You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
## Design option: RPC-Aware Ops
This was the first idea.
+[ ] Find `Op`s in the graph that use the `ArraysToArraysServiceClient` and can be parallelized (they must not depend on each other). This can be implemented by adding a mixin-interface by which an `isinstance(op, ArraysToArraysOp)` can be identified.
+[ ] Write a `ParallelArraysToArraysOp` that keeps a list of streams and runs evaluations in parallel.
+[ ] Do a subgraph replacement where the independent `ArraysToArraysOp`s nodes are substituted by a subgraph that routes the inputs to a new `ParallelArraysToArraysOp` node and redistributes the outputs.
Design option: Async Ops (preferred)
This would be RPC-unaware and more generic overall.
Old design idea
Design option: Async Ops (preferred)
This would be RPC-unaware and more generic overall.
AsyncOp
class #26AsyncArraysToArraysOp
,AsyncLogpOp
,AsyncLogpGradOp
.class ParallelAsyncOp(Op)
similar toaesara.graph.basic.Composite
that parallelizes the.perform_async()
calls of a bunch ofAsyncOp
s.AsyncOp
s and merges them into anParallelAsyncOp
The text was updated successfully, but these errors were encountered: