[Paddle-TRT] Implement MHA fp16 order same as training (#32629) (#32785)

* implement MHA order same as training

* fix fp16 compile issue on old architecture

Co-authored-by: zlsh80826 <rewang@nvidia.com>

Author: shangzhizhou

paddle/fluid/inference/tensorrt/plugin/qkv_to_context_plugin.cu

@@ -225,6 +225,14 @@ nvinfer1::DataType QkvToContextPluginDynamic::getOutputDataType(
   return input_types[0];
 }
 
+template <typename T>
+__global__ void apply_scale(T *data, T scale, int n) {
+#if CUDA_ARCH_FP16_SUPPORTED(__CUDA_ARCH__)
+  int tid = blockIdx.x * blockDim.x + threadIdx.x;
+  data[tid] = data[tid] * scale;
+#endif
+}
+
 int QkvToContextPluginDynamic::enqueue(
     const nvinfer1::PluginTensorDesc *input_desc,
     const nvinfer1::PluginTensorDesc *output_desc, const void *const *inputs,
@@ -291,10 +299,17 @@ int QkvToContextPluginDynamic::enqueue(
         platform::DeviceContextPool::Instance().Get(
             platform::CUDAPlace(device_id)));
 
+    int n_q = seq_len * head_number_ * head_size_;
+    constexpr int threads = 128;
+    int blocks = (n_q + threads - 1) / threads;
+
+    apply_scale<<<blocks, threads, 0, stream>>>(tptr, static_cast<half>(scale_),
+                                                n_q);
+
     const platform::CUDADeviceContext &dev_ctx = *device_ctx;
     operators::math::MultiHeadGPUComputeFunctor<half> multihead_compute_func;
     multihead_compute_func(dev_ctx, batch, seq_len, head_number_, head_size_,
-                           qkptr, input1_data, tptr, half(scale_), half(0.0));
+                           qkptr, input1_data, tptr, half(1.), half(0.0));
 
     int grid = batch * head_number_ * seq_len;
     int block = head_size_;