[Graph Optimization] Add full_cuda_graph to control subgraph split (#6027)

fastdeploy/model_executor/layers/attention/append_attn_backend.py

@@ -164,6 +164,20 @@ class AppendAttentionBackend(AttentionBackend):
 
         self.rank, self.device_id = init_rank_and_device_id(fd_config)
         self.use_output = not fd_config.graph_opt_config.full_cuda_graph
+        if self.use_output:
+            flag = "FLAGS_cuda_graph_blacklist"
+            paddle.set_flags(
+                {
+                    flag: ",".join(
+                        list(
+                            set(
+                                paddle.get_flags(flag)[flag].split(",")
+                                + ["custom_op.static_op_append_attention_with_output_"]
+                            )
+                        )
+                    )
+                }
+            )
         self.fd_config = fd_config
 
     def init_attention_metadata(self, forward_meta: ForwardMeta):

tests/ce/deploy/ernie45t_21b_cinn_fp8.yaml

@@ -6,3 +6,4 @@ graph_optimization_config:
   graph_opt_level: 2
   sot_warmup_sizes: [2,16,32,64]
   use_cudagraph: True
+  full_cuda_graph: False

tests/ce/deploy/ernie45t_21b_cinn_wint4.yaml

@@ -6,3 +6,4 @@ graph_optimization_config:
   graph_opt_level: 2
   sot_warmup_sizes: [2,16,32,64]
   use_cudagraph: True
+  full_cuda_graph: False

tests/ce/deploy/ernie45t_21b_sot_fp8.yaml

@@ -6,3 +6,4 @@ graph_optimization_config:
   graph_opt_level: 1
   sot_warmup_sizes: [2,16,32,64]
   use_cudagraph: True
+  full_cuda_graph: False

tests/ce/deploy/ernie45t_21b_sot_wint4.yaml

@@ -6,3 +6,4 @@ graph_optimization_config:
   graph_opt_level: 1
   sot_warmup_sizes: [2,16,32,64]
   use_cudagraph: True
+  full_cuda_graph: False

tests/graph_optimization/test_static_graph_cuda_graph_split.py

@@ -89,7 +89,9 @@ class TestStaticGraphCUDAGraphSplit(unittest.TestCase):
     def test(self):
         """Run test case"""
         # Set FastDeploy config
-        graph_opt_config = GraphOptimizationConfig({"use_cudagraph": True, "graph_opt_level": 1})
+        graph_opt_config = GraphOptimizationConfig(
+            {"use_cudagraph": True, "graph_opt_level": 1, "full_cuda_graph": False}
+        )
         scheduler_config = SchedulerConfig({"max_num_seqs": 1})
         graph_opt_config._set_cudagraph_sizes(max_capture_size=scheduler_config.max_num_seqs)
         graph_opt_config.init_with_cudagrpah_size(max_capture_size=scheduler_config.max_num_seqs)