biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Core] Support multi-node inference(eager and cuda graph) (#3686)

Browse Source
This commit is contained in:
Roy
2024-03-29 06:01:55 +08:00
committed by GitHub
parent a4075cba4d
commit 515386ef3c
7 changed files with 25 additions and 22 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
tests/distributed/test_comm_ops.py
View File

@@ -24,7 +24,7 @@ def all_reduce_test_worker(tensor_parallel_size: int, rank: int,
del os.environ["CUDA_VISIBLE_DEVICES"]
device = torch.device(f"cuda:{rank}")
torch.cuda.set_device(device)
init_test_distributed_environment(1, tensor_parallel_size, rank,
init_test_distributed_environment(1, tensor_parallel_size, rank, rank,
distributed_init_port)
num_elements = 8
all_tensors = [
@@ -46,7 +46,7 @@ def all_gather_test_worker(tensor_parallel_size: int, rank: int,
del os.environ["CUDA_VISIBLE_DEVICES"]
device = torch.device(f"cuda:{rank}")
torch.cuda.set_device(device)
init_test_distributed_environment(1, tensor_parallel_size, rank,
init_test_distributed_environment(1, tensor_parallel_size, rank, rank,
distributed_init_port)
num_dimensions = 3
tensor_size = list(range(2, num_dimensions + 2))
@@ -74,7 +74,7 @@ def broadcast_tensor_dict_test_worker(tensor_parallel_size: int, rank: int,
del os.environ["CUDA_VISIBLE_DEVICES"]
device = torch.device(f"cuda:{rank}")
torch.cuda.set_device(device)
init_test_distributed_environment(1, tensor_parallel_size, rank,
init_test_distributed_environment(1, tensor_parallel_size, rank, rank,
distributed_init_port)
test_dict = {
"a": torch.arange(8, dtype=torch.float32, device="cuda"),

4
tests/distributed/test_custom_all_reduce.py
View File

@@ -23,7 +23,7 @@ def graph_allreduce(world_size, rank, distributed_init_port):
del os.environ["CUDA_VISIBLE_DEVICES"]
device = torch.device(f"cuda:{rank}")
torch.cuda.set_device(device)
init_test_distributed_environment(1, world_size, rank,
init_test_distributed_environment(1, world_size, rank, rank,
distributed_init_port)
custom_ar.init_custom_ar()
@@ -58,7 +58,7 @@ def eager_allreduce(world_size, rank, distributed_init_port):
del os.environ["CUDA_VISIBLE_DEVICES"]
device = torch.device(f"cuda:{rank}")
torch.cuda.set_device(device)
init_test_distributed_environment(1, world_size, rank,
init_test_distributed_environment(1, world_size, rank, rank,
distributed_init_port)
sz = 1024