[Others] add assert and only count the actual load in cuda_graph (#5445)

fastdeploy/model_executor/graph_optimization/graph_optimization_backend.py

@@ -130,7 +130,6 @@ class GraphOptBackend:
         )
 
     def __call__(self, **kwargs):
-        self._debug_count_total_step += 1
         if not self.fd_config.graph_opt_config.use_cudagraph:
             return self.runnable(**kwargs)
         if self.cudagraph_piecewise_backend is None:
@@ -140,6 +139,9 @@ class GraphOptBackend:
 
         assert kwargs["forward_meta"].ids_remove_padding is not None
         real_shape = kwargs["forward_meta"].ids_remove_padding.shape[0]
+        if real_shape > 0:
+            # only count the actual load.
+            self._debug_count_total_step += 1
 
         if (not kwargs["forward_meta"].step_use_cudagraph) or (real_shape > self.cudagraph_switch_threshold):
             return self.dy_runnable(**kwargs)

fastdeploy/model_executor/layers/attention/append_attn_backend.py

@@ -233,9 +233,20 @@ class AppendAttentionBackend(AttentionBackend):
         forward_mixed
         """
         metadata = self.attention_metadata
-
         sliding_window = layer.sliding_window
 
+        if self.rope_3d:
+            assert len(forward_meta.rotary_embs.shape) == 6
+        else:
+            assert len(forward_meta.rotary_embs.shape) == 5
+            if layer.use_neox_rotary_style:
+                assert forward_meta.rotary_embs.shape[0:4] == [2, 1, self.max_seq_len, 1]
+                # 128 is qwen3
+                # 32 is glm
+                assert forward_meta.rotary_embs.shape[4] in [128, 32]
+            else:
+                assert forward_meta.rotary_embs.shape == [2, 1, self.max_seq_len, 1, 64]
+
         if self.pd_disaggregation_mode == "per_query":
             metadata.kv_signal_data_list[layer.layer_id] = init_signal_layerwise(
                 metadata.kv_signal_metadata,