[distributed] add function to create ipc buffers directly (#10064)

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

.buildkite/test-pipeline.yaml

@@ -510,6 +510,7 @@ steps:
   # NOTE: don't test llama model here, it seems hf implementation is buggy
   # see https://github.com/vllm-project/vllm/pull/5689 for details
   - pytest -v -s distributed/test_custom_all_reduce.py
+  - torchrun --nproc_per_node=2 distributed/test_ca_buffer_sharing.py
   - TARGET_TEST_SUITE=A100 pytest basic_correctness/ -v -s -m distributed_2_gpus
   - pytest -v -s -x lora/test_mixtral.py
 

tests/distributed/test_ca_buffer_sharing.py

@@ -0,0 +1,59 @@
+# can only run on machines with p2p access across GPUs
+# can only run with torchrun:
+# torchrun --nproc_per_node=2 tests/distributed/test_ca_buffer_sharing.py
+
+import ctypes
+
+import torch
+import torch.distributed as dist
+
+from vllm.distributed.device_communicators.cuda_wrapper import CudaRTLibrary
+from vllm.distributed.device_communicators.custom_all_reduce import (  # noqa
+    CustomAllreduce)
+
+# create a cpu process group for communicating metadata (ipc handle)
+dist.init_process_group(backend="gloo")
+rank = local_rank = dist.get_rank()
+world_size = dist.get_world_size()
+
+# every process sets its own device (differently)
+lib = CudaRTLibrary()
+lib.cudaSetDevice(rank)
+
+buffer_size_in_bytes = 1024
+byte_value = 2  # the value we write to the buffer for verification
+
+pointers = CustomAllreduce.create_shared_buffer(buffer_size_in_bytes)
+
+print(f"Rank {rank} has pointers {pointers}")
+
+dist.barrier()
+torch.cuda.synchronize()
+
+if rank == 0:
+    # the first rank tries to write to all buffers
+    for p in pointers:
+        pointer = ctypes.c_void_p(p)
+        lib.cudaMemset(pointer, byte_value, buffer_size_in_bytes)
+
+dist.barrier()
+torch.cuda.synchronize()
+
+host_data = (ctypes.c_char * buffer_size_in_bytes)()
+
+# all ranks read from all buffers, and check if the data is correct
+for p in pointers:
+    pointer = ctypes.c_void_p(p)
+    lib.cudaMemcpy(host_data, pointer, buffer_size_in_bytes)
+    for i in range(buffer_size_in_bytes):
+        assert ord(host_data[i]) == byte_value, (
+            f"Rank {rank} failed"
+            f" to verify buffer {p}. Expected {byte_value}, "
+            f"got {ord(host_data[i])}")
+
+print(f"Rank {rank} verified all buffers")
+
+dist.barrier()
+torch.cuda.synchronize()
+
+CustomAllreduce.free_shared_buffer(pointers)

vllm/distributed/device_communicators/custom_all_reduce.py

@@ -1,3 +1,4 @@
+import ctypes
 from contextlib import contextmanager
 from typing import Any, List, Optional, Union
 
@@ -7,6 +8,7 @@ from torch.distributed import ProcessGroup
 
 import vllm.envs as envs
 from vllm import _custom_ops as ops
+from vllm.distributed.device_communicators.cuda_wrapper import CudaRTLibrary
 from vllm.distributed.device_communicators.custom_all_reduce_utils import (
     gpu_p2p_access_check)
 from vllm.distributed.parallel_state import in_the_same_node_as
@@ -174,6 +176,35 @@ class CustomAllreduce:
                                        offsets, rank, self.full_nvlink)
         self.register_buffer(self.buffer)
 
+    @staticmethod
+    def create_shared_buffer(
+            size_in_bytes: int,
+            group: Optional[ProcessGroup] = None) -> List[int]:
+        lib = CudaRTLibrary()
+        pointer = lib.cudaMalloc(size_in_bytes)
+        handle = lib.cudaIpcGetMemHandle(pointer)
+        world_size = dist.get_world_size(group=group)
+        rank = dist.get_rank(group=group)
+        handles = [None] * world_size
+        dist.all_gather_object(handles, handle, group=group)
+
+        pointers: List[int] = []
+        for i, h in enumerate(handles):
+            if i == rank:
+                pointers.append(pointer.value)  # type: ignore
+            else:
+                pointers.append(
+                    lib.cudaIpcOpenMemHandle(h).value)  # type: ignore
+
+        return pointers
+
+    @staticmethod
+    def free_shared_buffer(pointers: List[int],
+                           group: Optional[ProcessGroup] = None) -> None:
+        rank = dist.get_rank(group=group)
+        lib = CudaRTLibrary()
+        lib.cudaFree(ctypes.c_void_p(pointers[rank]))
+
     @contextmanager
     def capture(self):
         """