[torchbench][optimus] Add backend optimus (#167357)

Summary: `--optimus [all | vertical_opt | horizontal_opt]` will kick off inductor compile with different fusion strategies.

Test Plan:
TorchBench Runner:

```
$ buck2 run mode/opt //pytorch/benchmark:run -- customized_optimus_illustrative -t train -d cuda
GPU Time per batch:   56.254 milliseconds
CPU Wall Time per batch:  56.326 milliseconds
CPU Wall Time:        56.326 milliseconds
Time to first batch:          420.0777 ms
GPU 0 Peak Memory:              0.0695 GB
CPU Peak Memory:              359.6362 GB
```

PT2 Benchmark Runner (comparing with eager):

```
buck2 run mode/opt //pytorch/benchmark:pt2 -- --only customized_optimus_illustrative  --performance --training --inductor

running benchmark: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 30/30 [00:02<00:00, 14.37it/s]
4.509x
```

eager latency: ~56 ms
inductor latency: ~11 ms

Optimus backend:

```
$ buck2 run mode/opt //pytorch/benchmark:pt2 -- --only customized_optimus_illustrative --performance --training --optimus all
11.02923508733511 ms, 13.884015614166856 ms, 0.794x
```

```
$ buck2 run mode/opt //pytorch/benchmark:pt2 -- --only customized_optimus_illustrative --performance --training --optimus vertical_opt
12.47156853787601 ms, 10.699485195800662 ms, 1.166x
```

```
$ buck2 run mode/opt //pytorch/benchmark:pt2 -- --only customized_optimus_illustrative --performance --training --optimus horizontal_opt
11.078484123572707 ms, 10.797873372212052 ms, 1.026x
```

optimus latency ~10 ms

Differential Revision: D86524903

Pull Request resolved: https://github.com/pytorch/pytorch/pull/167357
Approved by: https://github.com/mengluy0125

benchmarks/dynamo/common.py

@@ -952,7 +952,7 @@ def latency_experiment_summary(suite_name, args, model, timings, **kwargs):
         first_fields.append(kwargs["tag"])
     headers = first_headers + ["speedup", "abs_latency"]
     row = first_fields + [float(speedup), median[1] * 1000]
-    msg = f"{speedup:.3f}x"
+    msg = f"{median[0] * 1000} ms, {median[1] * 1000} ms, {speedup:.3f}x"
     if args.baseline:
         headers.extend(
             [
@@ -1010,7 +1010,7 @@ def latency_experiment_summary(suite_name, args, model, timings, **kwargs):
     # Hypothetically you can use this from other places, but it's currently
     # inaccessible, and when this assert fails you need to update the
     # event_name here to account for the other cases you are using this
-    assert args.quantization is not None
+    assert any([args.quantization, args.optimus])
     output_signpost(
         dict(zip(headers, row)),
         args,
@@ -2587,6 +2587,9 @@ class BenchmarkRunner:
                 **experiment_kwargs,
             )
 
+            # reset dynamo
+            torch._dynamo.reset()
+
             if self.args.export_aot_inductor:
                 optimized_model_iter_fn = optimize_ctx
             else:
@@ -2950,7 +2953,7 @@ class BenchmarkRunner:
             status = self.check_tolerance(name, model, example_inputs, optimize_ctx)
             print(status)
         elif self.args.performance:
-            if self.args.backend == "torchao":
+            if self.args.backend in ["torchao", "optimus"]:
                 status = self.run_performance_test_non_alternate(
                     name, model, example_inputs, optimize_ctx, experiment, tag
                 )
@@ -3526,6 +3529,12 @@ def parse_args(args=None):
         action="store_true",
         help="Measure speedup with TorchInductor",
     )
+    group.add_argument(
+        "--optimus",
+        choices=["vertical_opt", "horizontal_opt", "all"],
+        default=None,
+        help="Measure speedup of Optimus with TorchInductor baseline",
+    )
     group.add_argument(
         "--quantization",
         choices=[
@@ -3783,6 +3792,9 @@ def run(runner, args, original_dir=None):
     if args.inductor:
         assert args.backend is None
         args.backend = "inductor"
+    if args.optimus:
+        assert args.backend is None
+        args.backend = "optimus"
     if args.quantization:
         assert args.backend is None
         args.backend = "torchao"
@@ -4067,10 +4079,22 @@ def run(runner, args, original_dir=None):
 
             runner.model_iter_fn = model_iter_fn_and_mark_step
             optimize_ctx = torchao_optimize_ctx(args.quantization)
+        elif args.backend == "optimus":
+            from .optimus import get_baseline_ctx, get_optimus_optimize_ctx
+
+            baseline_ctx = get_baseline_ctx(
+                nopython=args.nopython, inductor_compile_mode=args.inductor_compile_mode
+            )
+            runner.model_iter_fn = baseline_ctx(runner.model_iter_fn)
+            optimize_ctx = get_optimus_optimize_ctx(
+                args.optimus, args.nopython, args.inductor_compile_mode
+            )
         else:
             optimize_ctx = torch._dynamo.optimize(args.backend, nopython=args.nopython)
         experiment = (
-            speedup_experiment if args.backend != "torchao" else latency_experiment
+            speedup_experiment
+            if args.backend not in ["torchao", "optimus"]
+            else latency_experiment
         )
         if args.accuracy:
             output_filename = f"accuracy_{args.backend}.csv"
@@ -4091,7 +4115,12 @@ def run(runner, args, original_dir=None):
     if args.only in runner.disable_cudagraph_models:
         args.disable_cudagraphs = True
 
-    if args.inductor or args.backend == "inductor" or args.export_aot_inductor:
+    if (
+        args.inductor
+        or args.backend == "inductor"
+        or args.export_aot_inductor
+        or args.backend == "optimus"
+    ):
         inductor_config.triton.cudagraphs = not args.disable_cudagraphs
         inductor_config.triton.persistent_reductions = (
             not args.disable_persistent_reductions

benchmarks/dynamo/optimus.py

@@ -0,0 +1,62 @@
+import functools
+
+import torch
+
+
+def get_baseline_ctx(nopython, inductor_compile_mode):
+    return functools.partial(
+        torch.compile,
+        backend="inductor",
+        fullgraph=nopython,
+        mode=inductor_compile_mode,
+    )
+
+
+def get_optimus_optimize_ctx(config, nopython, inductor_compile_mode):
+    if config == "vertical_opt":
+        optimus_inductor_config = {
+            "pre_grad_fusion_options": {
+                "normalization_pass": {},
+                "merge_splits_pass": {},
+                "split_cat_pass": {},
+                "unbind_stack_pass": {},
+                "unbind_cat_to_view_pass": {},
+            }
+        }
+    elif config == "horizontal_opt":
+        optimus_inductor_config = {
+            "pre_grad_fusion_options": {
+                "normalization_pass": {},
+                "batch_linear": {},
+                "batch_layernorm": {},
+            },
+        }
+    elif config == "all":
+        optimus_inductor_config = {
+            "pre_grad_fusion_options": {
+                "normalization_pass": {},
+                "batch_linear": {},
+                "batch_layernorm": {},
+                "merge_splits_pass": {},
+                "split_cat_pass": {},
+                "unbind_stack_pass": {},
+                "unbind_cat_to_view_pass": {},
+            },
+        }
+    else:
+        raise RuntimeError(f"Unknown optimus config: {config}")
+
+    def _inner(fn):
+        if "pre_grad_fusion_options" in optimus_inductor_config:
+            torch._inductor.config.pre_grad_fusion_options = optimus_inductor_config[
+                "pre_grad_fusion_options"
+            ]
+        if "post_grad_fusion_options" in optimus_inductor_config:
+            torch._inductor.config.post_grad_fusion_options = optimus_inductor_config[
+                "post_grad_fusion_options"
+            ]
+        return torch.compile(
+            fn, backend="inductor", fullgraph=nopython, mode=inductor_compile_mode
+        )
+
+    return _inner