Add description of prerequisites for running benchmarks (#9230)

Co-authored-by: Haifeng Jin <[email protected]>

Author: haifeng-jin

benchmarks/README.md

@@ -4,6 +4,39 @@ The two main benchmarking scripts are
   - `experiment_runner.py` to run benchmark experiments, and
   - `result_analyzer.py` to aggregate the benchmark result in CSV form.
 
+## Prerequisites
+
+Follow the [contributing
+guide](https://github.com/pytorch/xla/blob/master/CONTRIBUTING.md) to set up
+the repo for development. Using the devcontainer is recommended.
+
+Install the required libraries by the models.  Depending on the specific
+models, you may need to install `torchvision`, `torchaudio`, `torchtext`.
+
+Follow [this guide](https://pytorch.org/get-started/locally/) to install the
+`torchvision` and `torchaudio` version you need. Use the following command to
+install `torchtext`.
+
+```shell
+pip install torchtext
+```
+
+The following LibGL library is
+required by the Llava model.
+
+```shell
+apt install libgl1-mesa-glx
+```
+
+For the scripts to find the models to benchmark, clone the
+[`pytorch/benchmark`](https://github.com/pytorch/benchmark) repo into your
+workspace directory, which also contains the `pytorch` and the `vision` repo.
+
+```shell
+cd $WORKSPACE  # cd to the workspace directory containing `pytorch` and `vision`
+git clone [email protected]:pytorch/benchmark.git
+```
+
 
 ## Patching mismatched batch sizes