Yiqun Liu

[Not Review] Change the calculation of ips for benchmark.

1

本PR主要用来在高性能团队之间同步测试代码。

Fix a typo in README.cn.md.

1

fix #787

Call paddle.incubate.nn.functional.fused_gate_attention and set use_flash_attn to true.

1

rt

Optimize the use of no_recompute_layers.

6

### PR types Others ### PR changes Others ### Description RT

[Not Review] Enable profile with detail nvtx tag.

5

### PR types Others ### PR changes Others ### Description rt

Reorgnize the configs.

1

rt

rt

Enable to get the precise gpu_time for tf by using ctypes package to load libcudart.so and add add nchw support for batch_norm.

1

### 该PR主要包括以下工作： 1. 添加控制benchmark行为的`BENCHMARK_NEED_FEED`和`BENCHMARK_NEED_FETCH`两个环境变量，并且在`BENCHMARK_NEED_FETCH`开启时，解析nvprof log时会减掉DtoH的时间 2. 更新tf支持的版本到最新版本2.7.0