notes for ptpu cuda-graph

Author: cyx11111

vllm_fl/compilation/graph.py

@@ -207,10 +207,12 @@ def __call__(self, *args, **kwargs):
                 pass
 
             graph_kwargs: dict[str, Any] = {"pool": self.graph_pool}
+            # PTPU: pass the active stream explicitly. torch.ptpu.graph()
             if current_platform.device_type == "ptpu":
                 graph_kwargs["stream"] = (
                     current_platform.torch_device_fn.current_stream()
                 )
+            # FL-specific: use platform-agnostic graph capture
             with current_platform.torch_device_fn.graph(graph, **graph_kwargs):
                 # `output` is managed by pytorch's cudagraph pool
                 output = self.runnable(*args, **kwargs)