Created by: Meiyim

paddle 1.5.1 fluid_inference develop 预测网络代码（增量部分），增量前没有预测问题：

        src_ids, sent_ids = features
        zero = L.fill_constant([1], dtype='int64', value=0)
        input_mask = L.cast(L.equal(src_ids, zero), dtype) # assume pad id == 0
        d_shape = L.shape(src_ids)
        seqlen = d_shape[1]
        batch_size = d_shape[0]
        pos_ids = L.unsqueeze(L.range(0, seqlen, 1, dtype='int32'), axes=[0])
        pos_ids = L.expand(pos_ids, [batch_size, 1])
        pos_ids = L.unsqueeze(pos_ids, axes=[2])
        pos_ids = L.cast(pos_ids, 'int64')
        pos_ids.stop_gradient = True
        input_mask.stop_gradient = True
        task_ids = L.zeros_like(src_ids) 
        task_ids.stop_gradient = True
        bert = ErnieModel(
            src_ids=src_ids,
            position_ids=pos_ids,
            sentence_ids=sent_ids,
            task_ids=task_ids,
            input_mask=input_mask,
            config=self.hparam,
            use_fp16=self.hparam['fp16']
        )

paddle predictor 配置:

  paddle::AnalysisConfig config;
  config.SetModel(FLAGS_model_dir);
  config.EnableUseGpu(8000);
  config.SwitchSpecifyInputNames(true);

  auto predictor = CreatePaddlePredictor(config);

无论在多显存的机器上都会运行爆显存（后台观察到显卡瞬间被打满）：

I0828 15:42:42.367769 207866 buddy_allocator.cc:69] Allocate 150994944 bytes from chunk size 150995200
I0828 15:42:42.367838 207866 gpu_info.cc:195] GPU usage 8M/7611M, 8M available to allocate
I0828 15:42:42.367843 207866 gpu_info.cc:216] Alloc size is 4 MiB, is it Re-alloc: 1
W0828 15:42:42.368458 207866 system_allocator.cc:121] Cannot malloc 144 MB GPU memory. Please shrink FLAGS_fraction_of_gpu_memory_to_use or FLAGS_initial_gpu_memory_in_mb or FLAGS_reallocate_gpu_memory_in_mbenvironment variable to a lower value. Current FLAGS_fraction_of_gpu_memory_to_use value is 0.531531. Current FLAGS_initial_gpu_memory_in_mb value is 0. Current FLAGS_reallocate_gpu_memory_in_mb value is 0
F0828 15:42:42.368579 207866 naive_best_fit_allocator.cc:164] Cannot allocate 144.000000MB in GPU 0, available 8.937500MB, total 7.433533GB, GpuMinChunkSize 256.000000B, GpuMaxChunkSize 3.668286GB, GPU memory used: 7.089328GB
*** Check failure stack trace: ***
./gpu.sh: line 13: 207866 Aborted                 (core dumped) ./build/inference --logtostderr --model_dir $2 --data $1 --repeat 5 --output_prediction true --use_gpu true --device 0
                                                                                                                                                                                          49301,1       Bot