NVIDIA PhysicsNeMo Core (Latest Release)

deeplearning/physicsnemo/physicsnemo-core/_modules/physicsnemo/distributed/utils.html

Source code for physicsnemo.distributed.utils

# SPDX-FileCopyrightText: Copyright (c) 2023 - 2024 NVIDIA CORPORATION & AFFILIATES.
# SPDX-FileCopyrightText: All rights reserved.
# SPDX-License-Identifier: Apache-2.0
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

# TODO this also needs more docstrings
from typing import List, Optional

import torch
import torch.distributed as dist
import torch.nn as nn
import torch.nn.functional as F

from .manager import DistributedManager


def compute_split_shapes(size: int, num_chunks: int) -> List[int]:
    # treat trivial case first
    if num_chunks == 1:
        return [size]

    # first, check if we can split using div-up to balance the load:
    chunk_size = (size + num_chunks - 1) // num_chunks
    last_chunk_size = max(0, size - chunk_size * (num_chunks - 1))
    if last_chunk_size == 0:
        # in this case, the last shard would be empty, split with floor instead:
        chunk_size = size // num_chunks
        last_chunk_size = size - chunk_size * (num_chunks - 1)

    # generate sections list
    sections = [chunk_size for _ in range(num_chunks - 1)] + [last_chunk_size]

    return sections


[docs]def get_memory_format(tensor): """Gets format for tensor""" if tensor.is_contiguous(memory_format=torch.channels_last): return torch.channels_last else: return torch.contiguous_format
[docs]def pad_helper(tensor, dim, new_size, mode="zero"): """Util for padding tensors""" ndim = tensor.ndim dim = (dim + ndim) % ndim ndim_pad = ndim - dim output_shape = [0 for _ in range(2 * ndim_pad)] orig_size = tensor.shape[dim] output_shape[1] = new_size - orig_size tensor_pad = F.pad(tensor, output_shape, mode="constant", value=0.0) if mode == "conj": lhs_slice = [ slice(0, x) if idx != dim else slice(orig_size, new_size) for idx, x in enumerate(tensor.shape) ] rhs_slice = [ slice(0, x) if idx != dim else slice(1, output_shape[1] + 1) for idx, x in enumerate(tensor.shape) ] tensor_pad[lhs_slice] = torch.flip( torch.conj(tensor_pad[rhs_slice]), dims=[dim] ) return tensor_pad
[docs]def truncate_helper(tensor, dim, new_size): """Util for truncating""" input_format = get_memory_format(tensor) ndim = tensor.ndim dim = (dim + ndim) % ndim output_slice = [ slice(0, x) if idx != dim else slice(0, new_size) for idx, x in enumerate(tensor.shape) ] tensor_trunc = tensor[output_slice].contiguous(memory_format=input_format) return tensor_trunc

def split_tensor_along_dim(tensor, dim, num_chunks): if dim >= tensor.dim(): raise ValueError( f"Error, tensor dimension is {tensor.dim()} which cannot be split along {dim}" ) if tensor.shape[dim] < num_chunks: raise ValueError( "Error, cannot split dim {dim} of size {tensor.shape[dim]} into \ {num_chunks} chunks. Empty slices are currently not supported." ) # get split sections = compute_split_shapes(tensor.shape[dim], num_chunks) tensor_list = torch.split(tensor, sections, dim=dim) return tensor_list

[docs]@torch.no_grad() def reduce_loss(loss: float, dst_rank: int = 0, mean: bool = True): # pragma: no cover """Reduces loss from all processes to destination rank for logging. Parameters ---------- loss : float loss value dst_rank : int, Optional destination rank to redce to, by default 0. mean : bool, Optional Calculate the mean of the losses gathered, by default True. Raises ------ Exception If DistributedManager has yet to be initialized """ if not DistributedManager.is_initialized(): raise Exception( "Distributed manager should be initialized when using reduce_loss" ) distmng = DistributedManager() loss = torch.Tensor([loss]).to(distmng.device) # For serial runs, just return the current loss! if distmng.world_size == 1: return float(loss) op = torch.distributed.ReduceOp.SUM if not mean else torch.distributed.ReduceOp.AVG torch.distributed.reduce(loss, dst_rank, op, group=None) # Return loss if dst_rank, None otherwise if distmng.rank == dst_rank: return float(loss.cpu()) else: return None

# distributed primitives

[docs]def distributed_transpose(tensor, dim0, dim1, group=None, async_op=False): """Perform distributed transpose of tensor to switch sharding dimension""" # get input format input_format = get_memory_format(tensor) # get comm params comm_size = dist.get_world_size(group=group) # split and local transposition split_size = tensor.shape[dim0] // comm_size x_send = [ y.contiguous(memory_format=input_format) for y in torch.split(tensor, split_size, dim=dim0) ] x_recv = [torch.empty_like(x_send[0]) for _ in range(comm_size)] # global transposition req = dist.all_to_all(x_recv, x_send, group=group, async_op=async_op) return x_recv, req

def _reduce(input_, use_fp32=True, group=None): # pragma: no cover """All-reduce the input tensor across model parallel group.""" # Bypass the function if we are using only 1 GPU. if dist.get_world_size(group=group) == 1: return input_ # All-reduce, use_fp32 only relevant for lower precisions # if input is already in double precision, nothing changes if use_fp32 and (input_.dtype.itemsize < 4) and input_.dtype.is_floating_point: dtype = input_.dtype inputf_ = input_.float() dist.all_reduce(inputf_, group=group) input_ = inputf_.to(dtype) else: dist.all_reduce(input_, group=group) return input_ def _split(input_, dim_, group=None): # pragma: no cover """Split the tensor along its last dimension and keep the corresponding slice.""" # get input format input_format = get_memory_format(input_) # Bypass the function if we are using only 1 GPU. comm_size = dist.get_world_size(group=group) if comm_size == 1: return input_ # Split along last dimension. input_list = split_tensor_along_dim(input_, dim_, comm_size) # Note: torch.split does not create contiguous tensors by default. rank = dist.get_rank(group=group) output = input_list[rank].contiguous(memory_format=input_format) return output

[docs]def all_gather_v_wrapper( tensor: torch.Tensor, sizes: Optional[List[int]] = None, dim: int = 0, group: Optional[dist.ProcessGroup] = None, ) -> torch.Tensor: # pragma: no cover """ Implements a distributed AllGatherV primitive. It is based on the idea of a single global tensor which is distributed along a specified dimension into chunks of variable size. This primitive gathers all local tensors from each rank into the full global tensor onto each rank. Parameters ---------- tensor : "torch.Tensor" local tensor on each rank sizes : List[int], optional list of the sizes of each chunk on each rank along distributed dimension, valid and set on each rank, by default None. Can be single integer per rank (assuming all other dimensions except `dim` below are equal) or can be full dim : int, optional dimension along which global tensor is distributed, by default 0 group : Optional[dist.ProcessGroup], optional process group along which global tensor is shared, by default None Returns ------- torch.Tensor full global tensor, valid on each rank """ comm_size = dist.get_world_size(group=group) if (sizes is not None) and (len(sizes) != comm_size): raise ValueError(f"Mismatch in sizes {len(sizes)} and comm_size {comm_size}") if dim >= tensor.dim(): raise ValueError() if comm_size == 1: return tensor # This is valid if the the shape is a list of ints, but not if full tensor # shapes are passed on each rank. Check if each element of sizes itself is iterable: tensor_format = get_memory_format(tensor) if sizes is not None: full_shapes = False try: iterator = iter(sizes[0]) # noqa: F841 except TypeError: # Not iterable, use base tensor shape: tensor_shape = list(tensor.shape) else: # it is iterable, use shapes directly full_shapes = True tensor_shape = None # Catch and replace below tensor_list = [None] * comm_size for src in range(comm_size): if full_shapes: tensor_shape = sizes[src] else: tensor_shape[dim] = sizes[src] tensor_list[src] = torch.empty( tensor_shape, dtype=tensor.dtype, device=tensor.device, ) else: # assume equal shape on all ranks tensor_list = [torch.empty_like(tensor) for _ in range(comm_size)] dist.all_gather(tensor_list, tensor, group=group) output = torch.cat(tensor_list, dim=dim).contiguous(memory_format=tensor_format) return output
[docs]def all_gather_v_bwd_wrapper( tensor: torch.Tensor, sizes: List[int], dim: int = 0, use_fp32: bool = True, group: Optional[dist.ProcessGroup] = None, ) -> torch.Tensor: # pragma: no cover """ Implements a distributed AllReduceV primitive. It is based on the idea of a single global tensor which which can be distributed along a specified dimension into chunks of variable size. This primitive assumes different global tensors of the same shape on each rank. It then re-distributes chunks of all these tensors such that each rank receives all corresponding parts of a global tensor. Each rank then sums up the chunks after receiving it. By design, this primitive thus implements the backward pass of the "all_gather_v" primitive. In this case, the result would be a single global gradient tensor distributed onto different ranks. Parameters ---------- tensor : torch.Tensor global tensor on each rank (different one on each rank) sizes : List[int] list of the sizes of each chunk on each rank along distributed dimension, valid and set on each rank dim : int, optional dimension along which global tensor is distributed, by default 0 use_fp32 : bool, optional flag to specify reduction taking place at least in FP32 precision, by default True only acts on floating point inputs in lower precision group : Optional[dist.ProcessGroup], optional process group along which global tensor is shared, by default None Returns ------- torch.Tensor local tensor, i.e. result of reduction of all corresponding chunks from all global tensors for each rank separately """ comm_size = dist.get_world_size(group=group) rank = dist.get_rank(group=group) if len(sizes) != comm_size: raise ValueError() if dim >= tensor.dim(): raise ValueError() tensor_shape = list(tensor.shape) tensor_shape[dim] = sizes[rank] tmp = [ torch.empty( tensor_shape, dtype=tensor.dtype, device=tensor.device, ) for _ in range(comm_size) ] scatter_list = list(torch.split(tensor, sizes, dim=dim)) scatter_list = [t.contiguous() for t in scatter_list] dist.all_to_all(tmp, scatter_list, group=group) stack_dim = tensor.dim() tmp = torch.stack(tmp, dim=stack_dim) if use_fp32 and (tmp.dtype.itemsize < 4) and tmp.dtype.is_floating_point: # cast to float before sum and return float, then cast back output = tmp.sum(dim=stack_dim, dtype=torch.float32) output = output.to(dtype=tensor.dtype) else: # else: just do sum in native dtype output = tmp.sum(dim=stack_dim) return output
[docs]def gather_v_wrapper( tensor: torch.Tensor, sizes: List[int], dim: int = 0, dst: int = 0, group: Optional[dist.ProcessGroup] = None, ) -> torch.Tensor: # pragma: no cover """ Implements a distributed GatherV primitive. It is based on the idea of a single global tensor which is distributed along a specified dimension into chunks of variable size. This primitive assumes such a distributed tensor and gathers all local tensors from each rank into the full global tensor valid on the specified destination rank. Parameters ---------- tensor : torch.Tensor local tensor on each rank sizes : List[int] list of the sizes of each chunk on each rank along distributed dimension, valid and set on each rank dim : int, optional dimension along which global tensor is distributed, by default 0 dst : int, optional destination rank which contains the full global tensor after the operation, by default 0 group : Optional[dist.ProcessGroup], optional process group along which global tensor is shared, by default None Returns ------- torch.Tensor full global tensor, valid on destination rank """ comm_size = dist.get_world_size(group=group) rank = dist.get_rank(group=group) if len(sizes) != comm_size: raise ValueError() if dim >= tensor.dim(): raise ValueError() if not (0 <= dst < comm_size): raise ValueError() if tensor.size(dim) != sizes[rank]: raise ValueError() if comm_size == 1: return tensor tensor_shape = list(tensor.shape) x_recv = [None] * comm_size x_send = [None] * comm_size for r in range(comm_size): if rank == dst: tensor_shape[dim] = sizes[r] else: tensor_shape[dim] = 0 x_recv[r] = torch.empty( tensor_shape, dtype=tensor.dtype, device=tensor.device, ) if r == dst: x_send[r] = tensor else: tensor_shape[dim] = 0 x_send[r] = torch.empty( tensor_shape, dtype=tensor.dtype, device=tensor.device, ) dist.all_to_all(x_recv, x_send, group=group) # TODO: clean gather/scatter and some examples up # main question is around whether e.g. gather returns # None for rank != dst or an empty dummy or an dummy # containing meta-information like dtype/etc.. if rank != dst: for r in range(comm_size): tensor_shape[dim] = sizes[r] x_recv[r] = torch.empty( tensor_shape, dtype=tensor.dtype, device=tensor.device, ) output = torch.cat(x_recv, dim=dim) return output
[docs]def scatter_v_wrapper( tensor: torch.Tensor, sizes: List[int], dim: int = 0, src: int = 0, group: Optional[dist.ProcessGroup] = None, ) -> torch.Tensor: # pragma: no cover """ Implements a distributed ScatterV primitive. It is based on the idea of a single global tensor which is distributed along a specified dimension into chunks of variable size. This primitive scatters the global tensor from a specified source rank into local chunks onto each other rank. Parameters ---------- tensor : torch.Tensor global tensor, valid on source rank sizes : List[int] list of the sizes of each chunk on each rank along distributed dimension, valid and set each rank dim : int, optional dimension along which global tensor is distributed, by default 0 src : int, optional source rank of primitive, i.e. rank of original full global tensor, by default 0 group : Optional[dist.ProcessGroup], optional process group along which global tensor is shared, by default None Returns ------- torch.Tensor corresponding local part of the global tensor on each rank """ comm_size = dist.get_world_size(group=group) rank = dist.get_rank(group=group) if len(sizes) != comm_size: raise ValueError() if dist.get_rank(group=group) == 0 and dim >= tensor.dim(): raise ValueError() if not (0 <= src < comm_size): raise ValueError() # all_to_all is already all_to_all_v, use empty tensors to "mask"-out irrelevant parts tensor_shape = list(tensor.shape) x_send = [None] * comm_size x_recv = [None] * comm_size if rank == src: scatter_list = torch.split(tensor, sizes, dim=dim) scatter_list = [t.contiguous() for t in scatter_list] x_send = scatter_list else: for r in range(comm_size): tensor_shape[dim] = 0 x_send[r] = torch.empty( tensor_shape, device=tensor.device, dtype=tensor.dtype ) for r in range(comm_size): if r == src: tensor_shape[dim] = sizes[rank] else: tensor_shape[dim] = 0 x_recv[r] = torch.empty(tensor_shape, device=tensor.device, dtype=tensor.dtype) dist.all_to_all(x_recv, x_send, group=group) return x_recv[src]
[docs]def indexed_all_to_all_v_wrapper( tensor: torch.Tensor, indices: List[torch.Tensor], sizes: List[List[int]], dim: int = 0, group: Optional[dist.ProcessGroup] = None, ) -> torch.Tensor: # pragma: no cover """ Implements an indexed version of a distributed AllToAllV primitive. It is based on the idea of a single global tensor which is distributed along a specified dimension into chunks of variable size. This primitive assumes a set of indices into this dimension which indicate the corresponding slices sent to each other rank forming an indexed version of an AllToAllV primitive. Parameters ---------- tensor : torch.Tensor local part of global tensor on each rank indices : List[torch.Tensor] list of indices on each rank of slices being sent to each other rank from this rank sizes : List[List[int]] number of indices each rank sends to each other rank, valid and set on each rank, e.g. sizes[0][3] corresponds to the number of slices rank 0 sends to rank 3 dim : int dimension along which global tensor is distributed, by default 0 group : Optional[dist.ProcessGroup], optional process group along which global tensor is shared, by default None Returns ------- torch.Tensor local result of primitive corresponding to indexed global tensor """ comm_size = dist.get_world_size(group=group) rank = dist.get_rank(group=group) if len(sizes) != comm_size: raise ValueError() if dim >= tensor.dim(): raise ValueError() if len(sizes[rank]) != comm_size: raise ValueError() if len(indices) != comm_size: raise ValueError() x_send = [tensor[idx] for idx in indices] x_recv = [None] * comm_size tensor_shape = list(tensor.shape) for r in range(comm_size): tensor_shape[dim] = sizes[r][rank] x_recv[r] = torch.empty( tensor_shape, dtype=tensor.dtype, device=tensor.device, ) dist.all_to_all(x_recv, x_send, group=group) tensor_to_recv = torch.cat(x_recv, dim=dim) return tensor_to_recv
[docs]def indexed_all_to_all_v_wrapper_bwd( tensor: torch.Tensor, indices: List[torch.Tensor], sizes: List[List[int]], tensor_size_along_dim: int, use_fp32: bool = True, dim: int = 0, group: Optional[dist.ProcessGroup] = None, ) -> torch.Tensor: # pragma: no cover """ Implements the backward pass to the indexed version of a distributed AllToAllV primitive. Parameters ---------- tensor : torch.Tensor local tensor, i.e. gradient on resulting tensor from forward pass indices : List[torch.Tensor] list of indices on each rank of slices being sent to each other rank from this rank sizes : List[List[int]] list of the sizes of each chunk on each rank along distributed dimension, valid and set on each rank tensor_size_along_dim : int size of original local tensor along specified dimension, i.e. from the corresponding forward pass use_fp32 : bool, optional flag to specify reduction taking place at least in FP32 precision, by default True only acts on floating point inputs in lower precision dim : int, optional dimension along with global tensor is distributed, by default 0 group : Optional[dist.ProcessGroup], optional process group along which global tensor is shared, by default None Returns ------- torch.Tensor result of primitive corresponding to indexed global tensor """ comm_size = dist.get_world_size(group=group) rank = dist.get_rank(group=group) if len(sizes) != comm_size: raise ValueError() if dim >= tensor.dim(): raise ValueError() if len(sizes[rank]) != comm_size: raise ValueError() if len(indices) != comm_size: raise ValueError() tensor_shape = list(tensor.shape) # scatter gradients, roles reversed compared to forward pass # recv_sizes in forward pass recv_sizes = [sizes[i][rank] for i in range(comm_size)] # send_sizes in forward pass send_sizes = [sizes[rank][i] for i in range(comm_size)] x_send = torch.split(tensor, recv_sizes, dim=dim) x_send = [t.contiguous() for t in x_send] x_recv = [None] * comm_size for r in range(comm_size): tensor_shape[dim] = send_sizes[r] x_recv[r] = torch.empty( tensor_shape, dtype=tensor.dtype, device=tensor.device, ) dist.all_to_all(x_recv, x_send, group=group) tensor_to_recv = torch.cat(x_recv, dim=dim) # sum up gathered gradients and taking # care of precision handling as specified # by boolean flag indices = torch.cat(indices, dim=0) tensor_shape[dim] = tensor_size_along_dim if use_fp32 and (tensor.dtype.itemsize < 4) and tensor.dtype.is_floating_point: out = torch.zeros( tensor_shape, dtype=torch.float32, device=tensor.device, ) tensor_to_recv = tensor_to_recv.to(dtype=torch.float32) else: out = torch.zeros( tensor_shape, dtype=tensor.dtype, device=tensor.device, ) out.index_add_(source=tensor_to_recv, index=indices, dim=dim) if out.dtype != tensor.dtype: out = out.to(tensor.dtype) return out
[docs]def mark_module_as_shared( module: nn.Module, process_group: Optional[str], recurse: bool = True, use_fp32_reduction: bool = True, ) -> nn.Module: """ Helper function to mark parameters of a module as being shared across ranks by attaching gradient hooks to the corresponding tensors. Parameters ---------- module : nn.Module PyTorch module which is to be marked as having shared parameters. process_group : str | None str indicating process_group which contains ranks across which the module's parameters are shared. If passed as None, will default to the world group. recurse : bool, default=True Flag indicating whether the module's parameters are traversed in a recursive fashion, i.e. whether sub-modules are also considered as having shared parameters. use_fp32_reduction : bool, default=True Flag indicating whether the reduction for accumulating gradients will be done in at least FP32 or the native datatype. """ group = DistributedManager().group(process_group) handle_key = "_shared_weight_dist_hook" def hook(grad: torch.Tensor) -> torch.Tensor: # the documentation states that # "The hook should not modify its argument, but it can optionally return a new gradient # which will be used in place of grad." # as all_reduce is an in-place operation, need to copy gradient grad = _reduce(grad.clone(), group=group, use_fp32=use_fp32_reduction) return grad def hook_post_accum(param: torch.Tensor) -> None: # the documentation states that # "Note that, unlike other autograd hooks, this hook operates on the tensor that requires grad # and not the grad itself. The hook can in-place modify and access its Tensor argument, # including its .grad field." param.grad = _reduce(param.grad, group=group, use_fp32=use_fp32_reduction) for name, param in module.named_parameters(recurse=recurse): error_msg = f"Parameter {name} already marked as having shared weights, can't mark it again!" if hasattr(param, handle_key): raise RuntimeError(error_msg) if torch.__version__ < (2, 1): handle = param.register_hook(hook) else: handle = param.register_post_accumulate_grad_hook(hook_post_accum) setattr(param, handle_key, handle) return module
[docs]def unmark_module_as_shared( module: nn.Module, recurse: bool = True, ) -> nn.Module: """ Helper function to unmark parameters of a module as being shared across ranks by removing attached gradient hooks. Parameters ---------- module : nn.Module PyTorch module which is to be unmarked as having shared parameters. recurse : bool, default=True Flag indicating whether the module's parameters are traversed in a recursive fashion, i.e. whether sub-modules are also considered as having shared parameters. """ handle_key = "_shared_weight_dist_hook" for name, param in module.named_parameters(recurse=recurse): error_msg = ( f"Parameter {name} NOT marked as having shared weights, can't unmark it!" ) if not hasattr(param, handle_key): raise RuntimeError(error_msg) handle = getattr(param, handle_key) handle.remove() delattr(param, handle_key) return module
© Copyright 2023, NVIDIA PhysicsNeMo Team. Last updated on Jun 11, 2025.