cinn_launch op: fix dtype of tensor is always mutable_data<float> (#45835)

paddle/fluid/operators/cinn/cinn_launch_context.cc

@@ -270,8 +270,9 @@ void CinnLaunchContext::AssignExternalVariable(const std::string& var_name) {
       [this, var_name](void* ctx, cinn_buffer_t* buffer) {
         auto* tensor = cached_scope_->GetVar(var_name)->GetMutable<LoDTensor>();
         tensor->Resize(framework::DDim(buffer->dims, buffer->dimensions));
-        buffer->memory = reinterpret_cast<uint8_t*>(
-            tensor->mutable_data<float>(*cached_place_));
+        buffer->memory = reinterpret_cast<uint8_t*>(tensor->mutable_data(
+            *cached_place_,
+            framework::paddle2cinn::TransToPaddleDataType(buffer->type)));
         return 0;
       });
 
@@ -295,8 +296,9 @@ void CinnLaunchContext::AssignInternalVariable(const std::string& var_name) {
         auto* tensor =
             cached_temp_scope_->Var(var_name)->GetMutable<LoDTensor>();
         tensor->Resize(framework::DDim(buffer->dims, buffer->dimensions));
-        buffer->memory = reinterpret_cast<uint8_t*>(
-            tensor->mutable_data<float>(*cached_place_));
+        buffer->memory = reinterpret_cast<uint8_t*>(tensor->mutable_data(
+            *cached_place_,
+            framework::paddle2cinn::TransToPaddleDataType(buffer->type)));
         return 0;
       });
 
@@ -437,7 +439,8 @@ ParallelExecutor* CinnLaunchContext::InitializePE(const platform::Place& place,
     auto* buffer = GetCinnBufferOfVar(var_name);
     auto dim = framework::DDim(buffer->dims, buffer->dimensions);
     var->GetMutable<LoDTensor>()->Resize(dim);
-    var->GetMutable<LoDTensor>()->mutable_data<float>(place);
+    var->GetMutable<LoDTensor>()->mutable_data(
+        place, framework::paddle2cinn::TransToPaddleDataType(buffer->type));
   }
   return parallel_executor_.get();
 }