[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>

examples/offline_inference/rlhf_colocate.py

@@ -30,6 +30,7 @@ https://docs.ray.io/en/latest/placement-groups.html
 
 import gc
 import os
+import sys
 
 import ray
 import torch
@@ -40,6 +41,10 @@ from torch.multiprocessing.reductions import reduce_tensor
 
 from vllm import LLM
 
+if torch.version.hip is not None:
+    print("Skipping test for ROCm. Ray is unsupported on vLLM ROCm.")
+    sys.exit(0)
+
 
 class MyLLM(LLM):
     """Configure the vLLM worker for Ray placement group execution.

tests/distributed/test_utils.py

@@ -10,6 +10,7 @@ import torch
 import vllm.envs as envs
 from vllm.distributed.device_communicators.pynccl import PyNcclCommunicator
 from vllm.distributed.utils import StatelessProcessGroup
+from vllm.platforms import current_platform
 from vllm.utils.network_utils import get_open_port
 from vllm.utils.system_utils import update_environment_variables
 from vllm.utils.torch_utils import cuda_device_count_stateless
@@ -32,6 +33,8 @@ class _CUDADeviceCountStatelessTestActor:
 def test_cuda_device_count_stateless():
     """Test that cuda_device_count_stateless changes return value if
     CUDA_VISIBLE_DEVICES is changed."""
+    if current_platform.is_rocm():
+        pytest.skip("Skip for ROCm because Ray uses HIP_VISIBLE_DEVICES.")
     actor = _CUDADeviceCountStatelessTestActor.options(  # type: ignore
         num_gpus=2
     ).remote()