[Metax] update (#229) (#2528)

* Set MXCC_OVERRIDE_OPTIONS in compile script

Add MXCC_OVERRIDE_OPTIONS for metax GPU compilation.

* Add MXCC_OVERRIDE_OPTIONS for Metax GPU

* Update flash_attn_grad_kernel.cu

* Update compile.sh

* [Metax][feat] add top_p_sampling.patch. (#225)

* [Metax] Fix add flags

* [Metax] update

---------

Co-authored-by: duqimeng <77875733+duqimeng@users.noreply.github.com>
Co-authored-by: MingkunZhang <39252862+StareAtYou@users.noreply.github.com>

Author: 3 people

backends/metax_gpu/CMakeLists.txt

@@ -849,6 +849,8 @@ target_link_libraries(${TARGET_NAME} ${MACA_PATH}/lib/libmccl.so)
 target_link_libraries(${TARGET_NAME} ${MACA_PATH}/lib/libmcFlashAttn.so)
 target_link_libraries(${TARGET_NAME} ${MACA_PATH}/lib/libmcpti.so)
 
+target_link_options(${TARGET_NAME} PRIVATE "-T${CMAKE_SOURCE_DIR}/my_script.ld")
+
 if(WITH_CINN)
   message(STATUS "[MetaX] Linking CINN object library")
   target_link_libraries(${TARGET_NAME} $<TARGET_OBJECTS:metax_cinn_obj>)

backends/metax_gpu/my_script.lds

@@ -0,0 +1,8 @@
+SECTIONS
+{
+  OVERLAY :
+  {
+    .mc_fatbin { *(.mc_fatbin) }
+  }
+}
+INSERT AFTER .comment;

backends/metax_gpu/patch/paddle.patch

@@ -1117,3 +1117,76 @@ index 368cb21c21..f0f99fbd2f 100644
    return (getLaneId() == 0) ? 0ULL : (1ULL << getLaneId()) - 1ULL;
  #else
    unsigned mask;
+diff --git a/paddle/phi/kernels/gpu/multiclass_nms3_kernel.cu b/paddle/phi/kernels/gpu/multiclass_nms3_kernel.cu
+index 7ced1fdc17..e49759ebb4 100644
+--- a/paddle/phi/kernels/gpu/multiclass_nms3_kernel.cu
++++ b/paddle/phi/kernels/gpu/multiclass_nms3_kernel.cu
+@@ -302,11 +302,11 @@ void SortScoresPerClassGPU(gpuStream_t stream,
+       begin_bit,
+       end_bit,
+       stream);
+-#ifdef PADDLE_WITH_HIP
+-  PADDLE_ENFORCE_GPU_SUCCESS(hipGetLastError());
+-#else
+-  PADDLE_ENFORCE_GPU_SUCCESS(cudaGetLastError());
+-#endif
++// #ifdef PADDLE_WITH_HIP
++//   PADDLE_ENFORCE_GPU_SUCCESS(hipGetLastError());
++// #else
++//   PADDLE_ENFORCE_GPU_SUCCESS(cudaGetLastError());
++// #endif
+ }
+ 
+ /* ===========
+@@ -611,11 +611,11 @@ void AllClassNMSGPU(gpuStream_t stream,
+       score_shift,
+       caffe_semantics);
+ 
+-#ifdef PADDLE_WITH_HIP
+-  PADDLE_ENFORCE_GPU_SUCCESS(hipGetLastError());
+-#else
+-  PADDLE_ENFORCE_GPU_SUCCESS(cudaGetLastError());
+-#endif
++// #ifdef PADDLE_WITH_HIP
++//   PADDLE_ENFORCE_GPU_SUCCESS(hipGetLastError());
++// #else
++//   PADDLE_ENFORCE_GPU_SUCCESS(cudaGetLastError());
++// #endif
+ }
+ 
+ /* ==================
+@@ -769,11 +769,11 @@ void GatherNMSOutputsGPU(gpuStream_t stream,
+                               reinterpret_cast<int*>(nmsed_valid_mask),
+                               clip_boxes,
+                               T_SCORE(score_shift));
+-#ifdef PADDLE_WITH_HIP
+-  PADDLE_ENFORCE_GPU_SUCCESS(hipGetLastError());
+-#else
+-  PADDLE_ENFORCE_GPU_SUCCESS(cudaGetLastError());
+-#endif
++// #ifdef PADDLE_WITH_HIP
++//   PADDLE_ENFORCE_GPU_SUCCESS(hipGetLastError());
++// #else
++//   PADDLE_ENFORCE_GPU_SUCCESS(cudaGetLastError());
++// #endif
+ }
+ 
+ template <typename T_SCORE>
+@@ -818,11 +818,11 @@ void SortScoresPerImageGPU(gpuStream_t stream,
+       begin_bit,
+       end_bit,
+       stream);
+-#ifdef PADDLE_WITH_HIP
+-  PADDLE_ENFORCE_GPU_SUCCESS(hipGetLastError());
+-#else
+-  PADDLE_ENFORCE_GPU_SUCCESS(cudaGetLastError());
+-#endif
++// #ifdef PADDLE_WITH_HIP
++//   PADDLE_ENFORCE_GPU_SUCCESS(hipGetLastError());
++// #else
++//   PADDLE_ENFORCE_GPU_SUCCESS(cudaGetLastError());
++// #endif
+ }
+ 
+ template <typename T>
+