Generation backends

wrbench generate compiles payloads and sidecars by default. Real generation is enabled only with --no-dry-run and a runtime configuration.

Configure a backend

1. Copy wrbench.runtime.example.json to an untracked wrbench.runtime.json.
2. Fill in the upstream Python interpreter, repository, checkpoint, and model-specific paths.
3. Check the model record and adapter with wrbench doctor --model <model-key>.
4. Run the same compile command with --runtime-config and --no-dry-run.

IMAGE="$(python - <<'PY'
from wrbench.datasets import natural25_first_frame_path
print(natural25_first_frame_path("bedroom_cat_bed_jump"))
PY
)"

wrbench generate \
  --model easyanimate-v51-camera \
  --camera preset:yaw_LR \
  --image "$IMAGE" \
  --prompt "A living room." \
  --out out.mp4 \
  --runtime-config /path/to/wrbench.runtime.json \
  --no-dry-run

The runtime file is the only place for machine-specific paths. Do not add checkpoints, credentials, or upstream repository paths to model JSON files.

Built-in execution modes

Mode	Behavior
Compile-only (default)	Writes model-native controls and audit sidecars; no model environment is required
local_subprocess	Launches a supported upstream model in its configured environment

Supported launchers and their required runtime fields are defined by the model records and wrbench.runtime.example.json. Run wrbench models for the current registry instead of relying on a duplicated model list here.

Add backend support

1. Add or reuse a GenerationBackend under src/wrbench/backends/.
2. For subprocess execution, add a launcher under src/wrbench/backends/launchers/.
3. Connect it to the model's execution contract and runtime-config fields.
4. Add focused tests under tests/test_backends*.py.
5. Document only the new runtime fields in wrbench.runtime.example.json.

Model registry and adapter work is covered by Adding a model.