ValueHashedMasks

class t3toolbox.backend.common.ValueHashedMasks

Mixin giving uniform-layer mask holders VALUE-based __hash__/__eq__ over mask content.

Why this matters (PERFORMANCE, not cosmetics): a mask holder rides as jax aux_data, so its __hash__/__eq__ are part of the jit compilation cache key. Identity hash/eq (the bare @dataclass(eq=False) default) makes a fresh-but-array-identical holder a NEW key -> jit RECOMPILES. In a manifold-optimization loop the orthogonal frame is rebuilt every iteration (via ut3_orthogonal_representations), producing fresh holders whose rank structure is identical; identity hashing would recompile every step, dwarfing the actual compute – the opposite of the uniform layer’s whole point. Value-based hash/eq makes the cache key reflect the rank STRUCTURE: identical structure -> cache hit (no recompile); genuinely different structure -> recompile (correct). See docs/contributor/uniform_pytree_composition.md.

Subclasses must be @dataclass(frozen=True, eq=False) and expose .data as a tuple of the (HOST numpy, concrete) mask arrays. The content hash is cached (the holder is frozen/immutable); eq short-circuits on identity, then compares by np.array_equal.

Methods

__hash__()
__eq__(other)