[Paddle-TRT] Fix QkvToContextPluginDynamic bug (#50715)

* fix multihead * fix multihead

[Paddle-TRT] Fix QkvToContextPluginDynamic bug (#50715)
* fix multihead * fix multihead
612d5da0 · zhoutianzi666 · GitHub · 21c6eccf · 612d5da0
隐藏空白更改
内联并排

Showing with 2 addition and 0 deletion

paddle/fluid/inference/tensorrt/plugin/qkv_to_context_plugin.cu .../fluid/inference/tensorrt/plugin/qkv_to_context_plugin.cu +2 -0

未找到文件。
--- a/paddle/fluid/inference/tensorrt/plugin/qkv_to_context_plugin.cu
+++ b/paddle/fluid/inference/tensorrt/plugin/qkv_to_context_plugin.cu
@@ -479,6 +479,8 @@ int QkvToContextPluginDynamic::enqueue(
    const half *input1_data = static_cast<const half *>(qk_bias);
    // BxSx3xNxH => tptr: 3xBxNxSxH.
    if (need_padding) {
+      PADDLE_ENFORCE_GPU_SUCCESS(
+          cudaMemsetAsync(tptr, 0, sizeof(half) * input_num, stream));
      TransposePadding(input0_data,
                       tptr,
                       batch,