[BugFix] Revert skip capture (#5023)

* Revert "[BugFix][Metax] Fix metax compile issue in get_block_shape_and_split_kv_block (#5000)"

This reverts commit 05da8e3.

* Revert "skip DtoH capture (#4988)"

This reverts commit 5b24013.

Author: Sunny-bot1

custom_ops/gpu_ops/append_attn/get_block_shape_and_split_kv_block.cu

@@ -15,7 +15,6 @@
 #include "helper.h"
 #include "paddle/extension.h"
 #ifndef PADDLE_WITH_CUSTOM_DEVICE_METAX_GPU
-#include "paddle/phi/backends/gpu/cuda/cuda_graph_with_memory_pool.h"
 #include "paddle/phi/core/memory/memcpy.h"
 #endif
 #include "utils.cuh"
@@ -288,13 +287,9 @@ void GetBlockShapeAndSplitKVBlock(
                                                 seq_lens_encoder.data<int>(),
                                                 max_len_tensor_gpu.data<int>(),
                                                 bsz);
-  // Note (sunxin): Skip capturing the DtoH copy (it's time-consuming); CPU data
-  // is only for branching in attention.
-#ifndef PADDLE_WITH_CUSTOM_DEVICE_METAX_GPU
-  if (!phi::backends::gpu::IsCUDAGraphCapturing())
-#endif
-    max_len_tensor_cpu.copy_(
-        max_len_tensor_gpu, max_len_tensor_cpu.place(), false);
+
+  max_len_tensor_cpu.copy_(
+      max_len_tensor_gpu, max_len_tensor_cpu.place(), false);
 
   auto max_len_cpu_ptr = max_len_tensor_cpu.data<int>();
   int max_len_this_time = max_len_cpu_ptr[0];
@@ -403,13 +398,9 @@ void GetBlockShapeAndSplitKVBlock(
           bsz,
           decoder_block_shape_q,
           group_size);
-      // Note (sunxin): Skip capturing the DtoH copy (it's time-consuming); CPU
-      // data is only for branching in attention.
-#ifndef PADDLE_WITH_CUSTOM_DEVICE_METAX_GPU
-      if (!phi::backends::gpu::IsCUDAGraphCapturing())
-#endif
-        decoder_num_blocks_cpu.copy_(
-            decoder_num_blocks_device, decoder_num_blocks_cpu.place(), false);
+
+      decoder_num_blocks_cpu.copy_(
+          decoder_num_blocks_device, decoder_num_blocks_cpu.place(), false);
       PADDLE_ENFORCE_GPU_SUCCESS(cudaMemsetAsync(
           decoder_chunk_size_device.data<int>(), 64, sizeof(int32_t), stream));
     }