Create PT2-D MAST sample

[ghstack-poisoned]

userbenchmark/mast-sample/main.py

@@ -0,0 +1,105 @@
+import os
+import argparse
+import time
+
+import numpy as np
+import torch
+import torch.distributed as dist
+from torch._inductor.utils import maybe_profile
+
+
+"""
+Extensible script to measure PT2-D speedup on MAST
+
+Launch locally:
+    torchrun --nnodes=1 --nproc_per_node=8 \
+    userbenchmark/mast-sample/main.py --edir=$(pwd)
+
+Launch on MAST: see instructions in fbcode: run_pt2d_oss_benchmark.sh
+"""
+
+# Provided by torchx
+LOCAL_RANK = int(os.environ["LOCAL_RANK"])
+RANK = int(os.environ["RANK"])
+WORLD_SIZE = int(os.environ["WORLD_SIZE"])
+
+
+def fn():
+    linear = torch.nn.Linear(100, 200, bias=False, device="cuda")
+    x = torch.randn(100, 100, device="cuda")
+
+    # Supports collectives
+    dist.all_reduce(x)
+
+    # Supports functional collectives
+    x = dist._functional_collectives.all_reduce(x, "SUM", list(range(WORLD_SIZE)))
+
+    out = linear(x)
+    return out
+
+def bench(args):
+    # Supports torch.compile
+    compiled_fn = torch.compile(fn)
+
+    eager_times = []
+    compiled_times = []
+    n_repeat = 10
+    for _ in range(n_repeat):
+        torch.cuda.synchronize()
+        t1 = time.perf_counter()
+        fn()
+        torch.cuda.synchronize()
+        t2 = time.perf_counter()
+        compiled_fn()
+        torch.cuda.synchronize()
+        t3 = time.perf_counter()
+
+        eager_times.append(t2 - t1)
+        compiled_times.append(t3 - t2)
+
+    speedups = [e/c for e,c in zip(eager_times, compiled_times)]
+    speedup = np.median(speedups)
+
+    # Supports profile
+    with maybe_profile(True) as p:
+        eager_out = fn()
+        compiled_out = compiled_fn()
+
+    if RANK == 0:
+        print(f"eager_out={eager_out}")
+        print(f"compiled_out={compiled_out}")
+
+        trace_file = os.path.join(args.edir, f"trace_rank_{RANK}.json")
+        print(f"Writing trace to: {trace_file}")
+        p.export_chrome_trace(trace_file)
+
+        print(f"{speedup:.3f}x")
+
+        # Supports writing to a local file
+        output_file = os.path.join(args.edir, f"metrics_rank_{RANK}.csv")
+        print(f"Writing output content to: {output_file}")
+        with open(output_file, 'w') as f:
+            f.write("speedup\n")
+            f.write(f"{speedup}\n")
+
+
+if __name__ == "__main__":
+    assert torch.cuda.is_available()
+
+    dist.init_process_group(backend="nccl")
+    torch.cuda.set_device(LOCAL_RANK)
+    print(f"Hello from rank={RANK}")
+
+    parser = argparse.ArgumentParser(description="arg parser")
+    parser.add_argument(
+        "--edir",
+        type=str,
+        default=None,
+        help="directory to dump job outputs",
+    )
+    args = parser.parse_args()
+    assert args.edir is not None
+
+    bench(args)
+
+    print(f"Bye from rank={RANK}")