ArthurY's picture
update source
c3d0544
# SPDX-FileCopyrightText: Copyright (c) 2023 - 2025 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.
import logging
from dataclasses import dataclass
from typing import List, Optional
import torch
import torch.distributed as dist
from physicsnemo.distributed import (
DistributedManager,
all_gather_v,
gather_v,
indexed_all_to_all_v,
scatter_v,
)
logger = logging.getLogger(__name__)
@dataclass
class GraphPartition:
"""
Class acting as an "utility" structure to hold all relevant buffers and variables
to define a graph partition and facilitate exchange of necessary buffers for
message passing on a distributed graph.
A global graph is assumed to be defined through a global CSC structure
defining edges between source nodes and destination nodes which are assumed
to be numbered indexed by contiguous IDs. Hence, features associated to both
nodes and edges can be represented through dense feature tables globally.
When partitioning graph and features, we distribute destination nodes and all
their incoming edges on all ranks within the partition group based on a specified
mapping. Based on this scheme, there will a be a difference between
partitioned source nodes (partitioned features) and local source node
IDs which refer to the node IDs within the local graph defined by the
destination nodes on each rank. To allow message passing, communication
primitives have to ensure to gather all corresponding features for all
local source nodes based on the applied partitioning scheme. This also
leads to the distinction of local source node IDs and remote source node
IDs on each rank where the latter simply refers to the local row ID within
the dense partitioning of node features and the former indicates the source
of a message for each edge within each local graph.
Parameters
----------
partition_size : int
size of partition
partition_rank : int
local rank of this partition w.r.t. group of partitions
device : torch.device
device handle for buffers within this partition rank
"""
partition_size: int
partition_rank: int
device: torch.device
# flag to indicate using adj matrix 1-D row-decomp
matrix_decomp: bool = False
# data structures defining partition
# set in after initialization or during execution
# of desired partition scheme
# local CSR offsets defining local graph on each `partition_rank`
local_offsets: Optional[torch.Tensor] = None
# local CSR indices defining local graph on each `partition_rank`
local_indices: Optional[torch.Tensor] = None
# number of source nodes in local graph on each `partition_rank`
num_local_src_nodes: int = -1
# number of destination nodes in local graph on each `partition_rank`
num_local_dst_nodes: int = -1
# number of edges in local graph on each `partition_rank`
num_local_indices: int = -1
# mapping from global to local ID space (source node IDs)
map_partitioned_src_ids_to_global: Optional[torch.Tensor] = None
map_concatenated_local_src_ids_to_global: Optional[torch.Tensor] = None
# mapping from local to global ID space (destination node IDs)
map_partitioned_dst_ids_to_global: Optional[torch.Tensor] = None
map_concatenated_local_dst_ids_to_global: Optional[torch.Tensor] = None
# mapping from local to global ID space (edge IDs)
map_partitioned_edge_ids_to_global: Optional[torch.Tensor] = None
map_concatenated_local_edge_ids_to_global: Optional[torch.Tensor] = None
# reverse mappings
map_global_src_ids_to_concatenated_local: Optional[torch.Tensor] = None
map_global_dst_ids_to_concatenated_local: Optional[torch.Tensor] = None
map_global_edge_ids_to_concatenated_local: Optional[torch.Tensor] = None
# utility lists and sizes required for exchange of messages
# between graph partitions through distributed communication primitives
# number of IDs each rank potentially sends to all other ranks
sizes: Optional[List[List[int]]] = None
# local indices of IDs current rank sends to all other ranks
scatter_indices: Optional[List[torch.Tensor]] = None
# number of global source nodes for each `partition_rank`
num_src_nodes_in_each_partition: Optional[List[int]] = None
# number of global destination nodes for each `partition_rank`
num_dst_nodes_in_each_partition: Optional[List[int]] = None
# number of global indices for each `partition_rank`
num_indices_in_each_partition: Optional[List[int]] = None
def __post_init__(self):
# after partition_size has been set in __init__
if self.partition_size <= 0:
raise ValueError(f"Expected partition_size > 0, got {self.partition_size}")
if not (0 <= self.partition_rank < self.partition_size):
raise ValueError(
f"Expected 0 <= partition_rank < {self.partition_size}, got {self.partiton_rank}"
)
if self.sizes is None:
self.sizes = [
[None for _ in range(self.partition_size)]
for _ in range(self.partition_size)
]
if self.scatter_indices is None:
self.scatter_indices = [None] * self.partition_size
if self.num_src_nodes_in_each_partition is None:
self.num_src_nodes_in_each_partition = [None] * self.partition_size
if self.num_dst_nodes_in_each_partition is None:
self.num_dst_nodes_in_each_partition = [None] * self.partition_size
if self.num_indices_in_each_partition is None:
self.num_indices_in_each_partition = [None] * self.partition_size
def to(self, *args, **kwargs):
# move all tensors
for attr in dir(self):
attr_val = getattr(self, attr)
if isinstance(attr_val, torch.Tensor):
setattr(self, attr, attr_val.to(*args, **kwargs))
# handle scatter_indices separately
self.scatter_indices = [idx.to(*args, **kwargs) for idx in self.scatter_indices]
return self
def partition_graph_with_id_mapping(
global_offsets: torch.Tensor,
global_indices: torch.Tensor,
mapping_src_ids_to_ranks: torch.Tensor,
mapping_dst_ids_to_ranks: torch.Tensor,
partition_size: int,
partition_rank: int,
device: torch.device,
) -> GraphPartition:
"""
Utility function which partitions a global graph given as CSC structure.
It partitions both the global ID spaces for source nodes and destination nodes
based on the corresponding mappings as well as the graph structure and edge IDs.
For more details on partitioning in general see `GraphPartition`.
The function performs the partitioning based on a global graph in CPU
memory for each rank independently. It could be rewritten to e.g. only
do it one rank and exchange the partitions or to an algorithm that also
assumes an already distributed global graph, however, we expect global
graphs to fit in CPU memory. After the partitioning, we can get rid off
the larger one in CPU memory, only keep the local graphs on each GPU, and
avoid tedious gather/scatter routines for exchanging partitions in the process.
Note: It is up to the user to ensure that the provided mapping is valid. In particular,
we expect each rank to receive a non-empty partition of node IDs.
Parameters
----------
global_offsets : torch.Tensor
CSC offsets, can live on the CPU
global_indices : torch.Tensor
CSC indices, can live on the CPU
mapping_src_ids_to_ranks: torch.Tensor
maps each global ID from every source node to its partition rank
mapping_dst_ids_to_ranks: torch.Tensor
maps each global ID from every destination node to its partition rank
partition_size : int
number of process groups across which graph is partitioned,
i.e. the number of graph partitions
partition_rank : int
rank within process group managing the distributed graph, i.e.
the rank determining which partition the corresponding local rank
will manage
device : torch.device
device connected to the passed partition rank, i.e. the device
on which the local graph and related buffers will live on
"""
# initialize graph partition
graph_partition = GraphPartition(
partition_size=partition_size, partition_rank=partition_rank, device=device
)
# --------------------------------------------------------------
# initialize temporary variables used in computing the partition
# global IDs of in each partition
dst_nodes_in_each_partition = [None] * partition_size
src_nodes_in_each_partition = [None] * partition_size
num_dst_nodes_in_each_partition = [None] * partition_size
num_src_nodes_in_each_partition = [None] * partition_size
dtype = global_indices.dtype
input_device = global_indices.device
graph_partition.map_concatenated_local_src_ids_to_global = torch.empty_like(
mapping_src_ids_to_ranks
)
graph_partition.map_concatenated_local_dst_ids_to_global = torch.empty_like(
mapping_dst_ids_to_ranks
)
graph_partition.map_concatenated_local_edge_ids_to_global = torch.empty_like(
global_indices
)
graph_partition.map_global_src_ids_to_concatenated_local = torch.empty_like(
mapping_src_ids_to_ranks
)
graph_partition.map_global_dst_ids_to_concatenated_local = torch.empty_like(
mapping_dst_ids_to_ranks
)
graph_partition.map_global_edge_ids_to_concatenated_local = torch.empty_like(
global_indices
)
_map_global_src_ids_to_local = torch.empty_like(mapping_src_ids_to_ranks)
# temporarily track cum-sum of nodes per partition for "concatenated_local_ids"
_src_id_offset = 0
_dst_id_offset = 0
_edge_id_offset = 0
for rank in range(partition_size):
dst_nodes_in_each_partition[rank] = torch.nonzero(
mapping_dst_ids_to_ranks == rank
).view(-1)
src_nodes_in_each_partition[rank] = torch.nonzero(
mapping_src_ids_to_ranks == rank
).view(-1)
num_nodes = dst_nodes_in_each_partition[rank].numel()
if num_nodes == 0:
raise RuntimeError(
f"Aborting partitioning, rank {rank} has 0 destination nodes to work on."
)
num_dst_nodes_in_each_partition[rank] = num_nodes
num_nodes = src_nodes_in_each_partition[rank].numel()
num_src_nodes_in_each_partition[rank] = num_nodes
if num_nodes == 0:
raise RuntimeError(
f"Aborting partitioning, rank {rank} has 0 source nodes to work on."
)
# create mapping of global node IDs to/from "concatenated local" IDs
ids = src_nodes_in_each_partition[rank]
mapped_ids = torch.arange(
start=_src_id_offset,
end=_src_id_offset + ids.numel(),
dtype=dtype,
device=input_device,
)
_map_global_src_ids_to_local[ids] = mapped_ids - _src_id_offset
graph_partition.map_global_src_ids_to_concatenated_local[ids] = mapped_ids
graph_partition.map_concatenated_local_src_ids_to_global[mapped_ids] = ids
_src_id_offset += ids.numel()
ids = dst_nodes_in_each_partition[rank]
mapped_ids = torch.arange(
start=_dst_id_offset,
end=_dst_id_offset + ids.numel(),
dtype=dtype,
device=input_device,
)
graph_partition.map_global_dst_ids_to_concatenated_local[ids] = mapped_ids
graph_partition.map_concatenated_local_dst_ids_to_global[mapped_ids] = ids
_dst_id_offset += ids.numel()
graph_partition.num_src_nodes_in_each_partition = num_src_nodes_in_each_partition
graph_partition.num_dst_nodes_in_each_partition = num_dst_nodes_in_each_partition
# create local graph structures
for rank in range(partition_size):
offset_start = global_offsets[dst_nodes_in_each_partition[rank]].view(-1)
offset_end = global_offsets[dst_nodes_in_each_partition[rank] + 1].view(-1)
degree = offset_end - offset_start
local_offsets = degree.view(-1).cumsum(dim=0)
local_offsets = torch.cat(
[
torch.Tensor([0]).to(
dtype=dtype,
device=input_device,
),
local_offsets,
]
)
partitioned_edge_ids = torch.cat(
[
torch.arange(
start=offset_start[i],
end=offset_end[i],
dtype=dtype,
device=input_device,
)
for i in range(len(offset_start))
]
)
ids = partitioned_edge_ids
mapped_ids = torch.arange(
_edge_id_offset,
_edge_id_offset + ids.numel(),
device=ids.device,
dtype=ids.dtype,
)
graph_partition.map_global_edge_ids_to_concatenated_local[ids] = mapped_ids
graph_partition.map_concatenated_local_edge_ids_to_global[mapped_ids] = ids
_edge_id_offset += ids.numel()
partitioned_src_ids = torch.cat(
[
global_indices[offset_start[i] : offset_end[i]].clone()
for i in range(len(offset_start))
]
)
global_src_ids_on_rank, inverse_mapping = partitioned_src_ids.unique(
sorted=True, return_inverse=True
)
remote_local_src_ids_on_rank = _map_global_src_ids_to_local[
global_src_ids_on_rank
]
_map_global_src_ids_to_local_graph = torch.zeros_like(mapping_src_ids_to_ranks)
_num_local_indices = 0
for rank_offset in range(partition_size):
mask = mapping_src_ids_to_ranks[global_src_ids_on_rank] == rank_offset
if partition_rank == rank_offset:
# indices to send to this rank from this rank
graph_partition.scatter_indices[rank] = (
remote_local_src_ids_on_rank[mask]
.detach()
.clone()
.to(dtype=torch.int64)
)
numel_mask = mask.sum().item()
graph_partition.sizes[rank_offset][rank] = numel_mask
tmp_ids = torch.arange(
_num_local_indices,
_num_local_indices + numel_mask,
device=input_device,
dtype=dtype,
)
_num_local_indices += numel_mask
tmp_map = global_src_ids_on_rank[mask]
_map_global_src_ids_to_local_graph[tmp_map] = tmp_ids
local_indices = _map_global_src_ids_to_local_graph[partitioned_src_ids]
graph_partition.num_indices_in_each_partition[rank] = local_indices.size(0)
if rank == partition_rank:
# local graph
graph_partition.local_offsets = local_offsets
graph_partition.local_indices = local_indices
graph_partition.num_local_indices = graph_partition.local_indices.size(0)
graph_partition.num_local_dst_nodes = num_dst_nodes_in_each_partition[rank]
graph_partition.num_local_src_nodes = global_src_ids_on_rank.size(0)
# partition-local mappings (local IDs to global)
graph_partition.map_partitioned_src_ids_to_global = (
src_nodes_in_each_partition[rank]
)
graph_partition.map_partitioned_dst_ids_to_global = (
dst_nodes_in_each_partition[rank]
)
graph_partition.map_partitioned_edge_ids_to_global = partitioned_edge_ids
for r in range(graph_partition.partition_size):
err_msg = "error in graph partition: list containing sizes of exchanged indices does not match the tensor of indices to be exchanged"
if (
graph_partition.sizes[graph_partition.partition_rank][r]
!= graph_partition.scatter_indices[r].numel()
):
raise AssertionError(err_msg)
graph_partition = graph_partition.to(device=device)
return graph_partition
def partition_graph_with_matrix_decomposition(
global_offsets: torch.Tensor,
global_indices: torch.Tensor,
num_nodes: int,
partition_book: torch.Tensor,
partition_size: int,
partition_rank: int,
device: torch.device,
) -> GraphPartition:
"""
Utility function which partitions a global graph given as CSC structure based on its adjacency
matirx using 1-D row-wise decomposition. This approach ensures a 1D uniform distribution of nodes
and their associated 1-hop incoming edges. By treating source and destination nodes equivalently
during partitioning, this approach assumes the graph is not bipartite.
This decomposition also ensures that the graph convolution (spMM) remains local by maintaining a copy of
the local incoming edge features and the local node outputs from the graph convolution.
The memory complexity of this approach is O[(N/P + E/P)*hid_dim*L], where N/E are the number of nodes/edges.
The transformation from local node storage to local edge storage is achieved using nccl `alltoall`.
Key differences from the existing graph partition scheme (partition_graph_with_id_mapping):
(1) This function partitions the global node ID space uniformly, without distinguishing
between source and destination nodes (i.e., matrix row ordering or column ordering). Both
src/dst or row/col nodes are indexed consistently within the adjacency matrix.
(2) Each local graph (sub-matrix) can be defined/constructed by just node/edge offsets from
global graph.
(3) The partitioning is performed on a global graph stored in CPU memory, and then each device
(rank) constructs its local graph independently from the global csc matrix.
Parameters
----------
global_offsets : torch.Tensor
CSC offsets, can live on the CPU
global_indices : torch.Tensor
CSC indices, can live on the CPU
num_nodes : int
number of nodes in the global graph
partition_book : torch.Tensor
the boundaries of 1-D row-decomp of adj. matrix for all ranks
partition_size : int
number of process groups across which graph is partitioned,
i.e. the number of graph partitions
partition_rank : int
rank within process group managing the distributed graph, i.e.
the rank determining which partition the corresponding local rank
will manage
device : torch.device
device connected to the passed partition rank, i.e. the device
on which the local graph and related buffers will live on
"""
# initialize graph partition
graph_partition = GraphPartition(
partition_size=partition_size, partition_rank=partition_rank, device=device
)
dtype = global_indices.dtype
# --------------------------------------------------------------
# First partition the global row ptrs (dst nodes) to local row ptrs
num_edges = global_indices.size(0)
node_offset = partition_book[partition_rank]
num_local_nodes = (
partition_book[partition_rank + 1] - partition_book[partition_rank]
)
edge_partition_offset = global_offsets[node_offset]
if node_offset + num_local_nodes > num_nodes:
raise ValueError("Invalid node offset and number of local nodes")
local_offsets = global_offsets[node_offset : node_offset + num_local_nodes + 1].to(
device=device, non_blocking=True
)
graph_partition.local_offsets = local_offsets - edge_partition_offset
graph_partition.num_local_dst_nodes = num_local_nodes
# Scan through all partitions and compress the source nodes (edges) for each partition
# to fill the local send/recv buffers for all-to-all communications
partition_book = partition_book.to(device=device)
for to_partition in range(partition_size):
local_indices = global_indices[
global_offsets[partition_book[to_partition]] : global_offsets[
partition_book[to_partition + 1]
]
].to(device=device, non_blocking=True)
# compress the columns (src nodes or local_indices) for each partition and record mapping (inverse_indices)
global_src_node_at_partition, inverse_indices = local_indices.unique(
sorted=True, return_inverse=True
)
global_src_node_at_partition_rank = (
torch.bucketize(global_src_node_at_partition, partition_book, right=True)
- 1
)
src_node_indices = torch.nonzero(
global_src_node_at_partition_rank == partition_rank, as_tuple=False
).squeeze(1)
# fill local send buffer for alltoalls (scatter selected nodes to_partition rank)
graph_partition.scatter_indices[to_partition] = (
global_src_node_at_partition[src_node_indices] - node_offset
)
# fill the numbers of indices (edges), dst nodes and src nodes for each partition
graph_partition.num_indices_in_each_partition[to_partition] = (
local_indices.size(0)
)
graph_partition.num_dst_nodes_in_each_partition[to_partition] = (
partition_book[to_partition + 1] - partition_book[to_partition]
)
graph_partition.num_src_nodes_in_each_partition[to_partition] = (
global_src_node_at_partition.size(0)
)
if to_partition == partition_rank:
graph_partition.local_indices = inverse_indices
graph_partition.num_local_indices = graph_partition.local_indices.size(0)
graph_partition.num_local_src_nodes = global_src_node_at_partition.size(0)
# map from local (compressed) column indices [0, ..., num_local_src_nodes] to their global node IDs
graph_partition.map_partitioned_src_ids_to_global = (
global_src_node_at_partition
)
for from_partition in range(partition_size):
# fill all recv buffer sizes for alltoalls
graph_partition.sizes[from_partition][to_partition] = torch.count_nonzero(
global_src_node_at_partition_rank == from_partition
)
# trivial mappings due to 1D row-wise decomposition
graph_partition.map_partitioned_dst_ids_to_global = torch.arange(
node_offset, node_offset + num_local_nodes, dtype=dtype, device=device
)
graph_partition.map_partitioned_edge_ids_to_global = torch.arange(
edge_partition_offset,
edge_partition_offset + graph_partition.num_local_indices,
dtype=dtype,
device=device,
)
# trivial mappings due to 1D row-wise decomposition, with mem. cost O(E, N) at each dev; need to optimize
graph_partition.map_concatenated_local_src_ids_to_global = torch.arange(
num_nodes, dtype=dtype, device=device
)
graph_partition.map_concatenated_local_edge_ids_to_global = torch.arange(
num_edges, dtype=dtype, device=device
)
graph_partition.map_concatenated_local_dst_ids_to_global = (
graph_partition.map_concatenated_local_src_ids_to_global
)
graph_partition.map_global_src_ids_to_concatenated_local = (
graph_partition.map_concatenated_local_src_ids_to_global
)
graph_partition.map_global_dst_ids_to_concatenated_local = (
graph_partition.map_concatenated_local_src_ids_to_global
)
graph_partition.map_global_edge_ids_to_concatenated_local = (
graph_partition.map_concatenated_local_edge_ids_to_global
)
graph_partition.matrix_decomp = True
for r in range(graph_partition.partition_size):
err_msg = "error in graph partition: list containing sizes of exchanged indices does not match the tensor of indices to be exchanged"
if (
graph_partition.sizes[graph_partition.partition_rank][r]
!= graph_partition.scatter_indices[r].numel()
):
raise AssertionError(err_msg)
graph_partition = graph_partition.to(device=device)
return graph_partition
def partition_graph_nodewise(
global_offsets: torch.Tensor,
global_indices: torch.Tensor,
partition_size: int,
partition_rank: int,
device: torch.device,
matrix_decomp: bool = False,
) -> GraphPartition:
"""
Utility function which partitions a global graph given as CSC structure naively
by splitting both the IDs of source and destination nodes into chunks of equal
size. For more details on partitioning in general see `GraphPartition`.
The function performs the partitioning based on a global graph in CPU
memory for each rank independently. It could be rewritten to e.g. only
do it one rank and exchange the partitions or to an algorithm that also
assumes an already distributed global graph, however, we expect global
graphs to fit in CPU memory. After the partitioning, we can get rid off
the larger one in CPU memory, only keep the local graphs on each GPU, and
avoid tedious gather/scatter routines for exchanging partitions in the process.
Parameters
----------
global_offsets : torch.Tensor
CSC offsets, can live on the CPU
global_indices : torch.Tensor
CSC indices, can live on the CPU
partition_size : int
number of process groups across which graph is partitioned,
i.e. the number of graph partitions
partition_rank : int
rank within process group managing the distributed graph, i.e.
the rank determining which partition the corresponding local rank
will manage
device : torch.device
device connected to the passed partition rank, i.e. the device
on which the local graph and related buffers will live on
matrix_decomp : bool
flag to enable matrix decomposition for partitioning
"""
num_global_src_nodes = global_indices.max().item() + 1
num_global_dst_nodes = global_offsets.size(0) - 1
num_dst_nodes_per_partition = (
num_global_dst_nodes + partition_size - 1
) // partition_size
if matrix_decomp:
if num_global_src_nodes != num_global_dst_nodes:
raise ValueError(
"Must be square adj. matrix (num_src=num_dst) for matrix decomposition"
)
partition_book = torch.arange(
0,
num_global_dst_nodes,
num_dst_nodes_per_partition,
dtype=global_indices.dtype,
)
partition_book = torch.cat(
[
partition_book,
torch.tensor([num_global_dst_nodes], dtype=global_indices.dtype),
]
)
return partition_graph_with_matrix_decomposition(
global_offsets,
global_indices,
num_global_dst_nodes,
partition_book,
partition_size,
partition_rank,
device,
)
num_src_nodes_per_partition = (
num_global_src_nodes + partition_size - 1
) // partition_size
mapping_dst_ids_to_ranks = (
torch.arange(
num_global_dst_nodes,
dtype=global_offsets.dtype,
device=global_offsets.device,
)
// num_dst_nodes_per_partition
)
mapping_src_ids_to_ranks = (
torch.arange(
num_global_src_nodes,
dtype=global_offsets.dtype,
device=global_offsets.device,
)
// num_src_nodes_per_partition
)
return partition_graph_with_id_mapping(
global_offsets,
global_indices,
mapping_src_ids_to_ranks,
mapping_dst_ids_to_ranks,
partition_size,
partition_rank,
device,
)
def partition_graph_by_coordinate_bbox(
global_offsets: torch.Tensor,
global_indices: torch.Tensor,
src_coordinates: torch.Tensor,
dst_coordinates: torch.Tensor,
coordinate_separators_min: List[List[Optional[float]]],
coordinate_separators_max: List[List[Optional[float]]],
partition_size: int,
partition_rank: int,
device: torch.device,
) -> GraphPartition:
"""
Utility function which partitions a global graph given as CSC structure.
It partitions both the global ID spaces for source nodes and destination nodes
based on their corresponding coordinates. Each partition will manage points which
fulfill the boxconstraints specified by the specified coordinate separators. For each
rank one is expected to specify the minimum and maximum coordinate value for each dimension.
A partition the will manage all points for which ``min_val <= coord[d] < max_val`` holds. If one
of the constraints is passed as `None`, it is assumed to be non-binding and the partition is defined
by the corresponding half-space. Each rank maintains both a partition of the global source and
destination nodes resulting from this subspace division.
The function performs the partitioning based on a global graph in CPU
memory for each rank independently. It could be rewritten to e.g. only
do it one rank and exchange the partitions or to an algorithm that also
assumes an already distributed global graph, however, we expect global
graphs to fit in CPU memory. After the partitioning, we can get rid off
the larger one in CPU memory, only keep the local graphs on each GPU, and
avoid tedious gather/scatter routines for exchanging partitions in the process.
Note: It is up to the user to ensure that the provided partition is valid.
In particular, we expect each rank to receive a non-empty partition of node IDs.
Examples
--------
>>> import torch
>>> from physicsnemo.models.gnn_layers import partition_graph_by_coordinate_bbox
>>> # simple graph with a degree of 2 per node
>>> num_src_nodes = 8
>>> num_dst_nodes = 4
>>> offsets = torch.arange(num_dst_nodes + 1, dtype=torch.int64) * 2
>>> indices = torch.arange(num_src_nodes, dtype=torch.int64)
>>> # example with 2D coordinates
>>> # assuming partitioning a 2D problem into the 4 quadrants
>>> partition_size = 4
>>> partition_rank = 0
>>> coordinate_separators_min = [[0, 0], [None, 0], [None, None], [0, None]]
>>> coordinate_separators_max = [[None, None], [0, None], [0, 0], [None, 0]]
>>> device = "cuda:0"
>>> # dummy coordinates
>>> src_coordinates = torch.FloatTensor(
... [
... [-1.0, 1.0],
... [1.0, 1.0],
... [-1.0, -1.0],
... [1.0, -1.0],
... [-2.0, 2.0],
... [2.0, 2.0],
... [-2.0, -2.0],
... [2.0, -2.0],
... ]
... )
>>> dst_coordinates = torch.FloatTensor(
... [
... [-1.0, 1.0],
... [1.0, 1.0],
... [-1.0, -1.0],
... [1.0, -1.0],
... ]
... )
>>> # call partitioning routine
>>> pg = partition_graph_by_coordinate_bbox(
... offsets,
... indices,
... src_coordinates,
... dst_coordinates,
... coordinate_separators_min,
... coordinate_separators_max,
... partition_size,
... partition_rank,
... device,
... )
>>> pg.local_offsets
tensor([0, 2], device='cuda:0')
>>> pg.local_indices
tensor([0, 1], device='cuda:0')
>>> pg.sizes
[[0, 1, 1, 0], [0, 1, 1, 0], [1, 0, 0, 1], [1, 0, 0, 1]]
>>>
>>> # example with lat-long coordinates
>>> # dummy coordinates
>>> src_lat = torch.FloatTensor([-75, -60, -45, -30, 30, 45, 60, 75]).view(-1, 1)
>>> dst_lat = torch.FloatTensor([-60, -30, 30, 30]).view(-1, 1)
>>> src_long = torch.FloatTensor([-135, -135, 135, 135, -45, -45, 45, 45]).view(-1, 1)
>>> dst_long = torch.FloatTensor([-135, 135, -45, 45]).view(-1, 1)
>>> src_coordinates = torch.cat([src_lat, src_long], dim=1)
>>> dst_coordinates = torch.cat([dst_lat, dst_long], dim=1)
>>> # separate sphere at equator and 0 degree longitude into 4 parts
>>> coordinate_separators_min = [
... [-90, -180],
... [-90, 0],
... [0, -180],
... [0, 0],
... ]
>>> coordinate_separators_max = [
... [0, 0],
... [0, 180],
... [90, 0],
... [90, 180],
... ]
>>> # call partitioning routine
>>> partition_size = 4
>>> partition_rank = 0
>>> device = "cuda:0"
>>> pg = partition_graph_by_coordinate_bbox(
... offsets,
... indices,
... src_coordinates,
... dst_coordinates,
... coordinate_separators_min,
... coordinate_separators_max,
... partition_size,
... partition_rank,
... device,
... )
>>> pg.local_offsets
tensor([0, 2], device='cuda:0')
>>> pg.local_indices
tensor([0, 1], device='cuda:0')
>>> pg.sizes
[[2, 0, 0, 0], [0, 2, 0, 0], [0, 0, 2, 0], [0, 0, 0, 2]]
Parameters
----------
global_offsets : torch.Tensor
CSC offsets, can live on the CPU
global_indices : torch.Tensor
CSC indices, can live on the CPU
src_coordinates : torch.Tensor
coordinates of each source node
dst_coordinates : torch.Tensor
coordinates of each destination node
partition_size : int
number of process groups across which graph is partitioned,
i.e. the number of graph partitions
partition_rank : int
rank within process group managing the distributed graph, i.e.
the rank determining which partition the corresponding local rank
will manage
device : torch.device
device connected to the passed partition rank, i.e. the device
on which the local graph and related buffers will live on
"""
dim = src_coordinates.size(-1)
if dst_coordinates.size(-1) != dim:
raise ValueError()
if len(coordinate_separators_min) != partition_size:
a, b = len(coordinate_separators_min), partition_size
error_msg = "Expected len(coordinate_separators_min) == partition_size"
error_msg += f", but got {a} and {b} respectively"
raise ValueError(error_msg)
if len(coordinate_separators_max) != partition_size:
a, b = len(coordinate_separators_max), partition_size
error_msg = "Expected len(coordinate_separators_max) == partition_size"
error_msg += f", but got {a} and {b} respectively"
raise ValueError(error_msg)
for rank in range(partition_size):
if len(coordinate_separators_min[rank]) != dim:
a, b = len(coordinate_separators_min[rank]), dim
error_msg = f"Expected len(coordinate_separators_min[{rank}]) == dim"
error_msg += f", but got {a} and {b} respectively"
if len(coordinate_separators_max[rank]) != dim:
a, b = len(coordinate_separators_max[rank]), dim
error_msg = f"Expected len(coordinate_separators_max[{rank}]) == dim"
error_msg += f", but got {a} and {b} respectively"
num_global_src_nodes = global_indices.max().item() + 1
num_global_dst_nodes = global_offsets.size(0) - 1
mapping_dst_ids_to_ranks = torch.zeros(
num_global_dst_nodes, dtype=global_offsets.dtype, device=global_offsets.device
)
mapping_src_ids_to_ranks = torch.zeros(
num_global_src_nodes,
dtype=global_offsets.dtype,
device=global_offsets.device,
)
def _assign_ranks(mapping, coordinates):
for p in range(partition_size):
mask = torch.ones_like(mapping).to(dtype=torch.bool)
for d in range(dim):
min_val, max_val = (
coordinate_separators_min[p][d],
coordinate_separators_max[p][d],
)
if min_val is not None:
mask = mask & (coordinates[:, d] >= min_val)
if max_val is not None:
mask = mask & (coordinates[:, d] < max_val)
mapping[mask] = p
_assign_ranks(mapping_src_ids_to_ranks, src_coordinates)
_assign_ranks(mapping_dst_ids_to_ranks, dst_coordinates)
return partition_graph_with_id_mapping(
global_offsets,
global_indices,
mapping_src_ids_to_ranks,
mapping_dst_ids_to_ranks,
partition_size,
partition_rank,
device,
)
class DistributedGraph:
def __init__(
self,
global_offsets: torch.Tensor,
global_indices: torch.Tensor,
partition_size: int,
graph_partition_group_name: str = None,
graph_partition: Optional[GraphPartition] = None,
): # pragma: no cover
"""
Utility Class representing a distributed graph based on a given
partitioning of a CSC graph structure. By default, a naive node-wise
partitioning scheme is applied, see ``partition_graph_nodewise`` for
details on that. This class then wraps necessary communication primitives
to access all relevant feature buffers related to the graph.
Parameters
----------
global_offsets : torch.Tensor
CSC offsets, can live on the CPU
global_indices : torch.Tensor
CSC indices, can live on the CPU
partition_size : int
Number of process groups across which graphs are distributed, expected to
be larger than 1, i.e. an actual partition distributed among multiple ranks.
partition_group_name : str, default=None
Name of process group across which graphs are distributed. Passing no process
group name leads to a parallelism across the default process group.
Otherwise, the group size of a process group is expected to match partition_size.
graph_partition : GraphPartition, optional
Optional graph_partition, if passed as None, the naive
node-wise partitioning scheme will be applied to global_offsets and global_indices,
otherwise, these will be ignored and the passed partition will be used instead.
"""
dist_manager = DistributedManager()
self.device = dist_manager.device
self.partition_rank = dist_manager.group_rank(name=graph_partition_group_name)
self.partition_size = dist_manager.group_size(name=graph_partition_group_name)
error_msg = f"Passed partition_size does not correspond to size of process_group, got {partition_size} and {self.partition_size} respectively."
if self.partition_size != partition_size:
raise AssertionError(error_msg)
self.process_group = dist_manager.group(name=graph_partition_group_name)
if graph_partition is None:
# default partitioning scheme
self.graph_partition = partition_graph_nodewise(
global_offsets,
global_indices,
self.partition_size,
self.partition_rank,
self.device,
)
else:
error_msg = f"Passed graph_partition.partition_size does not correspond to size of process_group, got {graph_partition.partition_size} and {self.partition_size} respectively."
if graph_partition.partition_size != self.partition_size:
raise AssertionError(error_msg)
error_msg = f"Passed graph_partition.device does not correspond to device of this rank, got {graph_partition.device} and {self.device} respectively."
if graph_partition.device != self.device:
raise AssertionError(error_msg)
self.graph_partition = graph_partition
send_sizes = self.graph_partition.sizes[self.graph_partition.partition_rank]
recv_sizes = [
p[self.graph_partition.partition_rank] for p in self.graph_partition.sizes
]
msg = f"GraphPartition(rank={self.graph_partition.partition_rank}, "
msg += f"num_local_src_nodes={self.graph_partition.num_local_src_nodes}, "
msg += f"num_local_dst_nodes={self.graph_partition.num_local_dst_nodes}, "
msg += f"num_partitioned_src_nodes={self.graph_partition.num_src_nodes_in_each_partition[self.graph_partition.partition_rank]}, "
msg += f"num_partitioned_dst_nodes={self.graph_partition.num_dst_nodes_in_each_partition[self.graph_partition.partition_rank]}, "
msg += f"send_sizes={send_sizes}, recv_sizes={recv_sizes})"
print(msg)
dist.barrier(self.process_group)
def get_src_node_features_in_partition(
self,
global_node_features: torch.Tensor,
scatter_features: bool = False,
src_rank: int = 0,
) -> torch.Tensor: # pragma: no cover
# if global features only on local rank 0 also scatter, split them
# according to the partition and scatter them to other ranks
if self.graph_partition.matrix_decomp:
logger.warning(
"Matrix decomposition assumes one type of node feature partition, and the graph"
"adjacency matrix is square with identical src/dst node domains. "
"So, only `get_dst_node_features_in_partition` is used/needed to get src or dst"
"node features within a partition."
)
return self.get_dst_node_features_in_partition(
global_node_features,
scatter_features=scatter_features,
src_rank=src_rank,
)
if scatter_features:
global_node_features = global_node_features[
self.graph_partition.map_concatenated_local_src_ids_to_global
]
return scatter_v(
global_node_features,
self.graph_partition.num_src_nodes_in_each_partition,
dim=0,
src=src_rank,
group=self.process_group,
)
return global_node_features.to(device=self.device)[
self.graph_partition.map_partitioned_src_ids_to_global, :
]
def get_src_node_features_in_local_graph(
self, partitioned_src_node_features: torch.Tensor
) -> torch.Tensor: # pragma: no cover
# main primitive to gather all necessary src features
# which are required for a csc-based message passing step
return indexed_all_to_all_v(
partitioned_src_node_features,
indices=self.graph_partition.scatter_indices,
sizes=self.graph_partition.sizes,
use_fp32=True,
dim=0,
group=self.process_group,
)
def get_dst_node_features_in_partition(
self,
global_node_features: torch.Tensor,
scatter_features: bool = False,
src_rank: int = 0,
) -> torch.Tensor: # pragma: no cover
# if global features only on local rank 0 also scatter, split them
# according to the partition and scatter them to other ranks
if scatter_features:
global_node_features = global_node_features.to(device=self.device)[
self.graph_partition.map_concatenated_local_dst_ids_to_global
]
return scatter_v(
global_node_features,
self.graph_partition.num_dst_nodes_in_each_partition,
dim=0,
src=src_rank,
group=self.process_group,
)
return global_node_features.to(device=self.device)[
self.graph_partition.map_partitioned_dst_ids_to_global, :
]
def get_dst_node_features_in_local_graph(
self,
partitioned_dst_node_features: torch.Tensor,
) -> torch.Tensor: # pragma: no cover
# current partitioning scheme assumes that
# local graph is built from partitioned IDs
return partitioned_dst_node_features
def get_edge_features_in_partition(
self,
global_edge_features: torch.Tensor,
scatter_features: bool = False,
src_rank: int = 0,
) -> torch.Tensor: # pragma: no cover
# if global features only on local rank 0 also scatter, split them
# according to the partition and scatter them to other ranks
if scatter_features:
global_edge_features = global_edge_features[
self.graph_partition.map_concatenated_local_edge_ids_to_global
]
return scatter_v(
global_edge_features,
self.graph_partition.num_indices_in_each_partition,
dim=0,
src=src_rank,
group=self.process_group,
)
return global_edge_features.to(device=self.device)[
self.graph_partition.map_partitioned_edge_ids_to_global, :
]
def get_edge_features_in_local_graph(
self, partitioned_edge_features: torch.Tensor
) -> torch.Tensor: # pragma: no cover
# current partitioning scheme assumes that
# local graph is built from partitioned IDs
return partitioned_edge_features
def get_global_src_node_features(
self,
partitioned_node_features: torch.Tensor,
get_on_all_ranks: bool = True,
dst_rank: int = 0,
) -> torch.Tensor: # pragma: no cover
error_msg = f"Passed partitioned_node_features.device does not correspond to device of this rank, got {partitioned_node_features.device} and {self.device} respectively."
if partitioned_node_features.device != self.device:
raise AssertionError(error_msg)
if self.graph_partition.matrix_decomp:
logger.warning(
"Matrix decomposition assumes one type of node feature partition, and the graph"
"adjacency matrix is square with identical src/dst node domains. "
"So, only `get_global_dst_node_features` is used/needed to get global src or dst"
"node features."
)
return self.get_global_dst_node_features(
partitioned_node_features,
get_on_all_ranks=get_on_all_ranks,
dst_rank=dst_rank,
)
if not get_on_all_ranks:
global_node_feat = gather_v(
partitioned_node_features,
self.graph_partition.num_src_nodes_in_each_partition,
dim=0,
dst=dst_rank,
group=self.process_group,
)
if self.graph_partition.partition_rank == dst_rank:
global_node_feat = global_node_feat[
self.graph_partition.map_global_src_ids_to_concatenated_local
]
return global_node_feat
global_node_feat = all_gather_v(
partitioned_node_features,
self.graph_partition.num_src_nodes_in_each_partition,
dim=0,
use_fp32=True,
group=self.process_group,
)
global_node_feat = global_node_feat[
self.graph_partition.map_global_src_ids_to_concatenated_local
]
return global_node_feat
def get_global_dst_node_features(
self,
partitioned_node_features: torch.Tensor,
get_on_all_ranks: bool = True,
dst_rank: int = 0,
) -> torch.Tensor: # pragma: no cover
error_msg = f"Passed partitioned_node_features.device does not correspond to device of this rank, got {partitioned_node_features.device} and {self.device} respectively."
if partitioned_node_features.device != self.device:
raise AssertionError(error_msg)
if not get_on_all_ranks:
global_node_feat = gather_v(
partitioned_node_features,
self.graph_partition.num_dst_nodes_in_each_partition,
dim=0,
dst=dst_rank,
group=self.process_group,
)
if self.graph_partition.partition_rank == dst_rank:
global_node_feat = global_node_feat[
self.graph_partition.map_global_dst_ids_to_concatenated_local
]
return global_node_feat
global_node_feat = all_gather_v(
partitioned_node_features,
self.graph_partition.num_dst_nodes_in_each_partition,
dim=0,
use_fp32=True,
group=self.process_group,
)
global_node_feat = global_node_feat[
self.graph_partition.map_global_dst_ids_to_concatenated_local
]
return global_node_feat
def get_global_edge_features(
self,
partitioned_edge_features: torch.Tensor,
get_on_all_ranks: bool = True,
dst_rank: int = 0,
) -> torch.Tensor: # pragma: no cover
error_msg = f"Passed partitioned_edge_features.device does not correspond to device of this rank, got {partitioned_edge_features.device} and {self.device} respectively."
if partitioned_edge_features.device != self.device:
raise AssertionError(error_msg)
if not get_on_all_ranks:
global_edge_feat = gather_v(
partitioned_edge_features,
self.graph_partition.num_indices_in_each_partition,
dim=0,
dst=dst_rank,
group=self.process_group,
)
if self.graph_partition.partition_rank == dst_rank:
global_edge_feat = global_edge_feat[
self.graph_partition.map_global_edge_ids_to_concatenated_local
]
return global_edge_feat
global_edge_feat = all_gather_v(
partitioned_edge_features,
self.graph_partition.num_indices_in_each_partition,
dim=0,
use_fp32=True,
group=self.process_group,
)
global_edge_feat = global_edge_feat[
self.graph_partition.map_global_edge_ids_to_concatenated_local
]
return global_edge_feat