[Model Runner V2] Init cuda graph pool when necessary (#33217)

Signed-off-by: Xinyu Chen <xinyu1.chen@intel.com>

vllm/v1/worker/gpu/cudagraph_utils.py

@@ -45,6 +45,8 @@ class CudaGraphManager:
         )
 
         self.graphs: dict[int, torch.cuda.CUDAGraph] = {}
+        self.pool = None
+        if self.cudagraph_mode != CUDAGraphMode.NONE:
             self.pool = torch.cuda.graph_pool_handle()
         self.hidden_states: torch.Tensor | None = None
 

vllm/v1/worker/gpu/spec_decode/eagle_cudagraph.py

@@ -44,6 +44,8 @@ class EagleCudaGraphManager:
         )
 
         self.graphs: dict[int, torch.cuda.CUDAGraph] = {}
+        self.pool = None
+        if self.cudagraph_mode != CUDAGraphMode.NONE:
             self.pool = torch.cuda.graph_pool_handle()
 
     def get_cudagraph_size(self, num_tokens: int) -> int | None: