frozenleaves/vllm
1
0
Fork 0
mirror of https://github.com/vllm-project/vllm.git synced 2025-10-20 14:53:52 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Core][Test] fix function name typo in custom allreduce (#4750)

Browse Source
This commit is contained in:
youkaichao
2024-05-10 15:14:40 -07:00
committed by GitHub
parent fcc2994be6
commit 4e12131089
2 changed files with 6 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
tests/distributed/test_custom_all_reduce.py
View File

@ -25,7 +25,7 @@ def graph_allreduce(world_size, rank, distributed_init_port):
init_test_distributed_environment(1, world_size, rank,
distributed_init_port)
custom_all_reduce.init_custom_all_reduce()
custom_all_reduce.init_custom_ar()
for sz in test_sizes:
for dtype in [torch.float32, torch.float16, torch.bfloat16]:
with custom_all_reduce.capture():
@ -61,7 +61,7 @@ def eager_allreduce(world_size, rank, distributed_init_port):
distributed_init_port)
sz = 1024
custom_all_reduce.init_custom_all_reduce()
custom_all_reduce.init_custom_ar()
fa = custom_all_reduce.get_handle()
inp = torch.ones(sz, dtype=torch.float32, device=device)
out = fa.all_reduce_unreg(inp)

4
vllm/distributed/device_communicators/custom_all_reduce.py
View File

@ -52,6 +52,10 @@ def init_custom_ar() -> None:
"current process. This might be the case if 'CUDA_VISIBLE_DEVICES'"
" is set.")
return
# we only use a subset of GPUs here
# so we only need to check the nvlink connectivity of these GPUs
num_dev = world_size
# test nvlink first, this will filter out most of the cases
# where custom allreduce is not supported
cuda_visible_devices = envs.CUDA_VISIBLE_DEVICES