Shared-Memory Runtime — unilab.rl.runtime

The async IPC layer lives in the unilab distribution under unilab.rl.runtime (the former uni_rl.ipc / uni_rl.offpolicy modules are merged into this repository; issue #2372): async runner, shared-memory buffers, replay pipelines, inference exchange, DP launcher / sync, and weight sync.

Submodule

Role

unilab.rl.runtime.async_runner

The high-level orchestration loop

unilab.rl.runtime.ipc.tensor_rollout_ring_buffer

Tensor-backed rollout window used by APPO collectors

unilab.rl.runtime.ipc.replay_buffer

Bounded shared ingress for off-policy transitions

unilab.rl.runtime.ipc.replay_pipelines.*

Authoritative CUDA/MPS replay ring, device gather, and native H2D

unilab.rl.runtime.ipc.inference_ring

Bounded, ordered observation/action exchange for learner-owned off-policy inference

unilab.rl.runtime.ipc.weight_sync

Push learner weights to on-policy collector workers

unilab.rl.runtime.distributed.*

torchrun launch, rank context, CPU affinity, and DP parameter sync