Nightly Update: Paddle Submodule

.github/workflows/metax_work.yaml

@@ -5,12 +5,6 @@ on:
   pull_request:
     types: [opened, synchronize]
     branches: [develop, release/**]
-    paths:
-      - "**"
-      - "Paddle/**"
-      - "!backends/**"
-      - "backends/metax_gpu/**"
-
 permissions: read-all
 
 defaults:
@@ -40,16 +34,48 @@ jobs:
             git fetch origin pull/${{ github.event.pull_request.number }}/head:pull/${{ github.event.pull_request.number }}/head
             git checkout pull/${{ github.event.pull_request.number }}/head
 
-            git submodule update --init --recursive
+
+
+
+            paddle_branch=${{ github.base_ref || github.ref_name}}
+            echo $paddle_branch
+            # sleep 10000
+            change_numbers=$(git diff --name-only remotes/origin/${paddle_branch} | wc -l)
+            echo $change_numbers
+
+
+            change_backend=$(git diff --name-only remotes/origin/${paddle_branch} | grep -c "backends/" || true)
+            echo $change_backend
+            change_metax_only=$(git diff --name-only remotes/origin/${paddle_branch} | grep -c "backends/metax_gpu" || true)
+            echo $change_metax_only
+
+            # change_backend=$(git diff --name-only remotes/origin/${paddle_branch} | grep "backends/"| wc -l)
+            # echo $change_backend
+            # change_metax_only=$(git diff --name-only remotes/origin/${paddle_branch} | grep "backends/metax_gpu"| wc -l)
+            # echo $change_metax_only
+
+            git diff --name-only remotes/origin/${paddle_branch}
+
+            if [ $change_numbers -ne $change_backend ]; then
+              echo "Common file changed, continue to run metax FULL CI test ..."
+            elif [ $paddle_branch -eq 0 ] ; then
+              echo "NO metax backend changes found, skip metax FULL CI ....."
+              exit 0
+            fi
+
+
+            # git submodule update --init --recursive
           fi
 
 
       - name: compile
         run: |
+          # sleep 10000
           cd backends/metax_gpu
           bash build.sh
 
       - name: run test
+
         run: |
           cd backends/metax_gpu/tests
           bash run_test.sh -j 16

backends/metax_gpu/CMakeLists.txt

@@ -703,9 +703,9 @@ file(
 set(CUSTOM_DEVICE_SRCS ${CUDA_SRCS} ${CC_SRCS} ${ERNIE_CORE_SRCS})
 
 set_source_files_properties(${CUSTOM_DEVICE_SRCS} PROPERTIES LANGUAGE CUDA)
-
+set(MACA_PATH $ENV{MACA_PATH})
 set(CMAKE_CUCC_COMPILER "cucc")
-set(CMAKE_CUCC_FLAGS "-I /opt/maca/tools/cu-bridge/include/")
+set(CMAKE_CUCC_FLAGS "-I ${MACA_PATH}/tools/cu-bridge/include/")
 
 add_library(${TARGET_NAME} SHARED ${CUSTOM_DEVICE_SRCS})
 
@@ -734,9 +734,9 @@ target_link_libraries(
   ${WARPRNNT_LIBRARIES}
   ${PADDLE_CORE_LIB})
 
-target_link_libraries(${TARGET_NAME} /opt/maca/lib/libmccl.so)
-target_link_libraries(${TARGET_NAME} /opt/maca/lib/libmcFlashAttn.so)
-target_link_libraries(${TARGET_NAME} /opt/maca/lib/libmcpti.so)
+target_link_libraries(${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)
 
 include_directories(BEFORE ${PADDLE_SOURCE_DIR})
 

backends/metax_gpu/cmake/dgc.cmake

@@ -62,8 +62,8 @@ if(EXISTS ${DGC_DOWNLOAD_DIR}/${DGC_CACHE_FILENAME})
 else()
   download_dgc()
 endif()
-
-set(CU_BRIDGE_PATH "/opt/maca/tools/cu-bridge")
+set(MACA_PATH $ENV{MACA_PATH})
+set(CU_BRIDGE_PATH "${MACA_PATH}/tools/cu-bridge")
 
 add_custom_command(
   OUTPUT "${CU_BRIDGE_PATH}/bin/nvcc"

backends/metax_gpu/kernels/cuda_kernels/activation_grad_kernel_register.cu

@@ -119,7 +119,22 @@ void ActivationGradGPUImpl(const Context& dev_ctx,
     ActivationGradGPUImpl<T, Context, funcs::functor_class<T>>( \
         dev_ctx, &x, nullptr, &dout, dx, functor);              \
   }
-
+#define DEFINE_GPU_ACT_GRAD_KERNEL_WITH_TWO_DOUBLE_ATTRS_DEPX(  \
+    name, functor_class, attr1, attr2)                          \
+  template <typename T, typename Context>                       \
+  void name##GradKernel(const Context& dev_ctx,                 \
+                        const DenseTensor& x,                   \
+                        const DenseTensor& dout,                \
+                        double attr1,                           \
+                        double attr2,                           \
+                        DenseTensor* dx) {                      \
+    funcs::functor_class<T> functor;                            \
+    auto attrs = functor.GetAttrs();                            \
+    *(attrs[0].second) = attr1;                                 \
+    *(attrs[1].second) = attr2;                                 \
+    ActivationGradGPUImpl<T, Context, funcs::functor_class<T>>( \
+        dev_ctx, &x, nullptr, &dout, dx, functor);              \
+  }
 #define DEFINE_GPU_ACTIVATION_GRAD_KERNEL_DEPOUT(name, functor_class) \
   template <typename T, typename Context>                             \
   void name##GradKernel(const Context& dev_ctx,                       \
@@ -239,10 +254,10 @@ DEFINE_GPU_ACT_GRAD_KERNEL_WITH_TWO_ATTRS_DEPX(STanh,
                                                scale_a,
                                                scale_b);
 
-DEFINE_GPU_ACT_GRAD_KERNEL_WITH_TWO_ATTRS_DEPX(Softplus,
-                                               CudaSoftplusGradFunctor,
-                                               beta,
-                                               threshold);
+DEFINE_GPU_ACT_GRAD_KERNEL_WITH_TWO_DOUBLE_ATTRS_DEPX(Softplus,
+                                                      CudaSoftplusGradFunctor,
+                                                      beta,
+                                                      threshold);
 DEFINE_GPU_ACT_GRAD_KERNEL_WITH_TWO_ATTRS_DEPOUT(HardSigmoid,
                                                  CudaHardSigmoidGradFunctor,
                                                  slope,

backends/metax_gpu/kernels/cuda_kernels/activation_kernel_register.cu

@@ -90,7 +90,21 @@ void ActivationGPUImpl(const Context& dev_ctx,
     ActivationGPUImpl<T, Context, funcs::functor_class<T>>( \
         dev_ctx, x, out, functor);                          \
   }
-
+#define DEFINE_GPU_ACT_KERNEL_WITH_TWO_DOUBLE_ATTRS(        \
+    name, functor_class, attr1, attr2)                      \
+  template <typename T, typename Context>                   \
+  void name##Kernel(const Context& dev_ctx,                 \
+                    const DenseTensor& x,                   \
+                    double attr1,                           \
+                    double attr2,                           \
+                    DenseTensor* out) {                     \
+    funcs::functor_class<T> functor;                        \
+    auto attrs = functor.GetAttrs();                        \
+    *(attrs[0].second) = attr1;                             \
+    *(attrs[1].second) = attr2;                             \
+    ActivationGPUImpl<T, Context, funcs::functor_class<T>>( \
+        dev_ctx, x, out, functor);                          \
+  }
 DEFINE_GPU_ACTIVATION_KERNEL(Cos, CudaCosFunctor)
 DEFINE_GPU_ACTIVATION_KERNEL(Tan, CudaTanFunctor)
 DEFINE_GPU_ACTIVATION_KERNEL(Acos, CudaAcosFunctor)
@@ -139,10 +153,10 @@ DEFINE_GPU_ACT_KERNEL_WITH_TWO_ATTRS(HardTanh,
                                      t_min,
                                      t_max)
 DEFINE_GPU_ACT_KERNEL_WITH_TWO_ATTRS(Stanh, CudaSTanhFunctor, scale_a, scale_b)
-DEFINE_GPU_ACT_KERNEL_WITH_TWO_ATTRS(Softplus,
-                                     CudaSoftplusFunctor,
-                                     beta,
-                                     threshold)
+DEFINE_GPU_ACT_KERNEL_WITH_TWO_DOUBLE_ATTRS(Softplus,
+                                            CudaSoftplusFunctor,
+                                            beta,
+                                            threshold)
 DEFINE_GPU_ACT_KERNEL_WITH_TWO_ATTRS(HardSigmoid,
                                      CudaHardSigmoidFunctor,
                                      slope,

backends/metax_gpu/kernels/metax_kernel/weight_quantize_kernel_register.cu

@@ -115,11 +115,12 @@ void WeightQuantizeKernel(const Context& dev_ctx,
     dev_ctx.template Alloc<T>(scale);
     weight_quant_gpu<T, Context>(dev_ctx,
                                  x.data<T>(),
-                                 quanted_x.data<int8_t>(),
+                                 out->data<int8_t>(),
                                  scale->data<T>(),
                                  weight_shape,
                                  arch,
                                  algo);
+    out->Resize({m, n});
 #ifdef PADDLE_WITH_HIP
     DenseTensor x_int_tmp(out->type());
     x_int_tmp.Resize({m, n / 2});
@@ -133,12 +134,12 @@ void WeightQuantizeKernel(const Context& dev_ctx,
     funcs::Transpose<Context, int8_t, 2> trans;
     trans(dev_ctx, x_int_tmp, out, axis);
 #else
-    weight_permute_gpu<Context>(dev_ctx,
-                                quanted_x.data<int8_t>(),
-                                out->data<int8_t>(),
-                                weight_shape,
-                                arch,
-                                algo);
+    // weight_permute_gpu<Context>(dev_ctx,
+    //                             quanted_x.data<int8_t>(),
+    //                             out->data<int8_t>(),
+    //                             weight_shape,
+    //                             arch,
+    //                             algo);
 #endif
   } else if (algo == "w4a8") {
     weight_permute_gpu_w4a8<Context>(dev_ctx,

backends/metax_gpu/tests/CMakeLists.txt

@@ -49,7 +49,13 @@ foreach(test_name ${TEST_PROGRAMS})
     continue()
   endif()
 
-  set(CURRENT_TEST_PROGRAM ${PADDLE_LEGACY_TEST_PATH}/${test_name}.py)
+  string(FIND "${test_name}" "metax" METAX_SUFFIX_POS)
+  if(NOT METAX_SUFFIX_POS EQUAL -1)
+    set(CURRENT_TEST_PROGRAM ${METAX_UNIT_TEST_PATH}/${test_name}.py)
+  else()
+    set(CURRENT_TEST_PROGRAM ${PADDLE_LEGACY_TEST_PATH}/${test_name}.py)
+  endif()
+
   if(NOT EXISTS ${CURRENT_TEST_PROGRAM})
     message(WARNING "${CURRENT_TEST_PROGRAM} is not exist, skip it.")
   else()