Merge pull request PaddlePaddle#60 from MyAngelAyase/fix_bus_for_mmlu

fix bug for mmlu
bmers · Oct 25, 2023 · cd754a2 · cd754a2
2 parents f183ccc + 7c03a44
commit cd754a2
Show file tree

Hide file tree

Showing 2 changed files with 8 additions and 8 deletions.
diff --git a/backends/npu/custom_op/llama_decoder_layer_parallel_op.cc b/backends/npu/custom_op/llama_decoder_layer_parallel_op.cc
@@ -255,13 +255,13 @@ std::vector<paddle::Tensor> LlamaDecoderLayerParallelOp(
     std::vector<int32_t> layer_id_vec(1, 0);
     custom_kernel::TensorFromVector(*dev_ctx, layer_id_vec,
                                     *dev_ctx, &(g_llamaDecoderLayerParallelOp->layerIdTensor_));
-    g_llamaDecoderLayerParallelOp->output_ = std::make_shared<phi::DenseTensor>();
-    g_llamaDecoderLayerParallelOp->output_->Resize(phi::make_ddim(hidden.shape()));
-    dev_ctx->Alloc(g_llamaDecoderLayerParallelOp->output_.get(),
-        static_cast<const phi::DenseTensor *>(hidden.impl().get())->dtype());
   }
 
   if (executeCount % layer_num == 0) { // 每个token第一次进layer，更新stop flag
+    g_llamaDecoderLayerParallelOp->output_ = std::make_shared<phi::DenseTensor>();
+    g_llamaDecoderLayerParallelOp->output_->Resize(phi::make_ddim(hidden.shape()));
+    dev_ctx->Alloc(g_llamaDecoderLayerParallelOp->output_.get(),
+        static_cast<const phi::DenseTensor *>(hidden.impl().get())->dtype());  
     g_llamaDecoderLayerParallelOp->UpdateInputTensorAndParam(kv_seq_len);
   }
 

diff --git a/backends/npu/custom_op/llama_encoder_layer_parallel_op.cc b/backends/npu/custom_op/llama_encoder_layer_parallel_op.cc
@@ -221,13 +221,13 @@ std::vector<paddle::Tensor> LlamaEncoderLayerParallelOp(
     std::vector<int32_t> layer_id_vec(1, 0);
     custom_kernel::TensorFromVector(*dev_ctx, layer_id_vec,
                                     *dev_ctx, &(g_llamaEncoderLayerParallelOp->layerIdTensor_));
-    g_llamaEncoderLayerParallelOp->output_ = std::make_shared<phi::DenseTensor>();
-    g_llamaEncoderLayerParallelOp->output_->Resize(phi::make_ddim(hidden.shape()));
-    dev_ctx->Alloc(g_llamaEncoderLayerParallelOp->output_.get(), 
-        static_cast<const phi::DenseTensor *>(hidden.impl().get())->dtype());
   }
 
   if (executeCount % layer_num == 0) {
+    g_llamaEncoderLayerParallelOp->output_ = std::make_shared<phi::DenseTensor>();
+    g_llamaEncoderLayerParallelOp->output_->Resize(phi::make_ddim(hidden.shape()));
+    dev_ctx->Alloc(g_llamaEncoderLayerParallelOp->output_.get(), 
+        static_cast<const phi::DenseTensor *>(hidden.impl().get())->dtype());    
     g_llamaEncoderLayerParallelOp->UpdateInputTensorAndParam(kv_seq_len);
   }