Add a short README for the OPT examples and small tweaks. (#1793)

* Small changes to OPT example.

* Update opt README.

* Add a few modes to batch script.

* Update README.md

tank/examples/opt/README.md

@@ -1,3 +1,26 @@
-# Running Different OPT Variants
+# Run OPT for sentence completion through SHARK
 
-To run different sizes of OPT, change the string `OPT_MODEL` string in `opt_torch_test.py`. The default is 350m parameters. 66b cases also exist in the file, simply uncomment the test cases.
+From base SHARK directory, follow instructions to set up a virtual environment with SHARK. (`./setup_venv.sh` or `./setup_venv.ps1`)
+Then, you may run opt_causallm.py to get a very simple sentence completion application running through SHARK
+```
+python opt_causallm.py
+```
+
+# Run OPT performance comparison on SHARK vs. PyTorch
+
+```
+python opt_perf_comparison.py --max-seq-len=512 --model-name=facebook/opt-1.3b \
+        --platform=shark
+```
+Any OPT model from huggingface should work with this script, and you can choose between `--platform=shark` or `--platform=huggingface` to generate benchmarks of OPT inference on SHARK / PyTorch. 
+
+# Run a small suite of OPT models through the benchmark script
+
+```
+python opt_perf_comparison_batch.py
+```
+This script will run benchmarks from a suite of OPT configurations:
+- Sequence Lengths: 32, 128, 256, 512
+- Parameter Counts: 125m, 350m, 1.3b
+
+note: Most of these scripts are written for use on CPU, as perf comparisons against pytorch can be problematic across platforms otherwise.

tank/examples/opt/opt_perf_comparison.py

@@ -4,7 +4,10 @@ PyTorch.
 
 Usage Example:
 
-python opt_perf_comparision.py --max-seq-len=32 --model-name=facebook/opt-125m \
+python opt_perf_comparison.py --max-seq-len=32 --model-name=facebook/opt-125m \
+        --platform=shark
+
+python opt_perf_comparison.py --max-seq-len=512 --model-name=facebook/opt-1.3b \
         --platform=shark
 
 See parse_args() below for command line argument usage.

tank/examples/opt/opt_perf_comparison_batch.py

@@ -11,8 +11,8 @@ import subprocess
 
 def make_commands() -> Iterable[List[str]]:
     command = shlex.split("python opt_perf_comparison.py --no-save-json")
-    max_seq_lens = [32, 128, 512]
-    model_names = ["facebook/opt-" + e for e in ["125m", "350m"]]
+    max_seq_lens = [32, 128, 256, 512]
+    model_names = ["facebook/opt-" + e for e in ["125m", "350m", "1.3b"]]
     for max_seq_len in max_seq_lens:
         for model_name in model_names:
             yield command + [