[feat] OSS: adding a --profile option to the benchmark (#135)

benchmarks/oss.py

@@ -9,6 +9,7 @@ from typing import Any, List, Optional, cast
 
 import numpy as np
 import torch
+import torch.autograd.profiler as profiler
 import torch.distributed as dist
 import torch.multiprocessing as mp
 import torch.nn as nn
@@ -49,6 +50,13 @@ def get_problem(rank, data_size, batch_size):
     return model, dataloader, loss_fn
 
 
+class OptimType(str, Enum):
+    vanilla = "pytorch"
+    oss = "oss"
+    oss_sdp = "oss_sdp"
+    everyone = "everyone"
+
+
 def train(
     rank: int,
     world_size: int,
@@ -56,14 +64,13 @@ def train(
     batch_size: int = 32,
     data_size: int = 200,
     backend: str = "gloo",
-    use_oss: bool = True,
-    use_sdp: bool = False,
+    optim_type: OptimType = OptimType.vanilla,
+    profile: bool = False,
     check_regression: bool = True,
     reference_speed: float = -1.0,
     reference_memory: float = -1.0,
     reference_loss: float = -1.0,
 ):
-    assert not use_sdp or (use_sdp and use_oss), "ShardedDataParallel requires OSS"
     # DDP
     dist_init(rank=rank, world_size=world_size, backend=backend)
 
@@ -82,7 +89,7 @@ def train(
     # Shard the optimizer
     optimizer: Optional[torch.optim.Optimizer] = None
 
-    if use_sdp:
+    if optim_type == OptimType.oss_sdp:
         ddp = ShardedDataParallel(
             module=model,
             optimizer=OPTIM,
@@ -97,7 +104,7 @@ def train(
         model = DDP(model, device_ids=[rank], find_unused_parameters=True)  # type: ignore
         optimizer = (
             OSS(params=model.parameters(), optim=OPTIM, lr=1e-4, momentum=0.9)
-            if use_oss
+            if optim_type == OptimType.oss
             else OPTIM(model.parameters(), lr=1e-4, momentum=0.9)
         )
 
@@ -111,6 +118,7 @@ def train(
 
     measurements = []
     final_loss: Optional[float] = -1.0
+    need_profiling = profile
 
     for epoch in range(num_epochs):
         epoch_start = time.monotonic()
@@ -124,16 +132,29 @@ def train(
                 loss /= world_size
                 loss.backward()
 
-                if use_sdp:
+                if optim_type == OptimType.oss_sdp:
                     ddp.reduce()  # Send the gradients to the appropriate shards
 
                 return loss
 
-            final_loss = optimizer.step(closure)
+            if need_profiling:
+                print("Profiling the run")
+                with profiler.profile(use_cuda=True) as prof:  # type: ignore
+                    with profiler.record_function("batch"):
+                        final_loss = optimizer.step(closure)
+                        print("profiling done, final loss ", cast(float, final_loss))
+
+                if rank == 0:
+                    prof.export_chrome_trace(f"{optim_type}_trace.json")
+
+                need_profiling = False  # only profile once
+
+            else:
+                final_loss = optimizer.step(closure)
 
         epoch_end = time.monotonic()
 
-        if use_oss:
+        if optim_type == OptimType.oss:
             # Check the checkpointing in the case of the OSS optimizer
             # Memory usage could spill over from there
             optimizer = cast(OSS, optimizer)
@@ -160,7 +181,7 @@ def train(
     std = math.sqrt(sum(diff) / (len(measurements) - 1))
     print(f"[{dist.get_rank()}] : Mean speed: {mean:.2f} +/- {std:.2f}")
 
-    if use_oss and check_regression and dist.get_rank() == 0:
+    if check_regression and dist.get_rank() == 0:
         assert (mean + 3.0 * std) > reference_speed, "Speed regression detected"
         assert max_memory < 1.05 * reference_memory, "Memory use regression detected"
         assert abs(cast(float, final_loss) - reference_loss) < 1e-3, "Loss regression detected"
@@ -171,13 +192,6 @@ def train(
 
 
 if __name__ == "__main__":
-
-    class OptimType(str, Enum):
-        vanilla = "pytorch"
-        oss = "oss"
-        oss_sdp = "oss_sdp"
-        everyone = "everyone"
-
     parser = argparse.ArgumentParser(
         description="Benchmark the optimizer state sharding, on a typical computer vision workload"
     )
@@ -193,6 +207,7 @@ if __name__ == "__main__":
         "--optim_type", type=OptimType, choices=[o.value for o in OptimType], default=OptimType.everyone
     )
     parser.add_argument("--gloo", action="store_true", default=False)
+    parser.add_argument("--profile", action="store_true", default=False)
 
     args = parser.parse_args()
     print(f"Benchmark arguments: {args}")
@@ -209,8 +224,8 @@ if __name__ == "__main__":
                 args.batch_size,
                 args.data_size,
                 backend,
-                False,  # OSS
-                False,  # SDP
+                OptimType.vanilla,
+                args.profile,
                 False,  # no regression check
             ),
             nprocs=args.world_size,
@@ -227,8 +242,8 @@ if __name__ == "__main__":
                 args.batch_size,
                 args.data_size,
                 backend,
-                True,  # OSS
-                False,  # SDP
+                OptimType.oss,
+                args.profile,
                 args.check_regression,
                 args.reference_speed,
                 args.reference_memory,
@@ -248,8 +263,8 @@ if __name__ == "__main__":
                 args.batch_size,
                 args.data_size,
                 backend,
-                True,  # OSS
-                True,  # SDP
+                OptimType.oss_sdp,
+                args.profile,
                 False,  # FIXME: @lefaudeux - SDP should give the same results
                 -1,  # Not checking SDP for speed regression for now, still slower than OSS
                 args.reference_memory,