Concurrency

• get() and get_batch() are thread-safe and release the GIL. For method=1/2, calls on one variable of one store take turns (per-variable lock); for parallel reads give each thread a fork_reader(), as ThreadDataLoader does.

• method=0 get_batch() is collective: same number of calls, same order, one thread per rank. vae-ddp.py allows no workers with method=0.

• Only the main thread calls MPI; FUNNELED is the minimum thread level (mpi4py defaults to MULTIPLE; keep it if your threads call MPI).