Multiprocessing and Queues

Multiprocessing, Queues and Coordination

1. mp.get_context("spawn") creates a compatible worker process per device on Windows and CUDA.
2. A bounded task_queue receives length-ranked minibatches; a producer thread prevents the parent from blocking during initialization.
3. Workers all own the same result_queue and emit ready, batch, finished, or error messages.
4. load_lock serializes model initialization to avoid simultaneous downloads/loading peaks.
5. go_event coordinates the processing start after every worker has reported readiness.
6. receive() times out, detects process death, and propagates worker tracebacks.
7. Parent stores an answers dictionary keyed by original row index, detects duplicate indices, and writes output in order.

This is a local multiprocessing process pool, not inter-machine distributed inference. The producer's tasks and the result queue are process-local IPC mechanisms, not Kafka/RabbitMQ.