[Bugfix] Fix Stream usage in CPU model runner and OneDNN kernel check (#25046)

Signed-off-by: jiang1.li <jiang1.li@intel.com>

csrc/cpu/dnnl_kernels.cpp

@@ -523,7 +523,7 @@ void onednn_mm(torch::Tensor& c,        // [M, OC], row-major
   CPU_KERNEL_GUARD_IN(onednn_mm)
   TORCH_CHECK(a.dim() == 2);
   TORCH_CHECK(a.stride(-1) == 1);
-  TORCH_CHECK(c.is_contiguous());
+  TORCH_CHECK(c.stride(-1) == 1);
   MatMulPrimitiveHandler* ptr =
       reinterpret_cast<MatMulPrimitiveHandler*>(handler);
 

vllm/platforms/cpu.py

@@ -185,6 +185,11 @@ class CpuPlatform(Platform):
             parallel_config.distributed_executor_backend = "mp"
         if parallel_config.worker_cls == "auto":
             parallel_config.worker_cls = "vllm.v1.worker.cpu_worker.CPUWorker"
+        # Disable DBO
+        if parallel_config.enable_dbo:
+            logger.warning(
+                "Dual-Batch Overlap is not supported on CPU, disabled.")
+            parallel_config.enable_dbo = False
 
         # Note: workaround for v1 gpu_model_runner
         from vllm.config import CompilationLevel

vllm/v1/worker/cpu_model_runner.py

@@ -145,12 +145,20 @@ def _torch_cuda_wrapper():
             self.record = lambda: None
             self.synchronize = lambda: None
 
+    class _StreamPlaceholder:
+
+        def __init__(self, *args, **kwargs) -> None:
+            pass
+
     cuda_event = torch.cuda.Event
+    cuda_stream = torch.cuda.Stream
     try:
         torch.cuda.Event = _EventPlaceholder
+        torch.cuda.Stream = _StreamPlaceholder
         yield
     finally:
         torch.cuda.Event = cuda_event
+        torch.cuda.Stream = cuda_stream
 
 
 @contextmanager