---

name: cpu-mode-assertions-torch description: Assert CPU-only runtime inside a container using PyTorch (torch.cuda.is_available()==False) and optional env var checks. Use to prevent accidental GPU execution during CPU smoke tests.

Skill Instructions

Inputs

• SERVICE_CONTAINER (string): container name to exec into
• CPU_DEVICE_ENV (string, optional): env var to validate (e.g. ASR_DEVICE)
• CPU_DEVICE_EXPECTED (string, optional): expected value (e.g. cpu)

Procedure

Copyset -euo pipefail

service_container="${SERVICE_CONTAINER:?SERVICE_CONTAINER is required}"

# 1) Env assertion (optional)
if [ -n "${CPU_DEVICE_ENV:-}" ]; then
  docker exec "$service_container" printenv "$CPU_DEVICE_ENV"
fi

# 2) Runtime assertion (torch)
docker exec "$service_container" python - <<'PY'
import torch
print('torch_version=', torch.__version__)
print('torch_cuda_is_available=', torch.cuda.is_available())
print('torch_cuda_version=', getattr(torch.version, 'cuda', None))
PY

Acceptance Criteria

• torch_cuda_is_available prints False
• If CPU_DEVICE_ENV is provided, its value equals CPU_DEVICE_EXPECTED (or cpu if you standardize it)

Why

Prevents “silent GPU path” during CPU smoke testing (different behavior, heavier images, device-specific bugs).