hivemind/hivemind/server/expert_backend.py

from typing import Dict, Sequence, Any, Tuple, Union

import torch
from torch import nn

from hivemind.server.task_pool import TaskPool
from hivemind.utils import nested_flatten, nested_pack, nested_compare, nested_map,\
    BatchTensorDescriptor, DUMMY_BATCH_SIZE


class ExpertBackend(nn.Module):
    """
    ExpertBackend is a wrapper around torch module that allows it to run tasks asynchronously with Runtime
    By default, ExpertBackend handles three types of requests:

     - forward - receive inputs and compute outputs. Concurrent requests will be batched for better GPU utilization.
     - backward - receive gradients w.r.t. outputs, compute gradients w.r.t. inputs and **update expert**. Also batched.
     - get_info - return expert metadata. Not batched.

    :param expert: nn.Module to be wrapped into a backend. Arbitrary pytorch module with a few limitations:

     - Experts must always receive the same set of args and kwargs and produce output tensors of same type
     - All args, kwargs and outputs must be **tensors** where 0-th dimension represents to batch size
     - We recommend using experts that are ~invariant to the order in which they process batches
     - Using randomness (e.g. Dropout) leads to different samples at forward and backward. If you want consistency,
        you should explicitly register these random variables as model inputs or outputs.
        See hivemind.utils.custom_layers.DeterministicDropout for an example

    :param opt: torch optimizer to be applied on every backward call
    :param args_schema: description of positional arguments to expert.forward, list of BatchTensorProto
    :param kwargs_schema: description of keyword arguments to expert.forward, dict of BatchTensorProto
    :param outputs_schema: description of outputs from expert.forward, nested structure of BatchTensorProto
    :param kwargs: extra parameters to be forwarded into TaskPool.__init__
    """

    def __init__(self, name: str, expert: nn.Module, opt: torch.optim.Optimizer, *,
                 args_schema: Tuple[BatchTensorDescriptor, ...] = None,
                 kwargs_schema: Dict[str, BatchTensorDescriptor] = None,
                 outputs_schema: Union[BatchTensorDescriptor, Tuple[BatchTensorDescriptor, ...]] = None,
                 **kwargs):
        super().__init__()
        self.expert, self.opt, self.name = expert, opt, name

        self.args_schema = args_schema = tuple(args_schema or ())
        self.kwargs_schema = kwargs_schema = dict(kwargs_schema or {})
        assert args_schema or kwargs_schema, "expert must receive at least one positional or keyword input." \
                                             " Did you forget to provide args_schema/kwargs_schema?"

        if outputs_schema is None:
            # run expert once to get outputs schema
            dummy_args = tuple(sample.make_empty(DUMMY_BATCH_SIZE) for sample in args_schema)
            dummy_kwargs = {key: sample.make_empty(DUMMY_BATCH_SIZE) for key, sample in kwargs_schema.items()}
            dummy_outputs = self.expert(*dummy_args, **dummy_kwargs)
            outputs_schema = nested_map(BatchTensorDescriptor.from_tensor, dummy_outputs)

        self.outputs_schema = outputs_schema
        self.forward_schema = (self.args_schema, self.kwargs_schema)
        self.backward_schema = (self.forward_schema, self.outputs_schema)  # original inputs and grad w.r.t. outputs
        self.forward_pool = TaskPool(self.forward, uid=f'{self.name}_forward', **kwargs)
        self.backward_pool = TaskPool(self.backward, uid=f'{self.name}_backward', **kwargs)

    def forward(self, *inputs: torch.Tensor) -> Tuple[torch.Tensor, ...]:
        """
        Apply forward pass to an aggregated batch of requests. Used by Runtime, do not call this manually;
        To submit a request for asynchronous processing, please use ``ExpertBackend.forward_pool.submit_task``.

        Subclassing:
           This method receives a sequence of torch tensors following ``nested_flatten(self.forward_schema)``;

           It should return gradients w.r.t. inputs that follow ``nested_flatten(self.outputs_schema)``;

           .. todo we handle layer states (e.g. batchnorm stats) incorrectly, updating them twice.
           .. For now, either register all buffers as outputs or avoid stateful experts

        """
        args, kwargs = nested_pack(inputs, structure=self.forward_schema)

        with torch.no_grad():
            outputs = self.expert(*args, **kwargs)

        # Note: TaskPool requires function to accept and return a flat tuple of values, we pack/unpack it on client side
        return tuple(nested_flatten(outputs))

    def backward(self, *inputs: torch.Tensor) -> Tuple[torch.Tensor, ...]:
        """
        Apply backward pass to an aggregated batch of requests. Used by Runtime, do not call this manually
        To submit a request for asynchronous processing, please use ``ExpertBackend.backward_pool.submit_task``.

        Subclassing:
           This method receives a sequence of torch tensors following ``nested_flatten(self.backward_schema)``;

           It should return gradients w.r.t. inputs that follow ``nested_flatten(self.forward_schema)``;

           Runtime doesn't guarantee that backward will be performed in the same order and for the same data
           as forward, so we recommend stateless backward pass that re-runs expert forward pass inside backward.

           .. todo correct state handling (see forward)

           Please make sure to call ``ExpertBackend.apply_gradients`` here, otherwise the expert will not train
        """
        (args, kwargs), grad_outputs = nested_pack(inputs, structure=self.backward_schema)

        with torch.enable_grad():
            args = [tensor.detach().requires_grad_(True) if tensor.dtype in (torch.half, torch.float, torch.double)
                    else tensor.detach() for tensor in args]
            kwargs = {input_key: (tensor.detach().requires_grad_(True)
                                  if tensor.dtype in (torch.half, torch.float, torch.double)
                                  else tensor.detach())
                      for input_key, tensor in kwargs.items()}

            outputs = self.expert(*args, **kwargs)
            assert nested_compare(outputs, grad_outputs), "outputs and grad_outputs must have the same structure"

            outputs_flat = tuple(nested_flatten(outputs))

            grad_outputs_flat = tuple(map(
                lambda grad, out: grad.to(device=out.device, dtype=out.dtype, non_blocking=True),
                nested_flatten(grad_outputs), outputs_flat))
            torch.autograd.backward(outputs_flat, grad_tensors=grad_outputs_flat,
                                    create_graph=False, retain_graph=False)
            self.apply_gradients()

        return tuple(x.grad if isinstance(x.grad, torch.Tensor) else torch.zeros_like(x)
                     for x in nested_flatten((args, kwargs)))

    def apply_gradients(self) -> None:
        """
        Train the expert for one step. This method is called by ``ExpertBackend.backward`` after computing gradients.
        """
        self.opt.step()
        self.opt.zero_grad()

    def get_info(self) -> Dict[str, Any]:
        """ Get expert parameters and stats. Used by RemoteExpert to check shapes and for DMoE orchestration. """
        return dict(forward_schema=self.forward_schema, outputs_schema=self.outputs_schema,
                    keyword_names=tuple(self.kwargs_schema.keys()))

    def get_pools(self) -> Sequence[TaskPool]:
        """ return all pools that should be processed by ``Runtime`` """
        return self.forward_pool, self.backward_pool