serve.py

class YourPredictor:                          # any name — the yaml points at it

    def __init__(self, **init_args):          # REQUIRED
        # init_args come verbatim from inference.yaml → model.init_args.
        # Build the model, pick a device, load tokenizers. Runs once per replica.
        ...

    def predict(self, data: bytes) -> dict:   # REQUIRED — may be async def
        # `data` is the normalized request body. Return a JSON-serializable dict.
        ...

    def load_weights(self, path: str) -> None:  # optional
        # Called after __init__ IF the yaml sets model.path. `path` is the
        # local file the platform mounted for you.
        ...

    def warmup(self) -> None:                 # optional
        # Hit once per replica after deploy — pre-warm JIT / autotune caches.
        ...

    def health(self) -> dict:                 # optional
        # Returned verbatim on GET /health. Absent → {"status": "ok"}.
        ...

class Echo:
    def __init__(self, prefix: str = "echo: "):
        self.prefix = prefix

    def predict(self, data: bytes) -> dict:
        return {"echo": self.prefix + data.decode("utf-8", errors="replace")}