tensorplay.stax

Dynamic shapes

Use tensorplay.compile(fn, backend="stax", dynamic=True) to keep input extents symbolic. Shape reads through shape, size() and numel() remain expressions; generated CPU and CUDA kernels bind their extents and inferred strides from the tensors supplied on each call. Pointwise operations, reductions, reshapes and basic indexing can reuse an artifact across compatible sizes. Slices preserve symbolic lengths, strides and offsets, including positive steps, negative bounds and bounds computed from input sizes. Integer indexing and select guard index validity; returned views retain shared storage with their inputs.

import tensorplay as tp

def flatten_rows(x):
    return x.reshape(x.size(0), -1) * 2

compiled = tp.compile(flatten_rows, dynamic=True, strict_native=True)
compiled(tp.randn(3, 7))
compiled(tp.randn(5, 11))

Reuse checks the shape decisions and layout relations made during capture and lowering. A different branch, broadcast pattern or stride arrangement selects another guarded artifact. Sizes zero and one have separate specializations. strict_native=True raises when a region cannot be generated instead of executing its graph as a fallback.

Training preserves saved dimensions as symbolic scalar values for pointwise operations, sum/mean reductions, broadcasting, reshape/expand, squeeze, slice/select, unflatten, diagonal and repeat derivatives. Variance and standard deviation derivatives, including combined mean outputs, compute reduction counts from runtime dimensions. Zero standard deviation masks the incoming variance gradient, and zero repeats restore a zero gradient of the original input shape. Compatible sizes reuse one forward artifact and one backward artifact; each forward invocation saves its own dimensions for the later backward call. Gradient formulas that still save concrete shape metadata, scalar operands or shape-derived operator arguments select guarded specializations when that metadata changes. Variance corrections greater than one also select specializations to preserve decisions about nonpositive degrees of freedom. Backward input dimensions are independent even when their sample values happen to coincide. Symbolic graph capture is available through make_graph(..., tracing_mode="symbolic"); storage-free fake tracing and data-dependent output extents remain unsupported by that capture entry point. Tensor inputs nested in containers and call-outs with concrete output metadata use guarded specializations rather than sharing an artifact across sizes.

Functions