xoscar 0.9.0__cp312-cp312-macosx_10_13_x86_64.whl

xoscar/collective/process_group.py

@@ -0,0 +1,687 @@
+# Copyright 2022-2023 XProbe Inc.
+#
+# 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 os
+from abc import ABC, abstractmethod
+from typing import Any, List, Optional
+
+from ..utils import is_linux, lazy_import
+from . import xoscar_pygloo as xp
+from .common import (
+    RENDEZVOUS_MASTER_IP_ENV_KEY,
+    RENDEZVOUS_MASTER_PORT_ENV_KEY,
+    AllReduceAlgorithm,
+    AllReduceAlgorithmMappingGloo,
+    CollectiveReduceOp,
+    ReduceOpMappingGloo,
+    TypeMappingGloo,
+)
+from .utils import convert_data_to_cp_array, convert_data_to_np_array
+
+cupy = lazy_import("cupy")
+if cupy is not None:
+    from .backend.nccl_backend import TCPStore, XoscarNCCLBackend
+    from .common import ReduceOpMappingNCCL, ReduceOpMappingNCCLStr, TypeMappingNCCL
+
+
+class _World:
+    def __init__(self):
+        self._store = None
+        self._device = None
+        self._backend = None
+        self._nccl_size = 0
+
+    @property
+    def store(self):
+        return self._store
+
+    @property
+    def device(self):
+        return self._device
+
+    @store.setter  # type: ignore
+    def store(self, store):
+        self._store = store
+
+    @device.setter  # type: ignore
+    def device(self, device):
+        self._device = device
+
+
+_world = _World()
+
+
+class ProcessGroup(ABC):
+    class Options:
+        master_ip: Optional[str] = None
+        master_port: Optional[int] = None
+
+    def __init__(
+        self,
+        rank: int,
+        world_size: int,
+        group_name: Optional[str] = None,
+        pg_options: Optional[Options] = None,
+    ):
+        self._rank = rank
+        self._world_size = world_size
+        self._group_name = group_name
+        self._option = pg_options
+
+    @property
+    def rank(self):
+        return self._rank
+
+    @property
+    def world_size(self):
+        return self._world_size
+
+    @property
+    def name(self):
+        return self._group_name
+
+    @property
+    def options(self):
+        return self._option
+
+    @abstractmethod
+    def allreduce(self, *args, **kwargs):
+        """All reduce function"""
+
+    @abstractmethod
+    def reduce(self, *args, **kwargs):
+        """Reduce function"""
+
+    @abstractmethod
+    def allgather(self, *args, **kwargs):
+        """All gather function"""
+
+    @abstractmethod
+    def gather(self, *args, **kwargs):
+        """Gather function"""
+
+    @abstractmethod
+    def scatter(self, *args, **kwargs):
+        """Scatter function"""
+
+    @abstractmethod
+    def reduce_scatter(self, *args, **kwargs):
+        """Reduce scatter function"""
+
+    @abstractmethod
+    def alltoall(self, *args, **kwargs):
+        """All to all function"""
+
+    @abstractmethod
+    def broadcast(self, *args, **kwargs):
+        """Broadcast function"""
+
+
+class ProcessGroupGloo(ProcessGroup):
+    def __init__(
+        self,
+        ip: str,
+        rank: int,
+        world_size: int,
+        group_name: Optional[str] = None,
+        pg_options: Optional[ProcessGroup.Options] = None,
+    ):
+        super().__init__(rank, world_size, group_name, pg_options)
+        if _world.store is None:
+            master_ip = (
+                pg_options.master_ip
+                if pg_options is not None
+                else os.environ.get(RENDEZVOUS_MASTER_IP_ENV_KEY, None)
+            )
+            master_port = (
+                pg_options.master_port
+                if pg_options is not None
+                else os.environ.get(RENDEZVOUS_MASTER_PORT_ENV_KEY, None)
+            )
+            if master_ip is None or master_port is None:
+                raise ValueError("Cannot find master ip or port for rendezvous")
+
+            opt = xp.rendezvous.TCPStoreOptions()
+            opt.port = int(master_port)
+            opt.numWorkers = world_size
+            opt.isServer = rank == 0
+
+            store = xp.rendezvous.TCPStore(master_ip, opt)
+            if not is_linux():
+                attr = xp.transport.uv.attr(ip)  # type: ignore
+                dev = xp.transport.uv.CreateDevice(attr)  # type: ignore
+            else:
+                attr = xp.transport.tcp.attr(ip)
+                dev = xp.transport.tcp.CreateDevice(attr)  # type: ignore
+            _world.store = store  # type: ignore
+            _world.device = dev  # type: ignore
+        else:
+            store = _world.store
+            dev = _world.device
+
+        prefix_store = xp.rendezvous.PrefixStore(group_name or str(world_size), store)  # type: ignore
+        context = xp.rendezvous.Context(rank, world_size)
+        context.connectFullMesh(prefix_store, dev)
+        self._context = context
+
+    def reduce(
+        self,
+        send_data: Any,
+        recv_data: Any,
+        op: CollectiveReduceOp = CollectiveReduceOp.SUM,
+        root: Optional[int] = 0,
+        tag: Optional[int] = 0,
+    ):
+        send_buf = convert_data_to_np_array(send_data)
+        recv_buf = convert_data_to_np_array(recv_data)
+        size = send_buf.size
+        dtype = send_buf.dtype
+        sendptr = send_buf.ctypes.data
+        recvptr = recv_buf.ctypes.data
+        gloo_type = TypeMappingGloo[dtype.type]
+        xp.reduce(
+            self._context,
+            sendptr,
+            recvptr,
+            size,
+            gloo_type,
+            ReduceOpMappingGloo[op],
+            root,
+            tag,
+        )
+
+    def allreduce(
+        self,
+        send_data: Any,
+        recv_data: Any,
+        op: CollectiveReduceOp = CollectiveReduceOp.SUM,
+        algorithm: AllReduceAlgorithm = AllReduceAlgorithm.RING,
+        tag: Optional[int] = 0,
+    ):
+        send_buf = convert_data_to_np_array(send_data)
+        recv_buf = convert_data_to_np_array(recv_data)
+        size = send_buf.size
+        dtype = send_buf.dtype
+        sendptr = send_buf.ctypes.data
+        recvptr = recv_buf.ctypes.data
+        gloo_type = TypeMappingGloo[dtype.type]
+        xp.allreduce(
+            self._context,
+            sendptr,
+            recvptr,
+            size,
+            gloo_type,
+            ReduceOpMappingGloo[op],
+            AllReduceAlgorithmMappingGloo[algorithm],
+            tag,  # type: ignore
+        )
+
+    def gather(
+        self,
+        send_data: Any,
+        recv_data: Any,
+        root: Optional[int] = 0,
+        tag: Optional[int] = 0,
+    ):
+        send_buf = convert_data_to_np_array(send_data)
+        recv_buf = convert_data_to_np_array(recv_data)
+        size = send_buf.size
+        dtype = send_buf.dtype
+        sendptr = send_buf.ctypes.data
+        recvptr = recv_buf.ctypes.data
+        gloo_type = TypeMappingGloo[dtype.type]
+        xp.gather(self._context, sendptr, recvptr, size, gloo_type, root, tag)
+
+    def allgather(self, send_data: Any, recv_data: Any, tag: Optional[int] = 0):
+        send_buf = convert_data_to_np_array(send_data)
+        recv_buf = convert_data_to_np_array(recv_data)
+        size = send_buf.size
+        dtype = send_buf.dtype
+        sendptr = send_buf.ctypes.data
+        recvptr = recv_buf.ctypes.data
+        gloo_type = TypeMappingGloo[dtype.type]
+        xp.allgather(self._context, sendptr, recvptr, size, gloo_type, tag)
+
+    def scatter(
+        self,
+        send_data: List[Any],
+        recv_data: Any,
+        root: Optional[int] = 0,
+        tag: Optional[int] = 0,
+    ):
+        send_bufs = [convert_data_to_np_array(d) for d in send_data]
+        recv_buf = convert_data_to_np_array(recv_data)
+        size = sum([d.size for d in send_bufs])
+        dtype = recv_buf.dtype
+        sendptrs = [d.ctypes.data for d in send_bufs]
+        recvptr = recv_buf.ctypes.data
+        gloo_type = TypeMappingGloo[dtype.type]
+        xp.scatter(self._context, sendptrs, recvptr, size, gloo_type, root, tag)  # type: ignore
+
+    def reduce_scatter(
+        self,
+        send_data: Any,
+        recv_data: Any,
+        recv_elems: List[int],
+        op: CollectiveReduceOp = CollectiveReduceOp.SUM,
+    ):  # pragma: no cover
+        send_buf = convert_data_to_np_array(send_data)
+        recv_buf = convert_data_to_np_array(recv_data)
+        sendptr = send_buf.ctypes.data
+        recvptr = recv_buf.ctypes.data
+        size = send_buf.size
+        dtype = send_buf.dtype
+        gloo_type = TypeMappingGloo[dtype.type]
+        xp.reduce_scatter(
+            self._context,
+            sendptr,
+            recvptr,
+            size,
+            recv_elems,
+            gloo_type,
+            ReduceOpMappingGloo[op],
+        )
+
+    def alltoall(self, send_data: Any, recv_data: Any, tag: Optional[int] = 0):
+        send_buf = convert_data_to_np_array(send_data)
+        recv_buf = convert_data_to_np_array(recv_data)
+        size = send_buf.size
+        dtype = send_buf.dtype
+        sendptr = send_buf.ctypes.data
+        recvptr = recv_buf.ctypes.data
+        gloo_type = TypeMappingGloo[dtype.type]
+        xp.all_to_all(self._context, sendptr, recvptr, size, gloo_type, tag)
+
+    def broadcast(
+        self,
+        send_data: Any,
+        recv_data: Any,
+        root: Optional[int] = 0,
+        tag: Optional[int] = 0,
+    ):
+        if send_data is not None:
+            send_buf = convert_data_to_np_array(send_data)
+            sendptr = send_buf.ctypes.data
+        else:
+            sendptr = None
+        recv_buf = convert_data_to_np_array(recv_data)
+        size = recv_buf.size
+        dtype = recv_buf.dtype
+        recvptr = recv_buf.ctypes.data
+        gloo_type = TypeMappingGloo[dtype.type]
+        xp.broadcast(
+            self._context,
+            recvptr if sendptr is None else sendptr,
+            recvptr,
+            size,
+            gloo_type,
+            root,
+            tag,
+        )
+
+
+class ProcessGroupNCCL(ProcessGroup):
+    def __init__(
+        self,
+        ip: str,
+        rank: int,
+        device_id: int,
+        world_size: int,
+        group_name: Optional[str] = None,
+        pg_options: Optional[ProcessGroup.Options] = None,
+    ):
+        assert (
+            cupy != None
+        ), "cupy is required when creating a group using nccl as backend."
+        from cupy.cuda import nccl
+
+        super().__init__(rank, world_size, group_name, pg_options)
+        cupy.cuda.Device(device_id).use()
+        if _world._backend is None:
+            master_ip = (
+                pg_options.master_ip
+                if pg_options is not None
+                else os.environ.get(RENDEZVOUS_MASTER_IP_ENV_KEY, None)
+            )
+            master_port = (
+                pg_options.master_port
+                if pg_options is not None
+                else os.environ.get(RENDEZVOUS_MASTER_PORT_ENV_KEY, None)
+            )
+            if master_ip is None or master_port is None:
+                raise ValueError("Cannot find master ip or port for rendezvous")
+            store = TCPStore(world_size)
+            backend = XoscarNCCLBackend(
+                world_size, rank, store, master_ip, int(master_port)
+            )
+            _world._backend = backend
+            _world._nccl_size = world_size
+            self._is_world = True
+            self._backend = backend
+        else:
+            self._is_world = False
+            if rank == 0:
+                commId = nccl.get_unique_id()
+                ccid = cupy.array(commId)
+                for i in range(1, world_size):
+                    _world._backend.send(ccid, i, None)
+            else:
+                commId = (int(i) for i in range(128))
+                commId = tuple(commId)
+                ccid = cupy.array(commId, dtype="int64")
+                _world._backend.recv(ccid, 0, None)
+                commId = tuple(ccid.tolist())
+            self._backend = nccl.NcclCommunicator(world_size, commId, rank)
+
+    def reduce(
+        self,
+        send_buf: Any,
+        recv_buf: Any,
+        op: CollectiveReduceOp = CollectiveReduceOp.SUM,
+        root: Optional[int] = 0,
+        stream: Optional[Any] = None,
+    ):
+        send_buf = convert_data_to_cp_array(send_buf)
+        recv_buf = convert_data_to_cp_array(recv_buf)
+        dtype = send_buf.dtype
+        stream = (
+            stream
+            if stream is not None and isinstance(stream, cupy.cuda.Stream)
+            else cupy.cuda.Stream.null
+        )
+        if self._is_world:
+            self._backend.reduce(
+                send_buf, recv_buf, root, ReduceOpMappingNCCLStr[op], stream
+            )
+        else:
+            self._backend.reduce(
+                send_buf.data.ptr,
+                recv_buf.data.ptr,
+                send_buf.size,
+                TypeMappingNCCL[dtype.type],
+                ReduceOpMappingNCCL[op],
+                root,
+                stream.ptr,
+            )
+
+    def allreduce(
+        self,
+        send_buf: Any,
+        recv_buf: Any,
+        op: CollectiveReduceOp = CollectiveReduceOp.SUM,
+        stream: Optional[Any] = None,
+    ):
+        send_buf = convert_data_to_cp_array(send_buf)
+        recv_buf = convert_data_to_cp_array(recv_buf)
+        dtype = send_buf.dtype
+        stream = (
+            stream
+            if stream is not None and isinstance(stream, cupy.cuda.Stream)
+            else cupy.cuda.Stream.null
+        )
+        if self._is_world:
+            self._backend.all_reduce(
+                send_buf, recv_buf, ReduceOpMappingNCCLStr[op], stream
+            )
+        else:
+            self._backend.allReduce(
+                send_buf.data.ptr,
+                recv_buf.data.ptr,
+                send_buf.size,
+                TypeMappingNCCL[dtype.type],
+                ReduceOpMappingNCCL[op],
+                stream.ptr,
+            )
+
+    def gather(
+        self,
+        send_buf: Any,
+        recv_buf: Any,
+        root: Optional[int] = 0,
+        stream: Optional[Any] = None,
+    ):
+        assert (
+            send_buf.size * self.world_size == recv_buf.size
+        ), "Send_size * world_number must be equal to recv_size"
+        send_buf = convert_data_to_cp_array(send_buf)
+        recv_buf = convert_data_to_cp_array(recv_buf)
+        dtype = send_buf.dtype
+        stream = (
+            stream
+            if stream is not None and isinstance(stream, cupy.cuda.Stream)
+            else cupy.cuda.Stream.null
+        )
+        if self._is_world:
+            self._backend.gather(send_buf, recv_buf, root, stream)
+        else:
+            if self.rank == root:
+                cupy.cuda.nccl.groupStart()
+                for peer in range(self.world_size):
+                    if peer == self.rank:
+                        recv_buf[peer : peer + 1] = send_buf.reshape(1, -1)
+                    else:
+                        self._backend.recv(
+                            recv_buf[peer : peer + 1].data.ptr,
+                            recv_buf[peer : peer + 1].size,
+                            TypeMappingNCCL[dtype.type],
+                            peer,
+                            stream.ptr,
+                        )
+                cupy.cuda.nccl.groupEnd()
+            else:
+                send_buf = send_buf.reshape(1, -1)
+                self._backend.send(
+                    send_buf.data.ptr,
+                    send_buf.size,
+                    TypeMappingNCCL[dtype.type],
+                    root,
+                    stream.ptr,
+                )
+
+    def allgather(
+        self,
+        send_buf: Any,
+        recv_buf: Any,
+        stream: Optional[Any] = None,
+    ):
+        send_buf = convert_data_to_cp_array(send_buf)
+        recv_buf = convert_data_to_cp_array(recv_buf)
+        stream = (
+            stream
+            if stream is not None and isinstance(stream, cupy.cuda.Stream)
+            else cupy.cuda.Stream.null
+        )
+        dtype = send_buf.dtype
+        if self._is_world:
+            self._backend.all_gather(send_buf, recv_buf, send_buf.size, stream)
+        else:
+            self._backend.allGather(
+                send_buf.data.ptr,
+                recv_buf.data.ptr,
+                send_buf.size,
+                TypeMappingNCCL[dtype.type],
+                stream.ptr,
+            )
+
+    def scatter(
+        self,
+        send_buf: List[Any],
+        recv_buf: Any,
+        root: Optional[int] = 0,
+        stream: Optional[Any] = None,
+    ):
+        send_buf = [convert_data_to_cp_array(d) for d in send_buf]
+        recv_buf = convert_data_to_cp_array(recv_buf)
+        stream = (
+            stream
+            if stream is not None and isinstance(stream, cupy.cuda.Stream)
+            else cupy.cuda.Stream.null
+        )
+        if self._is_world:
+            send_buf = cupy.concatenate(send_buf).reshape(self.world_size, -1)
+            self._backend.scatter(send_buf, recv_buf, root, stream)
+        else:
+            if self.rank == root:
+                assert (
+                    len(send_buf) == self.world_size
+                ), "Scatter size must be equal to the size of group"
+                cupy.cuda.nccl.groupStart()
+                for peer in range(self.world_size):
+                    send_data = send_buf[peer]
+                    if peer == root:
+                        recv_buf[:] = send_data
+                    else:
+                        dtype = send_data.dtype
+                        self._backend.send(
+                            send_data.data.ptr,
+                            send_data.size,
+                            TypeMappingNCCL[dtype.type],
+                            peer,
+                            stream.ptr,
+                        )
+                cupy.cuda.nccl.groupEnd()
+            else:
+                dtype = recv_buf.dtype
+                self._backend.recv(
+                    recv_buf.data.ptr,
+                    recv_buf.size,
+                    TypeMappingNCCL[dtype.type],
+                    root,
+                    stream.ptr,
+                )
+
+    def reduce_scatter(
+        self,
+        send_buf: Any,
+        recv_buf: Any,
+        recv_elems: List[int],
+        op: CollectiveReduceOp = CollectiveReduceOp.SUM,
+        stream: Optional[Any] = None,
+    ):
+        send_buf = convert_data_to_cp_array(send_buf)
+        recv_buf = convert_data_to_cp_array(recv_buf)
+        dtype = send_buf.dtype
+        stream = (
+            stream
+            if stream is not None and isinstance(stream, cupy.cuda.Stream)
+            else cupy.cuda.Stream.null
+        )
+        if self._is_world:
+            self._backend.reduce_scatter(
+                send_buf,
+                recv_buf,
+                recv_elems[self.rank],
+                ReduceOpMappingNCCLStr[op],
+                stream,
+            )
+        else:
+            self._backend.reduceScatter(
+                send_buf.data.ptr,
+                recv_buf.data.ptr,
+                recv_elems[self.rank],
+                TypeMappingNCCL[dtype.type],
+                ReduceOpMappingNCCL[op],
+                stream.ptr,
+            )
+
+    def alltoall(
+        self,
+        send_buf: Any,
+        recv_buf: Any,
+        stream: Optional[Any] = None,
+    ):
+        assert (
+            self.world_size == send_buf.shape[0]
+        ), "The first dim of send data must be equal to world size."
+        assert (
+            recv_buf.size == send_buf.size
+        ), "The size of send data must be equal to the size of recv data."
+        send_buf = convert_data_to_cp_array(send_buf)
+        recv_buf = convert_data_to_cp_array(recv_buf)
+        dtype = send_buf.dtype
+        stream = (
+            stream
+            if stream is not None and isinstance(stream, cupy.cuda.Stream)
+            else cupy.cuda.Stream.null
+        )
+        if self._is_world:
+            self._backend.all_to_all(send_buf, recv_buf, stream)
+        else:
+            cupy.cuda.nccl.groupStart()
+            for peer in range(self.world_size):
+                if peer == self.rank:
+                    recv_buf[peer : peer + 1] = send_buf[peer : peer + 1]
+                else:
+                    if self.rank > peer:
+                        self._backend.recv(
+                            recv_buf[peer : peer + 1].data.ptr,
+                            recv_buf[peer : peer + 1].size,
+                            TypeMappingNCCL[dtype.type],
+                            peer,
+                            stream.ptr,
+                        )
+                        self._backend.send(
+                            send_buf[peer : peer + 1].data.ptr,
+                            send_buf[peer : peer + 1].size,
+                            TypeMappingNCCL[dtype.type],
+                            peer,
+                            stream.ptr,
+                        )
+                    else:
+                        self._backend.send(
+                            send_buf[peer : peer + 1].data.ptr,
+                            send_buf[peer : peer + 1].size,
+                            TypeMappingNCCL[dtype.type],
+                            peer,
+                            stream.ptr,
+                        )
+                        self._backend.recv(
+                            recv_buf[peer : peer + 1].data.ptr,
+                            recv_buf[peer : peer + 1].size,
+                            TypeMappingNCCL[dtype.type],
+                            peer,
+                            stream.ptr,
+                        )
+            cupy.cuda.nccl.groupEnd()
+
+    def broadcast(
+        self,
+        send_buf: Any,
+        recv_buf: Any,
+        root: Optional[int] = 0,
+        stream: Optional[Any] = None,
+    ):
+        send_buf = convert_data_to_cp_array(send_buf)
+        recv_buf = convert_data_to_cp_array(recv_buf)
+        dtype = send_buf.dtype
+        stream = (
+            stream
+            if stream is not None and isinstance(stream, cupy.cuda.Stream)
+            else cupy.cuda.Stream.null
+        )
+        if self._is_world:
+            if self._rank == root:
+                self._backend.broadcast(send_buf, root, stream)
+                if recv_buf is not None and (recv_buf != send_buf).any():
+                    recv_buf[:] = send_buf
+            else:
+                self._backend.broadcast(recv_buf, root, stream)
+        else:
+            self._backend.broadcast(
+                send_buf.data.ptr,
+                recv_buf.data.ptr,
+                send_buf.size,
+                TypeMappingNCCL[dtype.type],
+                root,
+                stream.ptr,
+            )