mstar.model.waypoint.ring_geometry

mstar.model.waypoint.ring_geometry#

Functions

describe_ring_memory(config, *[, ...])

Geometry and footprint of every ring, grouped local vs global, with the counterfactual under the opposite full_global_ring setting.

ring_memory_bytes(config, *[, num_sessions, ...])

Per-layer ring bytes for config, computed without allocating anything (so it can be called on a laptop while sizing a deployment).
mstar.model.waypoint.ring_geometry.describe_ring_memory(config, *, num_sessions=1, dtype=torch.bfloat16)[source]#

Geometry and footprint of every ring, grouped local vs global, with the counterfactual under the opposite full_global_ring setting.

The compacted default is what a reviewer should see; the full_global_ring line is the reference’s allocation, 8/9ths of whose global storage is permanently unwritten.

Parameters:
Return type:

str

mstar.model.waypoint.ring_geometry.ring_memory_bytes(config, *, num_sessions=1, dtype=torch.bfloat16)[source]#

Per-layer ring bytes for config, computed without allocating anything (so it can be called on a laptop while sizing a deployment).

num_sessions is a straight multiplier: the world dimension is folded into the token axis, so N worlds is N times one world’s slots and the per-slot arithmetic is untouched. That it is a multiplier is the whole reason num_sessions is a deployment sizing knob and the geometry is not.

Parameters:
Return type:

list[int]