xnn.gnn.models.blocks.RECOMPUTE_MIN_EDGES

xnn.gnn.models.blocks.RECOMPUTE_MIN_EDGES = 1000000

Rows (edges) from which a radial network in evaluation mode recomputes its hidden layers in the backward pass instead of keeping their activations (recompute="auto"). Measured on an L40S with the D4 of the production water model, 5k-41k water atoms: the peak falls by 22% for the production model (32 channels) and 5-7% for MACE-OFF23 medium and MACE-MP-0b2 large; the step takes 0-5% longer (one forward of the hidden layers, 4-5% of a step).