Configure workers and GPUs

Use worker_environment when workers in one pool need different process environment variables. A common use is assigning one GPU to each worker.

with environment.start(
    workers=2,
    worker_environment=lambda index: {
        "CUDA_VISIBLE_DEVICES": str(index),
    },
) as workers:
    # Worker 0 sees CUDA_VISIBLE_DEVICES=0.
    # Worker 1 sees CUDA_VISIBLE_DEVICES=1.
    ...

Wetlands calls the callback once for each zero-based worker index before launching any worker and copies the returned mappings. A replacement worker keeps the same index and receives the same copied mapping.

Each mapping must contain only string keys and string values. Names cannot be empty or contain = or null bytes, and values cannot contain null bytes.

The case-insensitive WETLANDS_* namespace is reserved. PYTHONEXECUTABLE, PYTHONHOME, and PYTHONPATH also cannot be set because Wetlands controls the worker interpreter and import environment.

Indexed environments cannot be combined with persistent=True. A later controller would not have the original callback configuration to validate a reconnected worker.