datasets.map(num_proc) can hang after torch ran in the parent

On Linux with the fork start method, calling a torch-backed preprocessing function (e.g. a transformers feature extractor) once in the parent process and then running datasets' map with num_proc=4 left all four forked workers idle at ~0% CPU indefinitely. That is the classic fork-after-threads deadlock: the parent's OpenMP/intra-op thread pool does not survive fork. Setting OMP_NUM_THREADS=1 before importing torch made the same script finish in seconds. If a num_proc map stalls with idle workers, check whether torch/BLAS threads were started in the parent before blaming the data. Also make sure you're not just waiting on a job-slot wrapper.