未验证 提交 5b814fd5 编写于 作者: W wenbin 提交者: GitHub

remove old data check (#35077)

上级 673bf719
...@@ -139,15 +139,9 @@ int EmbEltwiseLayernormPluginDynamicImpl<T>::enqueue( ...@@ -139,15 +139,9 @@ int EmbEltwiseLayernormPluginDynamicImpl<T>::enqueue(
auto emb_ptr_gpu_d = auto emb_ptr_gpu_d =
emb_ptr_tensor_.mutable_data<int64_t>(platform::CUDAPlace(device_id_)); emb_ptr_tensor_.mutable_data<int64_t>(platform::CUDAPlace(device_id_));
auto new_input_ptr = reinterpret_cast<uintptr_t>(inputs[0]); cudaMemcpyAsync(in_ptr_gpu_d, reinterpret_cast<const void *>(inputs),
sizeof(uintptr_t) * input_num, cudaMemcpyHostToDevice,
if (old_input_ptr_ != new_input_ptr) { stream);
old_input_ptr_ = new_input_ptr;
cudaMemcpyAsync(in_ptr_gpu_d, reinterpret_cast<const void *>(inputs),
sizeof(uintptr_t) * input_num, cudaMemcpyHostToDevice,
stream);
}
auto out_type = output_desc[0].type; auto out_type = output_desc[0].type;
......
...@@ -89,7 +89,6 @@ class EmbEltwiseLayernormPluginDynamicImpl ...@@ -89,7 +89,6 @@ class EmbEltwiseLayernormPluginDynamicImpl
framework::Tensor in_ptr_tensor_, emb_ptr_tensor_; framework::Tensor in_ptr_tensor_, emb_ptr_tensor_;
int device_id_{0}; int device_id_{0};
uintptr_t old_input_ptr_{0};
bool is_initialized_{false}; bool is_initialized_{false};
}; };
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册