[Ascend] Wx/reimpl index put

.github/workflows/_runs-on-nv-step1.yml

@@ -77,17 +77,13 @@ jobs:
             && source ${ENV_PATH}/pt2.0_diopi \
             && python main.py --mode gen_data" \
             || ( cd ${DEEPLINK_PATH}/${GITHUB_RUN_NUMBER}/${BUILD_TEST1} && git clean -xdf ${GEN_DATA} && exit 1 )
-            source ~/Aoss_env.sh
-            ads-cli cp ${DEEPLINK_PATH}/${GITHUB_RUN_NUMBER}/${BUILD_TEST1}/diopi_test/python/cache/ s3://${Platform_ci_aoss_name}:${Platform_ci_aoss_url}@platform.aoss.cn-sh-01c.sensecoreapi-oss.cn${DEEPLINK_PATH}/${GITHUB_RUN_NUMBER}/${BUILD_TEST1}/diopi_test/python/cache/ 
           elif [[ "${GETRUNNER}" == *diopi* ]];then
             ssh SH1424 """
             set -e
             source ${ENV_PATH}/pt2.0_diopi
             cd ${DEEPLINK_PATH}/${GITHUB_RUN_NUMBER} && cd ${BUILD_TEST1} && cd diopi_test/python &&
             srun --job-name=${GITHUB_JOB} --partition=${SLURM_PAR_V100} --time=20 --gres=gpu:1 bash -c 'python main.py --mode gen_data' \
             || ( cd ${DEEPLINK_PATH}/${GITHUB_RUN_NUMBER}/${BUILD_TEST1} && git clean -xdf ${GEN_DATA} && exit 1 )
-            source ~/Aoss_env.sh
-            ads-cli cp ${DEEPLINK_PATH}/${GITHUB_RUN_NUMBER}/${BUILD_TEST1}/diopi_test/python/cache/ s3://${Platform_ci_aoss_name}:${Platform_ci_aoss_url}@platform.aoss.cn-sh-01c.sensecoreapi-oss.cn${DEEPLINK_PATH}/${GITHUB_RUN_NUMBER}/${BUILD_TEST1}/diopi_test/python/cache/ 
             """
           else
             ln -s ${GEN_DATA_PATH}/${GEN_DATA}/diopi ${DEEPLINK_PATH}/${GITHUB_RUN_NUMBER}/${BUILD_TEST1}/diopi_test/python/cache/

.github/workflows/data-cron.yml

@@ -93,6 +93,7 @@ jobs:
           mkdir -p ${DATA_DIR}/source/${GEN_DATA}
           rsync -a --delete ${CLUSTER_V100}:${DATA_DIR}/source/${GEN_DATA}/diopi/ ${DATA_DIR}/source/${GEN_DATA}/diopi/
           """
+<<<<<<< HEAD
 #          ssh ${CLUSTER_ASCEND_910B} """
 #          mkdir -p ${DATA_DIR}/source/${GEN_DATA}
 #          rsync -a --delete ${CLUSTER_V100}:${DATA_DIR}/source/${GEN_DATA}/diopi/ ${DATA_DIR}/source/${GEN_DATA}/diopi/
@@ -114,4 +115,13 @@ jobs:
         run: |
           set -e
           source ~/Aoss_env.sh 
-          ads-cli cp s3://${Platform_ci_aoss_name}:${Platform_ci_aoss_url}@platform.aoss-internal.cn-sh-01c.sensecoreapi-oss.cn${DATA_DIR}/source/${GEN_DATA}/diopi/ ${DATA_DIR}/source/${GEN_DATA}/diopi/
+          ads-cli cp s3://${Platform_ci_aoss_name}:${Platform_ci_aoss_url}@platform.aoss-internal.cn-sh-01c.sensecoreapi-oss.cn${DATA_DIR}/source/${GEN_DATA}/diopi/ ${DATA_DIR}/source/${GEN_DATA}/diopi/
+=======
+          ssh ${CLUSTER_ASCEND_910B} """
+          mkdir -p ${DATA_DIR}/source/${GEN_DATA}
+          rsync -a --delete ${CLUSTER_V100}:${DATA_DIR}/source/${GEN_DATA}/diopi/ ${DATA_DIR}/source/${GEN_DATA}/diopi/
+          """
+          # ssh ${CLUSTER_SUPA} """
+          # rsync -a ${CLUSTER_V100}:${DATA_DIR}/source/${GEN_DATA}/diopi/ ${DATA_DIR}/source/${GEN_DATA}/diopi/
+          # """
+>>>>>>> 2b566db5 ([ci]fix Ascend CI (#1275))

.github/workflows/main.yml

@@ -96,10 +96,8 @@ jobs:
           && rsync -a --delete ${GITHUB_WORKSPACE}/source/ ${CLUSTER_SUPA}:${DEEPLINK_PATH}/${GITHUB_RUN_NUMBER}/source/ || echo "failure to connect to supa"
           ssh ${CLUSTER_1424} "mkdir -p ${DEEPLINK_PATH}/${GITHUB_RUN_NUMBER}/source" \
           && rsync -a --delete ${GITHUB_WORKSPACE}/source/ ${CLUSTER_1424}:${DEEPLINK_PATH}/${GITHUB_RUN_NUMBER}/source/ || echo "failure to connect to sh1424"
-          # ssh ${CLUSTER_ASCEND_910B} "mkdir -p ${DEEPLINK_PATH}/${GITHUB_RUN_NUMBER}/source" \
-          # && rsync -a --delete ${GITHUB_WORKSPACE}/source/ ${CLUSTER_ASCEND_910B}:${DEEPLINK_PATH}/${GITHUB_RUN_NUMBER}/source/ || echo "failure to connect to ascend 910b"
-          source ~/Aoss_env.sh
-          ads-cli cp ${GITHUB_WORKSPACE}/source/ s3://${Platform_ci_aoss_name}:${Platform_ci_aoss_url}@platform.aoss.cn-sh-01c.sensecoreapi-oss.cn${DEEPLINK_PATH}/${GITHUB_RUN_NUMBER}/source/
+          ssh ${CLUSTER_ASCEND_910B} "mkdir -p ${DEEPLINK_PATH}/${GITHUB_RUN_NUMBER}/source" \
+          && rsync -a --delete ${GITHUB_WORKSPACE}/source/ ${CLUSTER_ASCEND_910B}:${DEEPLINK_PATH}/${GITHUB_RUN_NUMBER}/source/ || echo "failure to connect to ascend 910b"
           # ssh ${CLUSTER_KLX} "mkdir -p ${DEEPLINK_PATH}/${GITHUB_RUN_NUMBER}/source" \
           # && rsync -a --delete ${GITHUB_WORKSPACE}/source/ ${CLUSTER_KLX}:${DEEPLINK_PATH}/${GITHUB_RUN_NUMBER}/source/ || echo "failure to connect to kunlunxin"
 
@@ -295,12 +293,6 @@ jobs:
     needs: [Rsync]
     if: ${{ contains( needs.Rsync.outputs.output, 'ASCEND' ) }}
     steps:
-      - name: COPY Source
-        run: |
-           set -e
-           source ~/Aoss_env.sh
-           ads-cli cp s3://${Platform_ci_aoss_name}:${Platform_ci_aoss_url}@platform.aoss-internal.cn-sh-01c.sensecoreapi-oss.cn${DEEPLINK_PATH}/${GITHUB_RUN_NUMBER}/source/ ${DEEPLINK_PATH}/${GITHUB_RUN_NUMBER}/source/
-           ads-cli  --dryrun --deleteSrc cp  s3://${Platform_ci_aoss_name}:${Platform_ci_aoss_url}@platform.aoss-internal.cn-sh-01c.sensecoreapi-oss.cn${DEEPLINK_PATH}/${GITHUB_RUN_NUMBER}/source/ /dev/null 2>&1 >/dev/null
       - name: build
         run: |
           set -e
@@ -327,10 +319,7 @@ jobs:
           set -e
           cd ${DEEPLINK_PATH}/${GITHUB_RUN_NUMBER}/${BUILD_TEST1}
           if [[ \"${{ needs.Rsync.outputs.output }}\" == *GENDATA* ]];then
-            # rsync -a ${CLUSTER_V100}:${DEEPLINK_PATH}/${GITHUB_RUN_NUMBER}/${BUILD_TEST1}/${GEN_DATA}/diopi ${DEEPLINK_PATH}/${GITHUB_RUN_NUMBER}/${BUILD_TEST1}/diopi_test/python/cache/
-            source ~/Aoss_env.sh
-            ads-cli cp s3://${Platform_ci_aoss_name}:${Platform_ci_aoss_url}@platform.aoss-internal.cn-sh-01c.sensecoreapi-oss.cn${DEEPLINK_PATH}/${GITHUB_RUN_NUMBER}/${BUILD_TEST1}/diopi_test/python/cache/ ${DEEPLINK_PATH}/${GITHUB_RUN_NUMBER}/${BUILD_TEST1}/diopi_test/python/cache/
-            ads-cli  --dryrun --deleteSrc cp  s3://${Platform_ci_aoss_name}:${Platform_ci_aoss_url}@platform.aoss-internal.cn-sh-01c.sensecoreapi-oss.cn${DEEPLINK_PATH}/${GITHUB_RUN_NUMBER}/${BUILD_TEST1}/diopi_test/python/cache/ /dev/null 2>&1 >/dev/null
+            rsync -a ${CLUSTER_V100}:${DEEPLINK_PATH}/${GITHUB_RUN_NUMBER}/${BUILD_TEST1}/${GEN_DATA}/diopi ${DEEPLINK_PATH}/${GITHUB_RUN_NUMBER}/${BUILD_TEST1}/diopi_test/python/cache/
           else
             ln -s ${GEN_DATA_PATH}/${GEN_DATA}/diopi ${DEEPLINK_PATH}/${GITHUB_RUN_NUMBER}/${BUILD_TEST1}/diopi_test/python/cache/
           fi

impl/ascend/functions/index.cpp

@@ -0,0 +1,41 @@
+/**
+ * @file
+ * @author DeepLink
+ * @copyright  (c) 2024, DeepLink.
+ */
+
+#include "../aclnn/adaptor.hpp"
+
+namespace impl {
+namespace ascend {
+
+diopiError_t diopiIndexBackward(diopiContextHandle_t ctx, diopiTensorHandle_t gradInput, diopiTensorHandle_t zerosLikeInput, diopiConstTensorHandle_t* indices,
+                                int64_t nums, diopiConstTensorHandle_t gradOutput) {
+    AscendTensor gradInputTensor(gradInput);
+    AscendTensor gradOutputTensor(gradOutput);
+    if (gradInputTensor.numel() == 0 || gradOutputTensor.numel() == 0) {
+        return diopiSuccess;
+    }
+
+    std::vector<diopiConstTensorHandle_t> indicesVec;
+    indicesVec.reserve(nums);
+
+    for (int i = 0; i < nums; i++) {
+        if (indices[i] != nullptr) {
+            indicesVec.emplace_back(indices[i]);
+        } else {
+            int64_t array[1] = {0};
+            diopiSize_t size = {array, 1};
+            diopiTensorHandle_t emptyTensor = nullptr;
+            diopiRequireTensor(ctx, &emptyTensor, &size, nullptr, gradOutputTensor.dtype(), diopi_device);
+            indicesVec.emplace_back(emptyTensor);
+        }
+    }
+
+    DIOPI_ASCEND_CALL_ACLNN(aclnnInplaceCopy, ctx, gradInput, zerosLikeInput);
+    DIOPI_ASCEND_CALL_ACLNN(aclnnIndexPutImpl, ctx, gradInput, indicesVec, gradOutput, true, false);
+    return diopiSuccess;
+}
+
+}  // namespace ascend
+}  // namespace impl

impl/ascend/functions/index_put.cpp

@@ -0,0 +1,68 @@
+/**
+ * @file
+ * @author DeepLink
+ * @copyright  (c) 2024, DeepLink.
+ */
+
+#include "../aclnn/adaptor.hpp"
+
+namespace impl {
+namespace ascend {
+
+diopiError_t diopiIndexPut(diopiContextHandle_t ctx, diopiTensorHandle_t out, diopiConstTensorHandle_t input, diopiConstTensorHandle_t values,
+                           diopiConstTensorHandle_t* indices, int64_t indicesCounts, bool accumulate) {
+    AscendTensor inputTensor(input);
+    AscendTensor valuesTensor(values);
+    if (inputTensor.numel() == 0 || valuesTensor.numel() == 0) {
+        return diopiSuccess;
+    }
+
+    std::vector<diopiConstTensorHandle_t> indicesVec;
+    indicesVec.reserve(indicesCounts);
+
+    for (int i = 0; i < indicesCounts; i++) {
+        if (indices[i] != nullptr) {
+            indicesVec.emplace_back(indices[i]);
+        } else {
+            int64_t array[1] = {0};
+            diopiSize_t size = {array, 1};
+            diopiTensorHandle_t emptyTensor = nullptr;
+            diopiRequireTensor(ctx, &emptyTensor, &size, nullptr, inputTensor.dtype(), diopi_device);
+            indicesVec.emplace_back(emptyTensor);
+        }
+    }
+
+    DIOPI_ASCEND_CALL_ACLNN(aclnnInplaceCopy, ctx, out, input);
+    DIOPI_ASCEND_CALL_ACLNN(aclnnIndexPutImpl, ctx, out, indicesVec, values, accumulate, false);
+    return diopiSuccess;
+}
+
+diopiError_t diopiIndexPutInp(diopiContextHandle_t ctx, diopiTensorHandle_t input, diopiConstTensorHandle_t values, diopiConstTensorHandle_t* indices,
+                              int64_t indicesCounts, bool accumulate) {
+    AscendTensor inputTensor(input);
+    AscendTensor valuesTensor(values);
+    if (inputTensor.numel() == 0 || valuesTensor.numel() == 0) {
+        return diopiSuccess;
+    }
+
+    std::vector<diopiConstTensorHandle_t> indicesVec;
+    indicesVec.reserve(indicesCounts);
+
+    for (int i = 0; i < indicesCounts; i++) {
+        if (indices[i] != nullptr) {
+            indicesVec.emplace_back(indices[i]);
+        } else {
+            int64_t array[1] = {0};
+            diopiSize_t size = {array, 1};
+            diopiTensorHandle_t emptyTensor = nullptr;
+            diopiRequireTensor(ctx, &emptyTensor, &size, nullptr, inputTensor.dtype(), diopi_device);
+            indicesVec.emplace_back(emptyTensor);
+        }
+    }
+
+    DIOPI_ASCEND_CALL_ACLNN(aclnnIndexPutImpl, ctx, input, indicesVec, values, accumulate, false);
+    return diopiSuccess;
+}
+
+}  // namespace ascend
+}  // namespace impl

impl/ascend_npu/CMakeLists.txt

@@ -174,6 +174,8 @@ set(OLD_IMPL_SRC
     ${OLD_IMPL_DIR}/functions/arange.cpp
     ${OLD_IMPL_DIR}/functions/gather.cpp
     ${OLD_IMPL_DIR}/functions/layer_norm.cpp
+    ${OLD_IMPL_DIR}/functions/index.cpp
+    ${OLD_IMPL_DIR}/functions/index_put.cpp
     ${OLD_IMPL_DIR}/functions/index_select.cpp
     ${OLD_IMPL_DIR}/functions/repeat.cpp
     ${OLD_IMPL_DIR}/functions/group_norm.cpp

impl/ascend_npu/ascend_config.yaml

@@ -100,6 +100,9 @@ ascend:
 - diopiHardtanh
 - diopiHardtanhBackward
 - diopiHardtanhInp
+- diopiIndexBackward
+- diopiIndexPut
+- diopiIndexPutInp
 - diopiIsNan
 - diopiIndexSelect
 - diopiIndexSelectBackward
@@ -244,10 +247,7 @@ ascend:
 - diopiZeros
 ascend_npu:
 - diopiCopyInp
-- diopiIndexPut
-- diopiIndexPutInp
 - diopiIndex
-- diopiIndexBackward
 - diopiDestIndexCopyKV
 - diopiTokenAttentionInference
 - diopiTokenSoftmaxReduceVInference