cjm-substrate-torch-utils

PyTorch helpers for cjm-substrate capabilities: GPU memory release, typed CUDA-OOM handling, and device selection.

Modules

• cjm_substrate_torch_utils.__init__
• cjm_substrate_torch_utils.device — Resolve a device spec (auto / cpu / cuda / cuda:N) to a concrete torch device string.
• cjm_substrate_torch_utils.memory — Robust move-to-CPU + drop-references + gc + CUDA-cache cleanup for releasing models, factored out of the per-capability reimplementations.
• cjm_substrate_torch_utils.oom — Convert torch CUDA out-of-memory exceptions into the substrate's typed CapabilityResourceError (SG-47 Track B) so CR-7 reactive retry can evict and reload.

API

cjm_substrate_torch_utils.device

• resolve_torch_device function — Resolve a device spec to a concrete torch device string.

cjm_substrate_torch_utils.memory

• release_model function — Release one or more model objects: move to CPU, drop references, gc, free CUDA cache.

cjm_substrate_torch_utils.oom

• cuda_oom_to_capability_resource_error function — Convert a CUDA out-of-memory exception into a substrate-typed CapabilityResourceError.
• is_cuda_oom function — Classify CUDA OOM across BOTH raise paths: the caching allocator's

Dependencies

Depends on: cjm-substrate, numpy, torch Used by: cjm-capability-demucs, cjm-capability-pyannote, cjm-capability-qwen3-forced-aligner, cjm-capability-voxtral-hf, cjm-capability-whisper, cjm-substrate-hf-utils