Upload hpo_train_shim.py with huggingface_hub

hpo_train_shim.py

@@ -45,12 +45,14 @@ from lerobot.datasets import EpisodeAwareSampler as _RealSampler
 # -> random segfaults mid-training. This is NOT a resource problem; more RAM or
 # a bigger GPU does not help.
 #
-# Starting workers with `forkserver` (or `spawn`) gives each worker a clean
-# process instead of a fork of the threaded parent, which removes the crash and
-# keeps num_workers > 0 usable. Override with HPO_MP_CONTEXT=spawn|fork.
+# `spawn` starts each worker as a fresh interpreter that inherits nothing, so the
+# threaded PyAV state never reaches it. (`forkserver` is NOT enough: its server
+# process is itself forked from the parent, so if PyAV/FFmpeg is already loaded
+# by then the server inherits the broken state and every worker forked from it
+# does too.) Override with HPO_MP_CONTEXT=forkserver|fork.
 # ---------------------------------------------------------------------------
 
-_MP_CONTEXT = os.environ.get("HPO_MP_CONTEXT", "forkserver")
+_MP_CONTEXT = os.environ.get("HPO_MP_CONTEXT", "spawn")
 
 _RealDataLoader = torch.utils.data.DataLoader