fix failed CI on CUDA spawn issue

torch.cuda.is_available() is not working well with multiprocessing case
so we switch to use is_cpu() to check the device

Signed-off-by: Yuan Zhou <yuan.zhou@intel.com>

Author: zhouyuan

tests/conftest.py

@@ -16,6 +16,7 @@
 from vllm.inputs import TextPrompt
 from vllm.logger import init_logger
 from vllm.sequence import MultiModalData, SampleLogprobs
+from vllm.utils import is_cpu
 
 logger = init_logger(__name__)
 
@@ -55,7 +56,7 @@ def cleanup():
     with contextlib.suppress(AssertionError):
         torch.distributed.destroy_process_group()
     gc.collect()
-    if torch.cuda.is_available():
+    if not is_cpu():
         torch.cuda.empty_cache()
 
 
@@ -144,7 +145,7 @@ def example_long_prompts() -> List[str]:
 class HfRunner:
 
     def wrap_device(self, input: any):
-        if torch.cuda.is_available():
+        if not is_cpu():
             return input.cuda()
         else:
             return input.cpu()