add ddp test

ghstack-source-id: 18074e0f320faf1065e1ca101bbb820d9644b7a7
Pull Request resolved: #1114

test/test_composability.py

@@ -18,6 +18,7 @@
 )
 from torch.distributed._tensor import DTensor
 from torch.distributed.device_mesh import DeviceMesh, init_device_mesh
+from torch.nn.parallel import DistributedDataParallel as DDP
 
 # torch.testing._internal.common_distributed requies "expecttest"
 from torch.testing._internal.common_distributed import MultiProcessTestCase
@@ -198,10 +199,12 @@ def test_manual_pipeline_with_manual_allreduce(self):
         ddp_pp_model.all_reduce(num_microbatches)
         print(f"{self.rank} finished all_reduce")
 
+    # pytest test/test_composability.py -vsk test_manual_pipeline_with_data_parallel_dp_type_DDP
+    @parametrize("dp_type", ["DDP", "FSDP"])
     @parametrize(
         "schedule_name", ["gpipe", "1f1b", "looped_bfs", "interleaved_1f1b"]
     )
-    def test_manual_pipeline_with_fsdp(self, schedule_name):
+    def test_manual_pipeline_with_data_parallel(self, dp_type, schedule_name):
         device_mesh, device = self._init_device_mesh(
             mesh_shape=(2, 2), mesh_dim_names=("dp", "pp")
         )
@@ -229,25 +232,32 @@ def build_stage(stage_idx, num_stages):
             )
             partial_model.to(device)
 
-            # apply FSDP
-            mp_policy = MixedPrecisionPolicy(
-                # TODO(whc) need to fix PP + FSDP-mixed-precision
-                # tracer for PP assumes f32 and is caught off guard when runtime FSDP interacts using bf16 inputs
-                # param_dtype=torch.bfloat16, reduce_dtype=torch.float32
-                param_dtype=torch.float32,
-                reduce_dtype=torch.float32,
-            )
-            fsdp_config = {"mesh": dp_mesh, "mp_policy": mp_policy}
-            for layer in partial_model.children():
-                fully_shard(
-                    layer,
-                    **fsdp_config,
-                    reshard_after_forward=False,
+            if dp_type == "FSDP":
+                # apply FSDP
+                mp_policy = MixedPrecisionPolicy(
+                    # TODO(whc) need to fix PP + FSDP-mixed-precision
+                    # tracer for PP assumes f32 and is caught off guard when runtime FSDP interacts using bf16 inputs
+                    # param_dtype=torch.bfloat16, reduce_dtype=torch.float32
+                    param_dtype=torch.float32,
+                    reduce_dtype=torch.float32,
+                )
+                fsdp_config = {"mesh": dp_mesh, "mp_policy": mp_policy}
+                for layer in partial_model.children():
+                    fully_shard(
+                        layer,
+                        **fsdp_config,
+                        reshard_after_forward=False,
+                    )
+                data_parallel_model = fully_shard(partial_model, **fsdp_config)
+            elif dp_type == "DDP":
+                data_parallel_model = DDP(
+                    partial_model, process_group=dp_mesh.get_group()
                 )
-            fsdp_model = fully_shard(partial_model, **fsdp_config)
+            else:
+                raise RuntimeError(f"unsupported dp type {dp_type}")
 
             stage = self._create_manual_pipeline_stage(
-                fsdp_model,
+                data_parallel_model,
                 stage_idx,
                 num_stages,
                 device,
@@ -339,14 +349,23 @@ def build_stage(stage_idx, num_stages):
         # Validate that whichever weights we have locally match that part of our local/full ref model
         # (we force FSDP's grads to be all-gathered (.full_tensor) to make it simpler)
         ref_parameters = dict(ref_model.named_parameters())
-        for partial_model, offset in zip(partial_models, offsets):
-            for name, p in partial_model.named_parameters():
-                parts = name.split(".")
-                parts[0] = str(int(parts[0]) + offset)
-                name = ".".join(parts)
-                ref_p = ref_parameters[name]
-                self.assertTrue(isinstance(p.grad, DTensor))
-                self.assertEqual(ref_p.grad, p.grad.full_tensor())
+        if dp_type == "FSDP":
+            for partial_model, offset in zip(partial_models, offsets):
+                for name, p in partial_model.named_parameters():
+                    parts = name.split(".")
+                    parts[0] = str(int(parts[0]) + offset)
+                    name = ".".join(parts)
+                    ref_p = ref_parameters[name]
+                    self.assertTrue(isinstance(p.grad, DTensor))
+                    self.assertEqual(ref_p.grad, p.grad.full_tensor())
+        elif dp_type == "DDP":
+            for partial_model, offset in zip(partial_models, offsets):
+                for name, p in partial_model.named_parameters():
+                    parts = name.split(".")[1:]  # remove the "module." prefix
+                    parts[0] = str(int(parts[0]) + offset)
+                    name = ".".join(parts)
+                    ref_p = ref_parameters[name]
+                    self.assertEqual(ref_p.grad, p.grad)
 
 
 instantiate_parametrized_tests(TestPipelineComposability)