Merge branch 'develop' into enable_leaky_relu

Author: YqGe585

Paddle

@@ -13,7 +13,7 @@
 // limitations under the License.
 
 #include "paddle/phi/core/kernel_registry.h"
-#include "paddle/phi/kernels/gpu/c_softmax_with_cross_entropy_grad_kernel.cu"  // NOLINT
+#include "paddle/phi/kernels/c_softmax_with_cross_entropy_grad_kernel.h"
 
 PD_CUSTOM_KERNEL_REGISTER(c_softmax_with_cross_entropy_grad,
                           iluvatar_gpu,

backends/iluvatar_gpu/kernels/cuda_kernels/c_softmax_with_cross_entropy_grad_kernel_register.cc

@@ -13,7 +13,7 @@
 // limitations under the License.
 
 #include "paddle/phi/core/kernel_registry.h"
-#include "paddle/phi/kernels/fusion/gpu/fused_transpose_split_quant_kernel.cu"  //NOLINT
+#include "paddle/phi/kernels/fusion/gpu/fused_transpose_split_quant_kernel.h"
 #include "paddle/phi/kernels/fusion/gpu/quant_utils.h"
 
 PD_CUSTOM_KERNEL_REGISTER(fused_transpose_split_quant,

backends/iluvatar_gpu/kernels/cuda_kernels/fused_transpose_split_quant_kernel_register.cu

@@ -180,16 +180,7 @@ index a943bbed9a..dc6384985f 100644
 
  #ifdef PADDLE_WITH_CUDNN_FRONTEND
  #define CUDNN_DNN_ROUTINE_EACH_FRONTEND(__macro) \
-@@ -191,7 +208,7 @@ CUDNN_DNN_ROUTINE_EACH_FRONTEND(DECLARE_DYNAMIC_LOAD_CUDNN_WRAP)
- CUDNN_DNN_ROUTINE_EACH_REMOVED_IN_E9(DECLARE_DYNAMIC_LOAD_CUDNN_WRAP)
- #endif
- 
--#if CUDNN_VERSION < 90000
-+#if CUDNN_VERSION < 90000 && CUDNN_VERSION >= 7201
- #define CUDNN_DNN_ROUTINE_EACH_AFTER_TWO_R7_REMOVED_IN_E9(__macro) \
-   __macro(cudnnSetRNNPaddingMode);                                 \
-   __macro(cudnnRNNForwardInferenceEx);                             \
-@@ -212,6 +229,15 @@ CUDNN_DNN_ROUTINE_EACH_AFTER_TWO_R7_REMOVED_IN_E9(
+@@ -210,6 +227,15 @@ CUDNN_DNN_ROUTINE_EACH_AFTER_TWO_R7_REMOVED_IN_E9(
    __macro(cudnnRNNBackwardWeights_v8);
  CUDNN_DNN_ROUTINE_EACH_R9(DECLARE_DYNAMIC_LOAD_CUDNN_WRAP)
  #endif

backends/iluvatar_gpu/patches/paddle-corex.patch

@@ -11,18 +11,37 @@ set(EXCLUDED_TESTS
     "test_ops_roi_align.py" "test_tensor_fill_diagonal_tensor_op_iluvatar.py"
     "test_tensor_fill_diagonal_op_iluvatar.py" "test_bincount_op_iluvatar.py")
 
-list(JOIN EXCLUDED_TESTS " or " EXCLUDE_EXPR)
-set(K_FILTER "not (${EXCLUDE_EXPR})")
-
-add_test(
-  NAME pytest_main
-  COMMAND ${Python_EXECUTABLE} -m pytest ${CMAKE_SOURCE_DIR}/unittests/ -k
-          ${K_FILTER} --tb=short -v
-  WORKING_DIRECTORY ${CMAKE_SOURCE_DIR})
-
-set_tests_properties(
-  pytest_main
-  PROPERTIES
-    ENVIRONMENT "PYTHONPATH=${LEGACY_TEST_PATH}:$ENV{PYTHONPATH}"
-    "LD_PRELOAD=${LD_LIBRARY_PATH}/libcuda.so.1"
-    "CUSTOM_DEVICE_ROOT=${CMAKE_BINARY_DIR}/python/paddle_custom_device/")
+# Find all test files
+file(GLOB TEST_FILES "${CMAKE_SOURCE_DIR}/unittests/test_*.py")
+
+# Create a separate CTest test for each test file
+foreach(TEST_FILE ${TEST_FILES})
+  get_filename_component(TEST_NAME ${TEST_FILE} NAME)
+
+  # Check if it's in the exclusion list
+  set(IS_EXCLUDED FALSE)
+  foreach(EXCLUDED_TEST ${EXCLUDED_TESTS})
+    if(TEST_NAME STREQUAL EXCLUDED_TEST)
+      set(IS_EXCLUDED TRUE)
+      break()
+    endif()
+  endforeach()
+
+  # If not in the exclusion list, add the test
+  if(NOT IS_EXCLUDED)
+    # Get the filename without extension as the test name
+    get_filename_component(TEST_BASE_NAME ${TEST_FILE} NAME_WE)
+
+    add_test(
+      NAME ${TEST_BASE_NAME}
+      COMMAND ${Python_EXECUTABLE} -m pytest ${TEST_FILE} --tb=short
+      WORKING_DIRECTORY ${CMAKE_SOURCE_DIR})
+
+    set_tests_properties(
+      ${TEST_BASE_NAME}
+      PROPERTIES
+        ENVIRONMENT "PYTHONPATH=${LEGACY_TEST_PATH}:$ENV{PYTHONPATH}"
+        "LD_PRELOAD=${LD_LIBRARY_PATH}/libcuda.so.1"
+        "CUSTOM_DEVICE_ROOT=${CMAKE_BINARY_DIR}/python/paddle_custom_device/")
+  endif()
+endforeach()

backends/iluvatar_gpu/tests/CMakeLists.txt

@@ -63,7 +63,7 @@ echo "=== Building project (if needed) ==="
 make -j$(nproc) || { echo "ERROR: Build failed"; exit 1; }
 
 echo "=== Running tests ==="
-ctest --output-on-failure -V || { 
+ctest --output-on-failure -V -j8 || { 
     echo "ERROR: Tests failed!" >&2
     echo "Exit code: $?" >&2
     exit 1

backends/iluvatar_gpu/tests/run_test.sh

@@ -13,7 +13,7 @@
 // limitations under the License.
 
 #include "paddle/phi/core/kernel_registry.h"
-#include "paddle/phi/kernels/fusion/gpu/fused_transpose_split_quant_kernel.cu"  //NOLINT
+#include "paddle/phi/kernels/fusion/gpu/fused_transpose_split_quant_kernel.h"
 #include "paddle/phi/kernels/fusion/gpu/quant_utils.h"
 
 PD_CUSTOM_KERNEL_REGISTER(fused_transpose_split_quant,