Set all num_workers to 0 to eliminate all forked subprocesses
preprocessing_num_workers=1 still forks (nch went 3→2 not 0), and
the CoW spike from even 1 fork at ~1GB parent PSS caused wav2vec2 to
still peak at 5.8GB. The script default is None (no fork) — passing
any integer forks that many workers.
dataloader_num_workers=0 loads data inline in the main process with no
persistent worker subprocesses. For tests running max_steps=2-10 on
tiny fixture datasets there is no throughput benefit to workers.
With all num_workers=0 no child processes are spawned → no CoW → PSS
reflects only the actual model/optimizer/data footprint.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>