Class incremental_save

lit_gpt/utils_old.py:365–404

class incremental_save:
    def __init__(self, name):
        self.name = name
        self.zipfile = torch._C.PyTorchFileWriter(str(name))
        self.has_saved = False
        self.next_key = 0

    def __enter__(self):
        return self

    def store_early(self, tensor):
        if isinstance(tensor, torch.Tensor):
            return SavingProxyForTensor(tensor, self)
        raise TypeError(f"can only store tensors early, not {type(tensor)}")

    def save(self, obj):
        if self.has_saved:
            raise RuntimeError("have already saved")
        # Write the pickle data for `obj`
        data_buf = BytesIO()
        pickler = IncrementalPyTorchPickler(self, data_buf, protocol=5)
        pickler.dump(obj)
        data_value = data_buf.getvalue()
        self.zipfile.write_record("data.pkl", data_value, len(data_value))
        self.has_saved = True

    def _write_storage_and_return_key(self, storage):
        if self.has_saved:
            raise RuntimeError("have already saved")
        key = self.next_key
        self.next_key += 1
        name = f"data/{key}"
        if storage.device.type != "cpu":
            storage = storage.cpu()
        num_bytes = storage.nbytes()
        self.zipfile.write_record(name, storage.data_ptr(), num_bytes)
        return key

    def __exit__(self, type, value, traceback):
        self.zipfile.write_end_of_file()


T = TypeVar("T")