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 9, 08:45 UTC
[CI][AMD] Skip 4 GPUs testgroup ray tests (#33305)
Signed-off-by: Ryan Rock <ryan.rock@amd.com> Co-authored-by: TJian <tunjian.tan@embeddedllm.com>
Ryan Rock committed
Jan 29, 2026 at 23:39 UTC
070c811d6f74c55302557878f5982411a3346b4d
2 files changed
+8
examples/offline_inference/rlhf_colocate.py
+5
@@ -30,6 +30,7 @@ https://docs.ray.io/en/latest/placement-groups.html
30
31
import gc
32
import os
33
+import sys
34
35
import ray
36
import torch
@@ -40,6 +41,10 @@ from torch.multiprocessing.reductions import reduce_tensor
41
42
from vllm import LLM
43
44
+if torch.version.hip is not None:
45
+ print("Skipping test for ROCm. Ray is unsupported on vLLM ROCm.")
46
+ sys.exit(0)
47
+
48
49
class MyLLM(LLM):
50
"""Configure the vLLM worker for Ray placement group execution.
tests/distributed/test_utils.py
+3
@@ -10,6 +10,7 @@ import torch
10
import vllm.envs as envs
11
from vllm.distributed.device_communicators.pynccl import PyNcclCommunicator
12
from vllm.distributed.utils import StatelessProcessGroup
13
+from vllm.platforms import current_platform
14
from vllm.utils.network_utils import get_open_port
15
from vllm.utils.system_utils import update_environment_variables
16
from vllm.utils.torch_utils import cuda_device_count_stateless
@@ -32,6 +33,8 @@ class _CUDADeviceCountStatelessTestActor:
33
def test_cuda_device_count_stateless():
34
"""Test that cuda_device_count_stateless changes return value if
35
CUDA_VISIBLE_DEVICES is changed."""
36
+ if current_platform.is_rocm():
37
+ pytest.skip("Skip for ROCm because Ray uses HIP_VISIBLE_DEVICES.")
38
actor = _CUDADeviceCountStatelessTestActor.options( # type: ignore
39
num_gpus=2
40
).remote()