Inference

Inference is a separate YAML-driven pipeline:

chimera-ml inference \
  --input video.mp4 \
  --output prediction.json \
  --config-path examples/oragen/configs/inference.yaml \
  --device auto \
  --work-dir work

The CLI accepts device auto, cpu, or cuda. auto selects CUDA when available, otherwise CPU. If work-dir is omitted, a temporary directory is created.

Config shape

pipeline:
  name: demo
  parallel: false

runtime:
  device: auto

steps:
  - name: extract_audio_step
    params:
      sample_rate: 16000
  - name: print_json_predictions_step

Every step is resolved from INFERENCE_STEPS and implements run(ctx: InferenceContext) -> InferenceContext. The context has input_path, work_dir, device, raw config, and shared artifacts. Use set_artifact and get_artifact; predictions is a convenience property for the predictions key.

Sequential and DAG modes

With parallel false or omitted, steps run in config order and each depends on the previous step. With parallel true, after is authoritative: steps without dependencies are roots and can start together. Step ids default to name; repeated names need explicit id values. Unknown dependencies, self-dependencies, duplicate ids, and cycles raise ValueError.

Each parallel step receives a copy of current artifacts. Results merge using keys written through set_artifact. A branch may overwrite an upstream artifact when it depends on that upstream node; unrelated branches may not write the same key.

Checkpoints and output

resolve_checkpoints_step accepts local files or http/https references, writes remote files to a work-dir-relative cache, and exposes absolute paths in artifacts["checkpoints"]. Cached files are reused unless force_download is true. Failed downloads remove partial files.

write_json_predictions_step writes input and predictions JSON; the print step prints the same payload. The CLI --output can create or override the write step in sequential mode. In parallel mode it cannot auto-create it: declare the step explicitly and set after dependencies in YAML.