A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 17, 03:00 UTC
[Core] Support multi-node inference(eager and cuda graph) (#3686)
Roy committed
Mar 29, 2024 at 06:01 UTC
515386ef3cacb44a2bcfab9d66eaee6143d94e95
7 files changed
+25
-22
tests/distributed/test_comm_ops.py
+3
-3
@@ -24,7 +24,7 @@ def all_reduce_test_worker(tensor_parallel_size: int, rank: int,
24
del os.environ["CUDA_VISIBLE_DEVICES"]
25
device = torch.device(f"cuda:{rank}")
26
torch.cuda.set_device(device)
27
- init_test_distributed_environment(1, tensor_parallel_size, rank,
27
+ init_test_distributed_environment(1, tensor_parallel_size, rank, rank,
28
distributed_init_port)
29
num_elements = 8
30
all_tensors = [
@@ -46,7 +46,7 @@ def all_gather_test_worker(tensor_parallel_size: int, rank: int,
46
del os.environ["CUDA_VISIBLE_DEVICES"]
47
device = torch.device(f"cuda:{rank}")
48
torch.cuda.set_device(device)
49
- init_test_distributed_environment(1, tensor_parallel_size, rank,
49
+ init_test_distributed_environment(1, tensor_parallel_size, rank, rank,
50
distributed_init_port)
51
num_dimensions = 3
52
tensor_size = list(range(2, num_dimensions + 2))
@@ -74,7 +74,7 @@ def broadcast_tensor_dict_test_worker(tensor_parallel_size: int, rank: int,
74
del os.environ["CUDA_VISIBLE_DEVICES"]
75
device = torch.device(f"cuda:{rank}")
76
torch.cuda.set_device(device)
77
- init_test_distributed_environment(1, tensor_parallel_size, rank,
77
+ init_test_distributed_environment(1, tensor_parallel_size, rank, rank,
78
distributed_init_port)
79
test_dict = {
80
"a": torch.arange(8, dtype=torch.float32, device="cuda"),
tests/distributed/test_custom_all_reduce.py
+2
-2
@@ -23,7 +23,7 @@ def graph_allreduce(world_size, rank, distributed_init_port):
23
del os.environ["CUDA_VISIBLE_DEVICES"]
24
device = torch.device(f"cuda:{rank}")
25
torch.cuda.set_device(device)
26
- init_test_distributed_environment(1, world_size, rank,
26
+ init_test_distributed_environment(1, world_size, rank, rank,
27
distributed_init_port)
28
29
custom_ar.init_custom_ar()
@@ -58,7 +58,7 @@ def eager_allreduce(world_size, rank, distributed_init_port):
58
del os.environ["CUDA_VISIBLE_DEVICES"]
59
device = torch.device(f"cuda:{rank}")
60
torch.cuda.set_device(device)
61
- init_test_distributed_environment(1, world_size, rank,
61
+ init_test_distributed_environment(1, world_size, rank, rank,
62
distributed_init_port)
63
64
sz = 1024
vllm/executor/ray_gpu_executor.py
-2
@@ -188,8 +188,6 @@ class RayGPUExecutor(ExecutorBase):
188
is_driver_worker=True,
189
)
190
191
- # FIXME(woosuk): We are not properly initializing pynccl when
192
- # we have multiple nodes.
191
self._run_workers("init_device")
192
self._run_workers(
193
"load_model",
vllm/model_executor/parallel_utils/pynccl.py
+8
-10
@@ -202,6 +202,7 @@ class NCCLCommunicator:
202
init_method=None,
203
timeout=datetime.timedelta(seconds=10),
204
world_size: int = -1,
205
+ local_rank: int = -1,
206
rank: int = -1,
207
store=None,
208
group_name: str = "",
@@ -219,25 +220,22 @@ class NCCLCommunicator:
220
store=store,
221
group_name=group_name,
222
pg_options=pg_options)
222
- self.world_size = dist.get_world_size()
223
- self.rank = dist.get_rank()
224
- torch.cuda.set_device(self.rank)
225
- if self.rank == 0:
223
+ torch.cuda.set_device(local_rank)
224
+ if rank == 0:
225
self.unique_id = ncclGetUniqueId()
226
else:
227
self.unique_id = NcclUniqueId()
229
- tensor = torch.ByteTensor(list(self.unique_id.internal)).cuda(
230
- self.rank)
228
+ tensor = torch.ByteTensor(list(
229
+ self.unique_id.internal)).cuda(local_rank)
230
dist.broadcast(tensor, src=0)
231
byte_list = tensor.cpu().tolist()
233
- self.unique_id = NcclUniqueId()
232
for i, byte in enumerate(byte_list):
233
self.unique_id.internal[i] = byte
234
self.comm = ctypes.c_void_p()
237
- result = _c_ncclCommInitRank(ctypes.byref(self.comm), self.world_size,
238
- self.unique_id, self.rank)
235
+ result = _c_ncclCommInitRank(ctypes.byref(self.comm), world_size,
236
+ self.unique_id, rank)
237
assert result == 0
240
- self.stream = torch.cuda.Stream(device=f"cuda:{self.rank}")
238
+ self.stream = torch.cuda.Stream(device=f"cuda:{local_rank}")
239
240
def all_reduce(self,
241
tensor: torch.Tensor,
vllm/model_executor/parallel_utils/pynccl_utils.py
+3
-1
@@ -36,11 +36,13 @@ def set_pynccl_stream(stream: torch.cuda.Stream):
36
pass
37
38
39
-def init_process_group(world_size: int, rank: int, init_method: str) -> None:
39
+def init_process_group(world_size: int, local_rank: int, rank: int,
40
+ init_method: str) -> None:
41
assert not is_initialized()
42
global comm
43
comm = NCCLCommunicator(init_method=init_method,
44
world_size=world_size,
45
+ local_rank=local_rank,
46
rank=rank)
47
48
vllm/test_utils.py
+5
-1
@@ -8,6 +8,7 @@ from vllm.worker.worker import init_distributed_environment
8
def init_test_distributed_environment(
9
pipeline_parallel_size: int,
10
tensor_parallel_size: int,
11
+ local_rank: int,
12
rank: int,
13
distributed_init_port: str,
14
) -> None:
@@ -16,7 +17,10 @@ def init_test_distributed_environment(
17
worker_use_ray=True)
18
distributed_init_method = f"tcp://localhost:{distributed_init_port}"
19
init_distributed_environment(
19
- parallel_config, rank, distributed_init_method=distributed_init_method)
20
+ parallel_config,
21
+ local_rank,
22
+ rank,
23
+ distributed_init_method=distributed_init_method)
24
25
26
def multi_process_tensor_parallel(
vllm/worker/worker.py
+4
-3
@@ -97,8 +97,8 @@ class Worker:
97
raise RuntimeError(
98
f"Not support device type: {self.device_config.device}")
99
# Initialize the distributed environment.
100
- init_distributed_environment(self.parallel_config, self.rank,
101
- self.distributed_init_method)
100
+ init_distributed_environment(self.parallel_config, self.local_rank,
101
+ self.rank, self.distributed_init_method)
102
# Set random seed.
103
set_random_seed(self.model_config.seed)
104
@@ -249,6 +249,7 @@ class Worker:
249
250
def init_distributed_environment(
251
parallel_config: ParallelConfig,
252
+ local_rank: int,
253
rank: int,
254
distributed_init_method: Optional[str] = None,
255
) -> None:
@@ -282,9 +283,9 @@ def init_distributed_environment(
283
elif parallel_config.world_size > 1:
284
# NOTE(woosuk): We don't initialize pynccl process group when world size
285
# is 1.
285
- # TODO(woosuk): Support multi-node connection.
286
pynccl_utils.init_process_group(
287
world_size=parallel_config.world_size,
288
+ local_rank=local_rank,
289
rank=rank,
290
init_method=distributed_init_method,
291
)