INNER CODE UNIT · Python

_TinygradProxy

onnxsim/onnxsim · onnxsim/rpc/server.py:235

class _TinygradProxy:
    """A tinygrad model living in its own worker process.

    One process per loaded model means every (device, codegen options) pair starts from clean
    kernel and schedule caches -- so a ``BEAM`` setting really applies instead of silently reusing
    kernels compiled earlier under another setting -- tinygrad's thread-bound state (its SQLite
    disk cache, device contexts) stays on the worker's main thread, and a GPU fault cannot take
    the server down.
    """

    def __init__(
        self, model_bytes: bytes, device: Optional[str], options, work_dir: str
    ):
        import importlib.util
        import multiprocessing

        if importlib.util.find_spec("tinygrad") is None:
            raise proto.RPCError(

View source record →