[core] set up data parallel communication (#13591)

Signed-off-by: youkaichao <youkaichao@gmail.com>

vllm/distributed/utils.py

@@ -11,7 +11,11 @@ from collections import deque
 from typing import Any, Deque, Dict, Optional, Sequence, Tuple
 
 import torch
-from torch.distributed import TCPStore
+from torch.distributed import ProcessGroup, TCPStore
+from torch.distributed.distributed_c10d import (Backend, PrefixStore,
+                                                _get_default_timeout,
+                                                is_nccl_available)
+from torch.distributed.rendezvous import rendezvous
 
 import vllm.envs as envs
 from vllm.logger import init_logger
@@ -227,3 +231,88 @@ class StatelessProcessGroup:
             world_size=world_size,
             store=store,
             data_expiration_seconds=data_expiration_seconds)
+
+
+def stateless_init_torch_distributed_process_group(
+        host: str, port: int, rank: int, world_size: int,
+        backend: str) -> ProcessGroup:
+    """
+    A replacement for `torch.distributed.init_process_group` that does not
+    pollute the global state. The created ProcessGroup object can be used for
+    some operations such as `allreduce`, because it does not depend on the
+    global rank. However, some operations such as `broadcast` cannot be used
+    because it depends on the global rank.
+
+    # TODO: ask for help from PyTorch team if we need the `broadcast` operation.
+
+    This function is useful when we are not sure about the total number of
+    processes in the process group. For example, we may have process
+    1, 2, ..., 8 who want to communicate, and process 9 might be the same
+    process as process 1, or it might be a different process; process 10
+    might be the same process as process 5, or it might be a different process.
+    In this case, how can we reliably form a communication channel within
+    process 9 and 10, without affecting the communication channel within
+    process 1, 2, ..., 8?
+
+    One possible solution is to figure out if process 9 and 10 are the same
+    as process 1 and 5 beforehand, and then form a communication channel
+    based on the information, adjusting the ranks and world_size etc. However,
+    figuring out the information is not always easy, and it will interfere
+    with the main communication channel.
+
+    Our solution is to always form a communication channel with process 1, 2,
+    ..., 8, and then use this function to form another communication channel
+    with process 9 and 10. This way, regardless of whether process 9 and 10
+    are the same as process 1 and 5, the main communication channel is
+    always formed with process 1, 2, ..., 8, and the additional communication
+    channel is formed with process 9 and 10.
+    """
+    init_method = f"tcp://{host}:{port}"
+    backend = Backend(backend)  # it is basically string
+    timeout = _get_default_timeout(backend)
+
+    store, rank, world_size = next(
+        rendezvous(init_method, rank, world_size, timeout=timeout))
+    store.set_timeout(timeout)
+
+    group_rank = rank
+    group_size = world_size
+
+    # Use a PrefixStore to avoid accidental overrides of keys used by
+    # different systems (e.g. RPC) in case the store is multi-tenant.
+    prefix_store = PrefixStore(init_method, store)
+
+    pg_options = ProcessGroup.Options(backend=backend, timeout=timeout)
+
+    pg: ProcessGroup = ProcessGroup(
+        prefix_store,
+        group_rank,
+        group_size,
+        pg_options,
+    )
+
+    if backend == "gloo":
+        from torch.distributed.distributed_c10d import ProcessGroupGloo
+        backend_class = ProcessGroupGloo(prefix_store,
+                                         group_rank,
+                                         group_size,
+                                         timeout=timeout)
+        backend_type = ProcessGroup.BackendType.GLOO
+        device = torch.device("cpu")
+    elif backend == "nccl":
+        assert is_nccl_available()
+        from torch.distributed.distributed_c10d import ProcessGroupNCCL
+
+        backend_options = ProcessGroupNCCL.Options()
+        backend_options._timeout = timeout
+
+        backend_class = ProcessGroupNCCL(prefix_store, group_rank, group_size,
+                                         backend_options)
+        backend_type = ProcessGroup.BackendType.NCCL
+        device = torch.device("cuda")
+
+    backend_class._set_sequence_number_for_group()
+
+    pg._register_backend(device, backend_type, backend_class)
+
+    return pg