Nightly Update: Paddle Submodule

.github/workflows/metax_work.yaml

@@ -0,0 +1,54 @@
+name: padlle metax gpu test
+
+on:
+  workflow_dispatch:
+  pull_request:
+    types: [opened, synchronize]
+    branches: [develop, release/**]
+    paths:
+      - "**"
+      - "!backends/**"
+      - "backends/metax_gpu/**"
+
+permissions: read-all
+
+defaults:
+  run:
+    shell: bash
+
+jobs:
+  metax-gpu-test:
+    runs-on: paddle-metax-runner-set
+    # runs-on: debug-paddle-runner-set
+    steps:
+      - name: Checkout repository
+        run: |
+          git config --global user.name "GitHub Actions"
+          git config --global user.email "[email protected]"
+
+          git clone \
+            --reference-if-able /home/runner/PaddleCustomDevice \
+            --depth=1 \
+            --shallow-submodules \
+            --jobs=8 \
+            --branch ${{ github.base_ref || github.ref_name}} \
+            --recurse-submodules \
+            https://${{ github.actor }}:${{ secrets.GITHUB_TOKEN }}@github.com/${{ github.repository }}.git .
+
+          if [ "${{ github.event_name }}" == "pull_request" ]; then
+            git fetch origin pull/${{ 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
+          fi
+
+
+      - name: compile
+        run: |
+          cd backends/metax_gpu
+          bash build.sh
+
+      - name: run test
+        run: |
+          cd backends/metax_gpu/tests
+          bash run_test.sh

backends/metax_gpu/CMakeLists.txt

@@ -37,6 +37,8 @@ include(cblas)
 include(flashattn)
 include(cutlass)
 include(dgc)
+include(warpctc)
+include(warprnnt)
 
 set(PLUGIN_VERSION ${PADDLE_VERSION})
 
@@ -612,12 +614,9 @@ file(
   ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/funcs/math_function.cc
   ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/gpu/log_softmax_kernel.cu
   ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/gpu/log_softmax_grad_kernel.cu
-  # ${PADDLE_SOURCE_DIR}/paddle/phi/backends/context_pool.cc
   ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/funcs/repeat_tensor2index_tensor.cu
   ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/gpu/binomial_kernel.cu
   ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/gpu/bernoulli_kernel.cu
-  # ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/gpu/bmm_grad_kernel_impl.h
-  # ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/gpu/bmm_kernel.cu
   ${PADDLE_SOURCE_DIR}/paddle/phi/backends/dynload/cufft.cc
   ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/gpu/box_coder_kernel.cu
   ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/gpu/broadcast_tensors_kernel.cu
@@ -640,31 +639,12 @@ file(
   ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/gpu/gather_tree_kernel.cu
   ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/gpu/graph_reindex_kernel.cu
   ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/gpu/graph_sample_neighbors_kernel.cu
-  # ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/gpu/group_norm_kernel.cu
   ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/gpu/group_norm_grad_kernel.cu
   ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/gpu/gumbel_softmax_grad_kernel.cu
   ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/gpu/gumbel_softmax_kernel.cu
-  # ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/fusion/gpu/fused_act_dequant_kernel.cu
-  # ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/fusion/gpu/block_multi_head_attention_kernel.cu
-  # ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/fusion/gpu/fused_weighted_swiglu_act_quant_kernel.cu
-  # ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/fusion/gpu/fused_elemwise_activation_kernel.cu
-  # ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/fusion/gpu/fused_softmax_mask_upper_triangle_grad_kernel.cu
-  # ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/fusion/fp8_gemm/fp8_gemm_with_cublasLt/fp8_fp8_half_gemm.cu
-  # ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/fusion/cutlass/memory_efficient_attention_grad_kernel.cu
-  # ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/fusion/cutlass/fused_conv2d_add_act_kernel.cu
-  # ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/fusion/cutlass/variable_length_memory_efficient_attention_kernel.cu
-  # ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/fusion/cutlass/memory_efficient_attention_kernel.cu
-  # ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/fusion/cutlass/gemm_epilogue_kernel.cu
-  # ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/fusion/gpu/blha_get_max_len.cu
-  # ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/fusion/gpu/fused_elemwise_activation_grad_kernel.cu
-  # ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/stride/as_real_kernel.cc
-  # ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/stride/as_complex_kernel.cc
-  # ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/stride/complex_grad_kernel.cc
-  # ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/stride/complex_kernel.cc
-  # ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/selected_rows/shape_kernel.cc
-  # ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/sparse/gpu/conv_kernel_igemm.cu
+  ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/gpu/top_p_sampling_kernel.cu
+  ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/gpu/rms_norm_kernel.cu
   # ############################################################################
-  # kernels/fusion kernels/selected_rows
   ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/selected_rows/gpu/adamw_kernel.cu
   # kernels/kps
   ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/kps/elementwise_kernel.cu
@@ -696,7 +676,6 @@ file(
   ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/legacy/gpu/cal_aux_loss_grad_kernel.cu
   ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/legacy/gpu/expand_modality_expert_id_kernel.cu
   ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/legacy/gpu/int_bincount_kernel.cu
-  ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/gpu/top_p_sampling_kernel.cu
   ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/fusion/gpu/fused_bias_act_kernel.cu)
 
 file(
@@ -706,15 +685,14 @@ file(
   passes/*.cc
   kernels/*.cc
   kernels/*.cu
+  kernels/fusion/*.cc
+  kernels/fusion/*.cu
   kernels/gpudnn/*.cc
   kernels/gpudnn/*.cu
   kernels/cuda_kernels/*.cc
   kernels/cuda_kernels/*.cu
   kernels/funcs/blas/*.cc
-  kernels/ernie_core/*.cu
-  kernels/ernie_core/rms_norm_kernel_register.cu
-  kernels/ernie_core/top_p_sampling_kernel_register.cu
-  kernels/ernie_core/fused_bias_act_kernel_register.cu)
+  kernels/ernie_core/*.cu)
 
 set(CUSTOM_DEVICE_SRCS ${CUDA_SRCS} ${CC_SRCS} ${ERNIE_CORE_SRCS})
 
@@ -723,18 +701,17 @@ set_source_files_properties(${CUSTOM_DEVICE_SRCS} PROPERTIES LANGUAGE CUDA)
 set(CMAKE_CUCC_COMPILER "cucc")
 set(CMAKE_CUCC_FLAGS "-I /opt/maca/tools/cu-bridge/include/")
 
-set_source_files_properties(
-  ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/gpu/rms_norm_kernel.cu
-  PROPERTIES LANGUAGE CUDA)
-add_library(
-  ${TARGET_NAME} SHARED
-  ${CUSTOM_DEVICE_SRCS}
-  ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/gpu/rms_norm_kernel.cu)
+add_library(${TARGET_NAME} SHARED ${CUSTOM_DEVICE_SRCS})
 
 target_include_directories(
   ${TARGET_NAME}
-  PRIVATE ${PADDLE_SOURCE_DIR} ${CMAKE_SOURCE_DIR} ${CMAKE_SOURCE_DIR}/kernels
-          ${CUDA_INCLUDE_DIRS} ${PADDLE_SOURCE_DIR}/third_party/pybind/include
+  PRIVATE ${PADDLE_SOURCE_DIR}
+          ${CMAKE_SOURCE_DIR}
+          ${CMAKE_SOURCE_DIR}/kernels
+          ${CUDA_INCLUDE_DIRS}
+          ${WARPCTC_INCLUDE_DIR}
+          ${WARPRNNT_INCLUDE_DIR}
+          ${PADDLE_SOURCE_DIR}/third_party/pybind/include
           ${PADDLE_SOURCE_DIR}/paddle/phi/api/include/compat)
 
 target_link_libraries(
@@ -747,16 +724,16 @@ target_link_libraries(
   protobuf
   external_error_proto
   dgc
+  ${WARPCTC_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)
 include_directories(BEFORE ${PADDLE_SOURCE_DIR})
 
 target_compile_definitions(
   ${TARGET_NAME}
   PUBLIC PADDLE_WITH_CUDA=1
          PADDLE_WITH_CUSTOM_DEVICE=1
+         mcblasContext=cublasContext
          GPUContext=CustomContext
          KPSContext=CustomContext
          STREAM_TYPE=cudaStream_t

backends/metax_gpu/build.sh

@@ -20,15 +20,23 @@ set -e
 pip  uninstall paddlepaddle -y
 
 
-export http_proxy=http://10.2.192.21:1080 https_proxy=http://10.2.192.21:1080
+# init paddle
+# git submodule sync --recursive && git submodule update --init --recursive
+
+# sleep 1000000
+# unset http_proxy https_proxy
+
+
+# export http_proxy=https://172.17.0.1:1080 https_proxy=http://10.2.192.21:1080
+# export
 pip install safetensors==0.6.2 -i https://mirrors.tuna.tsinghua.edu.cn/pypi/web/simple some-package
 # install paddle
+
+
 python -m pip install --pre paddlepaddle -i https://www.paddlepaddle.org.cn/packages/nightly/cpu/
 
-# exit 1
-# init paddle
-git submodule sync --recursive && git submodule update --init --recursive
 
+# unset http_proxy https_proxy
 
 # apply patch
 bash change_patch.sh
@@ -49,8 +57,8 @@ fi
 
 echo "make_maca"
 cd build
-cmake_maca .. -DPython3_EXECUTABLE=$(which python3) -DWITH_GPU=ON
-make_maca -j8
+cmake_maca .. -DCMAKE_BUILD_TYPE=Release -DPython3_EXECUTABLE=$(which python3) -DWITH_GPU=ON
+make_maca -j60
 
 echo "install whl"
 pip install dist/paddle_metax_gpu*.whl --force-reinstall

backends/metax_gpu/change_patch.sh

@@ -21,7 +21,9 @@ unzip mcEigen_3.4.0_paddle_final.zip
 mv mcEigen_3.4.0_paddle_final eigen3
 cd ..
 cp -r patch/eigen3/ ../../Paddle/third_party/eigen3
+rm -r patch/eigen3
 cp patch/tmp/mixed_vector* ../../Paddle/paddle/phi/core
 cd ../../Paddle/
 git apply --verbose ../backends/metax_gpu/patch/paddle.patch
 cd -
+# cp -r patch/intrinsics.cuh ../../Paddle/third_party/warpctc/include/contrib/moderngpu/include/device/