embedding lookup (dim=0, indices=token IDs).
Element-wise greater than or equal: x >= value → 1.0, else 0.0
Select elements from input along dim at the given indices. Similar to gather but for general index selection patterns.
Element-wise less than: x < value → 1.0, else 0.0
Matrix product of two tensors. For 2-D inputs this is the standard matrix multiplication C = A · B; for higher-rank inputs the leading dimensions are broadcast and the last two are contracted.
C = A · B
Extract a sub-tensor along dimension dim starting at start with the given length.
Zero-pad a 4D tensor N, C, H, W on the spatial dimensions.
Scaled dot-product attention.
Element-wise sign: -1 for negative, 0 for zero, +1 for positive. Non-differentiable.
Extract sliding windows of size size along dimension dim with stride step. Result has one extra dimension appended containing the window elements.