Language

unilab.algos.torch.appo.staging¶

Bounded rollout staging for APPO learners.

Classes

RolloutStagingPool

Preallocated learner-device storage for a bounded set of rollouts.

class unilab.algos.torch.appo.staging.RolloutStagingPool[source]¶

Bases: object

Preallocated learner-device storage for a bounded set of rollouts.

Raw IPC rollout slots are env-major: [N, T, …]. Learners consume time-major combined batches: [T, K*N, …]. The pool owns the destination tensors and exposes active views without rebuilding them via torch.cat.

Parameters:

capacity (int)
num_envs (int)
slot_shapes (Mapping[str, tuple[int, ...]])
device (str | device)

__init__(*, capacity, num_envs, slot_shapes, device)[source]¶

Parameters:

capacity (int)
num_envs (int)
slot_shapes (Mapping[str, tuple[int, ...]])
device (str | device)

property active_count: int¶

property slot_versions: tuple[int, ...]¶

stage_numpy_views(raw_views)[source]¶

Copy one raw shared-memory rollout into the next staging slot.

Parameters:: raw_views (Mapping[str, ndarray])
Return type:: int

batch()[source]¶

Return active learner-ready views backed by the staging pool.

Return type:: dict[str, Tensor]