Add support of DDP and experimental CompiledAutograd

Summary:
Address the comments in #319 and resubmit the PR to fit the current code base.

Test Plan:
```
CONFIG_FILE=./train_configs/debug_model.toml ./run_llama_train.sh --comm.train_timeout_seconds=3600   --training.tensor_parallel_degree=1 --training.data_parallel_degree=8 --experimental.data_parallel_type=ddp --training.steps=1000 --metrics.log_freq=10 --profiling.profile_freq=1000
```

ghstack-source-id: 81dc85d
Pull Request resolved: #432

Author: fegin

estimation.py

@@ -71,6 +71,7 @@ def estimate_memory(job_config: JobConfig):
         pp=job_config.experimental.pipeline_parallel_degree,
         world_size=world_size,
         enable_loss_parallel=job_config.training.enable_loss_parallel,
+        dp_type=job_config.training.data_parallel_type,
     )
 
     device = torch.device(f"cuda:{int(os.environ['LOCAL_RANK'])}")

test_runner.py

@@ -304,6 +304,15 @@ def build_test_list():
             ],
             "FSDP2 with float8 all-gather and precomputed dynamic scales",
             "fsdp2_float8_all_gather_precompute_dynamic_scales",
+        ),
+        OverrideDefinitions(
+            [
+                [
+                    "--training.data_parallel_type ddp",
+                ]
+            ],
+            "DDP",
+            "ddp",
             ngpu=4,
         ),
     ]

torchtitan/config_manager.py

@@ -312,6 +312,17 @@ def __init__(self):
                 The default value will be the number of pipeline stages, if unspecified.
             """,
         )
+        self.parser.add_argument(
+            "--training.data_parallel_type",
+            type=str,
+            default="fsdp",
+            help="Data parallelism type. TorchTitan currently supports FSDP and DDP.",
+        )
+        self.parser.add_argument(
+            "--experimental.enable_compiled_autograd",
+            action="store_true",
+            help="Enable CompiledAutograd to compile the backward.",
+        )
         self.parser.add_argument(
             "--training.mixed_precision_param",
             type=str,

torchtitan/parallelisms/__init__.py

@@ -28,8 +28,10 @@ class ParallelDims:
     pp: int
     world_size: int
     enable_loss_parallel: bool
+    dp_type: str
 
     def __post_init__(self):
+        self.dp_type = self.dp_type.lower()
         self._validate()
 
     def _validate(self):
@@ -42,6 +44,7 @@ def _validate(self):
         assert (
             dp * tp * pp == self.world_size
         ), f"Invalid parallel dims: dp({dp}) * tp({tp}) * pp({pp}) != WORLD_SIZE({self.world_size})"
+        assert self.dp_type in ("fsdp", "ddp")
 
     def build_mesh(self, device_type):
         dims = []

torchtitan/parallelisms/parallelize_llama.py

@@ -16,6 +16,8 @@
 from torch.distributed import DeviceMesh
 
 from torch.distributed._composable.fsdp import fully_shard, MixedPrecisionPolicy
+
+from torch.distributed._composable.replicate import replicate
 from torch.distributed._tensor import Replicate, Shard
 from torch.distributed.algorithms._checkpoint.checkpoint_wrapper import (
     checkpoint_wrapper as ptd_checkpoint_wrapper,
@@ -453,13 +455,15 @@ def apply_compile(model: nn.Module, job_config: JobConfig):
     return model
 
 
-def apply_dp(
+def apply_fsdp(
     model: nn.Module,
     world_mesh: DeviceMesh,
     parallel_dims: "ParallelDims",
     job_config: JobConfig,
 ):
-    """Apply data parallelism (FSDP2) to the model."""
+    """
+    Apply data parallelism to the model. FSDP2 is used here.
+    """
 
     dp_mesh = world_mesh["dp"] if world_mesh.ndim > 1 else world_mesh
     assert dp_mesh.mesh_dim_names == ("dp",), dp_mesh.mesh_dim_names
@@ -492,6 +496,29 @@ def apply_dp(
     return model
 
 
+def apply_ddp(
+    model: nn.Module,
+    world_mesh: DeviceMesh,
+    parallel_dims: "ParallelDims",
+    job_config: JobConfig,
+):
+    if world_mesh.ndim > 1:
+        raise RuntimeError("DDP has not supported > 1D parallelism.")
+
+    if job_config.training.compile:
+        if job_config.experimental.enable_compiled_autograd:
+            torch._dynamo.config.optimize_ddp = (
+                "python_reducer_without_compiled_forward"
+            )
+        else:
+            torch._dynamo.config.optimize_ddp = "ddp_optimizer"
+
+    model = replicate(model, device_mesh=world_mesh, bucket_cap_mb=100)
+
+    logger.info("Applied DDP to the model")
+    return model
+
+
 def parallelize_llama(
     model: nn.Module,
     world_mesh: DeviceMesh,
@@ -516,6 +543,9 @@ def parallelize_llama(
         model = apply_compile(model, job_config)
 
     if parallel_dims.dp_enabled:
-        model = apply_dp(model, world_mesh, parallel_dims, job_config)
+        if parallel_dims.dp_type == "fsdp":
+            model = apply_fsdp(model, world_mesh, parallel_dims, job_config)
+        else:
+            model = apply_ddp(model, world_mesh, parallel_dims, job_config)
 
     return model

train.py

@@ -138,6 +138,22 @@ def zero_grad(self):
     return OptimizersContainer([_build_optimizer(model) for model in model_parts])
 
 
+def get_train_context(enable_loss_parallel: bool, enable_compiled_autograd: bool):
+    @contextlib.contextmanager
+    def context():
+        with contextlib.ExitStack() as stack:
+            if enable_loss_parallel:
+                stack.enter_context(loss_parallel())
+            if enable_compiled_autograd:
+                stack.enter_context(
+                    torch._dynamo.utils.maybe_enable_compiled_autograd(True)
+                )
+
+            yield
+
+    return context
+
+
 # Enable debug tracing on failure: https://pytorch.org/docs/stable/elastic/errors.html
 @record
 def main(job_config: JobConfig):
@@ -160,6 +176,7 @@ def main(job_config: JobConfig):
         pp=job_config.experimental.pipeline_parallel_degree,
         world_size=world_size,
         enable_loss_parallel=job_config.training.enable_loss_parallel,
+        dp_type=job_config.training.data_parallel_type,
     )
     device = torch.device(f"cuda:{int(os.environ['LOCAL_RANK'])}")
     torch.cuda.set_device(device)
@@ -194,9 +211,9 @@ def main(job_config: JobConfig):
         dp_rank,
     )
 
-    # loss_parallel enables dispatching to efficient loss operators
-    loss_parallel_ctx = (
-        loss_parallel if parallel_dims.loss_parallel_enabled else contextlib.nullcontext
+    train_context = get_train_context(
+        parallel_dims.loss_parallel_enabled,
+        job_config.experimental.enable_compiled_autograd,
     )
 
     # loss fn can be shared by pipeline-parallel or non-pp execution
@@ -364,7 +381,7 @@ def loss_fn(pred, labels):
                 # pipeline parallel forward / backward inside step() call
                 is_last_stage = pp_mesh.get_local_rank() == pp_mesh.size() - 1
 
-                with loss_parallel_ctx():
+                with train_context():
                     if pp_mesh.get_local_rank() == 0:
                         pp_schedule.step(input_ids)
                     elif is_last_stage:
@@ -381,7 +398,7 @@ def loss_fn(pred, labels):
                 )
             else:
                 # Non-PP forward / backward
-                with loss_parallel_ctx():
+                with train_context():
                     pred = model(input_ids)
                     loss = loss_fn(pred, labels)
                     # pred.shape=(bs, seq_len, vocab_size)