RayOrch

Lightweight orchestration utilities for building asynchronous Ray pipelines with RayModule, overlapped microbatch execution, and DAG-style scheduling.

Install

pip install rayorch

For development:

pip install -r requirements-dev.txt

Core Concepts

• RayModule: wraps an operator class into Ray actors with optional replica dispatch and collect.
• OverlappedPipeline: graphless microbatch overlap with backpressure.
• DagPipeline / DagPipelineExecutor: explicit dependency DAG scheduling.

Minimal Example

fromrayorchimportOverlappedPipeline, RayModuleclassAddOne:
defrun(self, x):
returnx+1classPipe(OverlappedPipeline):
def__init__(self):
self.a=RayModule(AddOne, replicas=1).pre_init()
self.b=RayModule(AddOne, replicas=1).pre_init()
super().__init__(max_inflight=4)
defforward(self, x):
returnself.b(self.a(x))
pipe=Pipe()
print(pipe([1, 2, 3])) # [3, 4, 5]

License

Apache-2.0. See LICENSE.