fix_qkv_plugin: half_scale

Author: Wangzheee

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

@@ -347,8 +347,8 @@ int QkvToContextPluginDynamic::enqueue(
  platform::CUDAPlace(device_id)));
 
  int n_q = seq_len * head_number_ * head_size_ * batch;
- constexpr int threads = 128;
- int blocks = (n_q + threads - 1) / threads;
+ int threads = head_number_ * head_size_ * batch;
+ int blocks = seq_len;
 
  apply_scale<<<blocks, threads, 0, stream>>>(tptr, static_cast<half>(scale_),
  n_q);