2D sharding

Classic height / width / block sharding as four steps from TensorSpec + TensorLayout, then Buffer / BufferDistributionSpec for shard→core. Default demo: logical [2,3,5] TILE 2×2 → padded [2,4,6] → physical [8,6] → block shards on a 2×2 core grid. Step 1 is always the original logical input.

Presets
Tensor
N-D, e.g. 2,3,5 or 64,96
×
Sharding
Pick one — the other is derived (1 shard / core).
×
Sets how many cores; shard shape is even-split auto.
×
Height: H; width: W; block: H×W. Core grid is derived from physical ÷ shard.
How the shard grid maps onto the core grid (and auto-split for block).
Alignment{} → create_default_alignment (tile or RM row/shard width).
Step 1 / 6 — input

Tensor

Cores

stats