Crate monarch_tensor_worker

A hyperactor-based implementation of a PyTorch worker actor.

The worker is responsible for executing PyTorch operations on a local device. It assumes it has exclusive access to device resources, and manages concurrency internally via device-specific constructs (CUDA stream, threads, etc.).

This is a port of monarch/python/controller/worker.py but does have gaps due to drift that needs to be reconciled. This mainly includes:

• Support for record and replay
• debugger support
• general drift in exisitng messages

Modules

bootstrap

device_mesh

pipe

py_pipe

stream

test_util

Structs

WorkerActor

A PyTorch runtime instance, operating on a single accelerator device, controlled via hyperactor messaging.

Enums

AssignRankMessage

Worker messages. These define the observable behavior of the worker, so the documentations here

Traits

AssignRankMessageClient

The custom client trait for this message type.

AssignRankMessageHandler

The custom handler trait for this message type.