Shared-Memory Runtime — moved to uni_rl¶
The async IPC layer moved out of the unilab package into the independently
released uni_rl package (issue #1480): uni_rl.ipc hosts the async
runner, shared-memory buffers, replay pipelines, inference slot, DP launcher /
sync, and weight sync.
Submodule |
Role |
|---|---|
|
The high-level orchestration loop |
|
NumPy-backed shared-memory ring/buffer |
|
Rollout window used by on-policy collectors |
|
Bounded shared ingress for off-policy transitions |
|
Authoritative CUDA/MPS replay ring, device gather, and native H2D |
|
Fixed shared observation/action slot for learner-owned off-policy inference |
|
Push learner weights to on-policy collector workers |