PartitionModel function => add another abstract partition interface to support user-specific features( such as partition between same type device )

[jackiechen19940708]

PartitionModel function => add another abstract partition interface to support user-specific features( such as partition model between same type device to enable parallelism )

for the following function ,maybe it's good to provide an abstract interface for user to make a specific device partition

int Compilation::PartitionModel(
    Context* context,
    Model* model,
    std::vector<std::pair<
        Context::DeviceContext*,
        std::tuple<core::Model*, bool, std::vector<int>, std::vector<int>>>>*
        models) {
...
    for (auto& operation : model->model_.operations) {
            if (_flags_[operation_index] && !flags[operation_index]) {                             <----- here may be provide another abstraction for user specificed partition and all way down into other revelant code (such as user specificed ModelPartitioner)
              flags[operation_index] = true;
              // Only the operations which are not supported by the previous devices
              // are added.
              supported_operations[i].second.insert(&operation);                                                  
            }
            operation_index++;
          }



// For example we have <model> as follows:
//        conv2d_0                       conv2d_1
//           |                                        |
//      batch_norm_0                batch_norm_1
//           |                                        |
//         relu_0                              relu_1
//              \                                  /
//                    \                        /
//                        leave out ops(such as operator)


// and two devices with support operators listed 
//   <device_id=0, [conv2d_0,batch_norm_0,relu_0, conv2d_1,batch_norm_1,relu_1]>
//   <device_id=1, [conv2d_0,batch_norm_0,relu_0, conv2d_1,batch_norm_1,relu_1]>

// we need to use a interface to partition the model into devices as follows
//   <device_id=0, {[conv2d_0,batch_norm_0,relu_0], false, [...], [...]}>>
//   <device_id=1, {[conv2d_1,batch_norm_1,relu_1], false, [...], [...]>>

• 版本、环境信息
  1）Paddle Lite 版本：3c61295 (https://github.com/PaddlePaddle/Paddle-Lite/blob/3c61295edf66eb5c24e7c98b604c14a1bd4457d7/lite/backends/nnadapter/nnadapter/src/runtime/compilation.cc#L380C1-L386C18)
     2）Host 环境：Not Revelant
     3）预测后端信息：Not Revelant
• 复现信息：Nor Relevant
• 建议描述：Nor Relevant

[qili93]

您好，请问310B是昇腾23年新出的芯片类型对吗？这个芯片类型Lite尚未适配，Lite会将此需求纳入需求池进行评审，感谢您的反馈！

[hong19860320]

PartitionModel function => add another abstract partition interface to support user-specific features( such as partition model between same type device to enable parallelism )

for the following function ,maybe it's good to provide an abstract interface for user to make a specific device partition

int Compilation::PartitionModel(
    Context* context,
    Model* model,
    std::vector<std::pair<
        Context::DeviceContext*,
        std::tuple<core::Model*, bool, std::vector<int>, std::vector<int>>>>*
        models) {
...
    for (auto& operation : model->model_.operations) {
            if (_flags_[operation_index] && !flags[operation_index]) {                             <----- here may be provide another abstraction for user specificed partition and all way down into other revelant code (such as user specificed ModelPartitioner)
              flags[operation_index] = true;
              // Only the operations which are not supported by the previous devices
              // are added.
              supported_operations[i].second.insert(&operation);                                                  
            }
            operation_index++;
          }



// For example we have <model> as follows:
//        conv2d_0                       conv2d_1
//           |                                        |
//      batch_norm_0                batch_norm_1
//           |                                        |
//         relu_0                              relu_1
//              \                                  /
//                    \                        /
//                        leave out ops(such as operator)


// and two devices with support operators listed 
//   <device_id=0, [conv2d_0,batch_norm_0,relu_0, conv2d_1,batch_norm_1,relu_1]>
//   <device_id=1, [conv2d_0,batch_norm_0,relu_0, conv2d_1,batch_norm_1,relu_1]>

// we need to use a interface to partition the model into devices as follows
//   <device_id=0, {[conv2d_0,batch_norm_0,relu_0], false, [...], [...]}>>
//   <device_id=1, {[conv2d_1,batch_norm_1,relu_1], false, [...], [...]>>

• 版本、环境信息
  1）Paddle Lite 版本：3c61295 (https://github.com/PaddlePaddle/Paddle-Lite/blob/3c61295edf66eb5c24e7c98b604c14a1bd4457d7/lite/backends/nnadapter/nnadapter/src/runtime/compilation.cc#L380C1-L386C18)
     2）Host 环境：Not Revelant
     3）预测后端信息：Not Revelant
• 复现信息：Nor Relevant
• 建议描述：Nor Relevant

It's good idea, we will consider your suggestion in the future.
Now we provide another solution, please check https://www.paddlepaddle.org.cn/lite/v2.12/develop_guides/nnadapter.html