[SOT][Cudagraph] Remove BreakGraph of #3302 && update CustomOp (#3694)

* rm inplace info && to(gpu)

* update append_attention

* unpin paddle version

* add full_cuda_graph=False

* add blank line

---------

Co-authored-by: SigureMo <sigure.qaq@gmail.com>

custom_ops/gpu_ops/append_attention.cu

@@ -593,7 +593,7 @@ std::vector<paddle::Tensor> AppendAttention(
   return {paddle::Tensor{}};
 }
 
-void AppendAttentionWithOutput(
+std::vector<paddle::Tensor> AppendAttentionWithOutput(
     const paddle::Tensor& qkv,
     const paddle::Tensor& key_cache,
     const paddle::Tensor& value_cache,
@@ -756,6 +756,8 @@ void AppendAttentionWithOutput(
       break;
     }
   }
+
+  return {fmha_out};
 }
 
 
@@ -1112,10 +1114,8 @@ PD_BUILD_STATIC_OP(append_attention_with_output)
              paddle::Optional("kv_signal_data"),
              paddle::Optional("q_norm_weight"),
              paddle::Optional("k_norm_weight")})
-    .Outputs({"fmha_out_out", "qkv_out", "key_cache_out", "value_cache_out"})
-    .SetInplaceMap({{"fmha_out", "fmha_out_out"},
-                    {"key_cache", "key_cache_out"},
-                    {"value_cache", "value_cache_out"}})
+    .Outputs({"fmha_out_out"})
+    .SetInplaceMap({{"fmha_out", "fmha_out_out"}})
     .Attrs({"rms_norm_eps: float",
             "compute_type: std::string",
             "cache_quant_type: std::string",

custom_ops/gpu_ops/cpp_extensions.cc

@@ -91,7 +91,7 @@ std::vector<paddle::Tensor> AppendAttention(
     const int speculate_max_draft_token_num, const bool causal,
     const bool speculate_decoder);
 
-void AppendAttentionWithOutput(
+std::vector<paddle::Tensor> AppendAttentionWithOutput(
     const paddle::Tensor &qkv, const paddle::Tensor &key_cache,
     const paddle::Tensor &value_cache, const paddle::Tensor &seq_lens_encoder,
     const paddle::Tensor &seq_lens_decoder,