xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-18 04:15:01 +08:00
Code Issues Packages Projects Releases Wiki Activity

[BugFix] Pop instead of del CUDA_VISIBLE_DEVICES (#12962)

Browse Source 
Signed-off-by: Hollow Man <hollowman@opensuse.org>
This commit is contained in:
ℍ𝕠𝕝𝕝𝕠𝕨 𝕄𝕒𝕟 2025-02-11 18:21:50 +02:00 committed by GitHub
parent 21f5d50fa5
commit 6c4dbe23eb
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
4 changed files with 9 additions and 9 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
examples/offline_inference/rlhf.py
View File

@ -92,7 +92,7 @@ class MyLLM(LLM):
# a hack to make the script work. # a hack to make the script work.
# stop ray from manipulating CUDA_VISIBLE_DEVICES # stop ray from manipulating CUDA_VISIBLE_DEVICES
# at the top-level # at the top-level
del os.environ["CUDA_VISIBLE_DEVICES"] os.environ.pop("CUDA_VISIBLE_DEVICES", None)
super().__init__(*args, **kwargs) super().__init__(*args, **kwargs)

2
examples/offline_inference/rlhf_colocate.py
View File

@ -59,7 +59,7 @@ class MyLLM(LLM):
# a hack to make the script work. # a hack to make the script work.
# stop ray from manipulating CUDA_VISIBLE_DEVICES # stop ray from manipulating CUDA_VISIBLE_DEVICES
# at the top-level # at the top-level
del os.environ["CUDA_VISIBLE_DEVICES"] os.environ.pop("CUDA_VISIBLE_DEVICES", None)
# every worker will use 0.4 GPU, so that we can schedule # every worker will use 0.4 GPU, so that we can schedule
# 2 instances on the same GPUs. # 2 instances on the same GPUs.
os.environ["VLLM_RAY_PER_WORKER_GPUS"] = "0.4" os.environ["VLLM_RAY_PER_WORKER_GPUS"] = "0.4"

10
tests/distributed/test_comm_ops.py
View File

@ -22,7 +22,7 @@ def all_reduce_test_worker(tp_size: int, pp_size: int, rank: int,
# it is important to delete the CUDA_VISIBLE_DEVICES environment variable # it is important to delete the CUDA_VISIBLE_DEVICES environment variable
# so that each worker can see all the GPUs # so that each worker can see all the GPUs
# they will be able to set the device to the correct GPU # they will be able to set the device to the correct GPU
del os.environ["CUDA_VISIBLE_DEVICES"] os.environ.pop("CUDA_VISIBLE_DEVICES", None)
device = torch.device(f"cuda:{rank}") device = torch.device(f"cuda:{rank}")
torch.cuda.set_device(device) torch.cuda.set_device(device)
init_test_distributed_environment(tp_size, pp_size, rank, init_test_distributed_environment(tp_size, pp_size, rank,
@ -44,7 +44,7 @@ def all_gather_test_worker(tp_size: int, pp_size: int, rank: int,
# it is important to delete the CUDA_VISIBLE_DEVICES environment variable # it is important to delete the CUDA_VISIBLE_DEVICES environment variable
# so that each worker can see all the GPUs # so that each worker can see all the GPUs
# they will be able to set the device to the correct GPU # they will be able to set the device to the correct GPU
del os.environ["CUDA_VISIBLE_DEVICES"] os.environ.pop("CUDA_VISIBLE_DEVICES", None)
device = torch.device(f"cuda:{rank}") device = torch.device(f"cuda:{rank}")
torch.cuda.set_device(device) torch.cuda.set_device(device)
init_test_distributed_environment(tp_size, pp_size, rank, init_test_distributed_environment(tp_size, pp_size, rank,
@ -72,7 +72,7 @@ def broadcast_tensor_dict_test_worker(tp_size: int, pp_size: int, rank: int,
# it is important to delete the CUDA_VISIBLE_DEVICES environment variable # it is important to delete the CUDA_VISIBLE_DEVICES environment variable
# so that each worker can see all the GPUs # so that each worker can see all the GPUs
# they will be able to set the device to the correct GPU # they will be able to set the device to the correct GPU
del os.environ["CUDA_VISIBLE_DEVICES"] os.environ.pop("CUDA_VISIBLE_DEVICES", None)
device = torch.device(f"cuda:{rank}") device = torch.device(f"cuda:{rank}")
torch.cuda.set_device(device) torch.cuda.set_device(device)
init_test_distributed_environment(tp_size, pp_size, rank, init_test_distributed_environment(tp_size, pp_size, rank,
@ -108,7 +108,7 @@ def broadcast_tensor_dict_test_worker(tp_size: int, pp_size: int, rank: int,
@ray.remote(num_gpus=1, max_calls=1) @ray.remote(num_gpus=1, max_calls=1)
def send_recv_tensor_dict_test_worker(tp_size: int, pp_size: int, rank: int, def send_recv_tensor_dict_test_worker(tp_size: int, pp_size: int, rank: int,
distributed_init_port: str): distributed_init_port: str):
del os.environ["CUDA_VISIBLE_DEVICES"] os.environ.pop("CUDA_VISIBLE_DEVICES", None)
device = torch.device(f"cuda:{rank}") device = torch.device(f"cuda:{rank}")
torch.cuda.set_device(device) torch.cuda.set_device(device)
init_test_distributed_environment(tp_size, pp_size, rank, init_test_distributed_environment(tp_size, pp_size, rank,
@ -148,7 +148,7 @@ def send_recv_tensor_dict_test_worker(tp_size: int, pp_size: int, rank: int,
@ray.remote(num_gpus=1, max_calls=1) @ray.remote(num_gpus=1, max_calls=1)
def send_recv_test_worker(tp_size: int, pp_size: int, rank: int, def send_recv_test_worker(tp_size: int, pp_size: int, rank: int,
distributed_init_port: str): distributed_init_port: str):
del os.environ["CUDA_VISIBLE_DEVICES"] os.environ.pop("CUDA_VISIBLE_DEVICES", None)
device = torch.device(f"cuda:{rank}") device = torch.device(f"cuda:{rank}")
torch.cuda.set_device(device) torch.cuda.set_device(device)
init_test_distributed_environment(tp_size, pp_size, rank, init_test_distributed_environment(tp_size, pp_size, rank,

4
tests/distributed/test_custom_all_reduce.py
View File

@ -24,7 +24,7 @@ for i, v in enumerate(test_sizes):
@ray.remote(num_gpus=1, max_calls=1) @ray.remote(num_gpus=1, max_calls=1)
def graph_allreduce(tp_size, pp_size, rank, distributed_init_port): def graph_allreduce(tp_size, pp_size, rank, distributed_init_port):
del os.environ["CUDA_VISIBLE_DEVICES"] os.environ.pop("CUDA_VISIBLE_DEVICES", None)
device = torch.device(f"cuda:{rank}") device = torch.device(f"cuda:{rank}")
torch.cuda.set_device(device) torch.cuda.set_device(device)
init_test_distributed_environment(tp_size, pp_size, rank, init_test_distributed_environment(tp_size, pp_size, rank,
@ -80,7 +80,7 @@ def graph_allreduce(tp_size, pp_size, rank, distributed_init_port):
@ray.remote(num_gpus=1, max_calls=1) @ray.remote(num_gpus=1, max_calls=1)
def eager_allreduce(tp_size, pp_size, rank, distributed_init_port): def eager_allreduce(tp_size, pp_size, rank, distributed_init_port):
del os.environ["CUDA_VISIBLE_DEVICES"] os.environ.pop("CUDA_VISIBLE_DEVICES", None)
device = torch.device(f"cuda:{rank}") device = torch.device(f"cuda:{rank}")
torch.cuda.set_device(device) torch.cuda.set_device(device)
init_test_distributed_environment(tp_size, pp_size, rank, init_test_distributed_environment(tp_size, pp_size, rank,