More improvements to FSDP, benchmark against DDP (#13)

* leave root params in mem

* fix logic

* don't show mem usage all the time

* fix

* make configurable

* add alloc/free for unsharded data

* add alloc/free for unsharded grad

* fix

* record for

* revert 69d74c4 - alloc/free for unsharded grad

* revert alloc/free trick for unsharded params data

* add support for DDP in benchmark

* set device ids explicitly

* fix

* change up how weights are initialized

* fix test

* Add back alloc/free hack for unsharded data

* Revert "Add back alloc/free hack for unsharded data"

This reverts commit 0386841.

* Handle frozen layers with reshard-only post-backward hook

* Revert "Handle frozen layers with reshard-only post-backward hook"

This reverts commit 0f408d2.

* add to test

* add to test

* Fixes for frozen modules

* Divide grad before and after reducing for stability

* Add support for hybrid sharding

* make grad clipping optional

* clean up

* calculate grad norm more efficiently

* Revert "calculate grad norm more efficiently"

This reverts commit d66a683.

* fix

Author: epwalsh

docs/source/distributed/fsdp.rst

@@ -2,5 +2,5 @@
 ====================
 
 .. automodule:: olmo_core.distributed.fsdp
-   :members: FSDP, FSDPPrecision
+   :members: FSDP, FSDPPrecision, FSDPShardingStrategy
    :member-order: bysource

src/benchmarks/fsdp/common.py

@@ -151,7 +151,7 @@ def build_components(
     config: TransformerConfig,
     batch_size: int,
     num_batches: int = 100,
-    fsdp_wrapper: Literal["torch", "olmo_core"] = "olmo_core",
+    fsdp_wrapper: Literal["torch", "olmo_core", "ddp"] = "olmo_core",
     wrap_blocks: bool = True,
     mixed_precision: bool = True,
     max_prefetch_count: int = 1,
@@ -204,6 +204,11 @@ def auto_wrap_policy(module: nn.Module, recurse: bool, *args, **kwargs) -> bool:
         )
 
         model.apply(init_function)  # just in case
+    elif fsdp_wrapper == "ddp":
+        from torch.nn.parallel import DistributedDataParallel as DDP
+
+        model = DDP(model.cuda(), device_ids=[dist.get_rank()])
+        model.apply(init_function)
     else:
         raise NotImplementedError(fsdp_wrapper)
 

src/benchmarks/fsdp/train.py

@@ -13,6 +13,7 @@
 
 import torch
 import torch.distributed as dist
+from torch.nn.utils import clip_grad_norm_
 
 from olmo_core.distributed.checkpoint import (
     load_model_and_optim_state,
@@ -28,13 +29,14 @@ def main(
     config: TransformerConfig,
     batch_size: int,
     num_batches: int = 100,
-    fsdp_wrapper: Literal["torch", "olmo_core"] = "olmo_core",
+    fsdp_wrapper: Literal["torch", "olmo_core", "ddp"] = "olmo_core",
     dry_run: bool = False,
     save_path: Optional[str] = None,
     load_path: Optional[str] = None,
     mixed_precision: bool = True,
     profile: bool = False,
     trace_output: str = "/tmp/traces/olmo_core.chrome_trace.json.gz",
+    max_grad_norm: Optional[float] = None,
     **kwargs,
 ):
     model, optim, dataloader = build_components(
@@ -98,21 +100,28 @@ def on_trace_ready(p):
             loss.backward()
 
             # Clip gradient norms.
-            model.clip_grad_norm_(1.0)
+            norm: Optional[torch.Tensor] = None
+            if max_grad_norm is not None:
+                if hasattr(model, "clip_grad_norm_"):
+                    norm = model.clip_grad_norm_(max_grad_norm)
+                else:
+                    norm = clip_grad_norm_(model.parameters(), max_grad_norm)
 
             # Take optimizer step.
             optim.step()
 
             batch_time = time.monotonic() - batch_start
             if i > 0:
                 batch_times.append(batch_time)
+            norm_str = f"{norm.item():.3f}" if norm is not None else "n/a"
             print_rank0(
                 f"Batch [{i+1}/{num_batches}]:\n"
                 f"  loss={loss.item():.3f}\n"
-                f"  throughput/seconds_per_batch={batch_time:.3f}",
+                f"  throughput/seconds_per_batch={batch_time:.3f}\n"
+                f"  grad/total_norm={norm_str}"
             )
 
-            if i == 2:
+            if profile and i == 2:
                 print_rank0(torch.cuda.memory_summary())
 
             if p is not None:
@@ -134,7 +143,7 @@ def on_trace_ready(p):
     parser = argparse.ArgumentParser(prog="train.py", description="Train an FSDP model")
     parser.add_argument(
         "--fsdp",
-        choices=["torch", "olmo_core"],
+        choices=["torch", "olmo_core", "ddp"],
         default="olmo_core",
         help="""The FSDP implementation.""",
     )
@@ -190,6 +199,10 @@ def on_trace_ready(p):
         type=int,
         default=1,
     )
+    parser.add_argument(
+        "--max-grad-norm",
+        type=float,
+    )
     parser.add_argument(
         "--lr",
         type=float,
@@ -237,5 +250,6 @@ def on_trace_ready(p):
         mixed_precision=mixed_precision,
         max_prefetch_count=args.max_prefetch_count,
         learning_rate=args.lr,
+        max_grad_norm=args.max_grad_norm,
         seed=args.seed,
     )

src/olmo_core/distributed/fsdp/__init__.py

@@ -77,6 +77,6 @@
 -------------
 """
 
-from .fsdp import FSDP, FSDPDebugConfig, FSDPPrecision
+from .fsdp import FSDP, FSDPDebugConfig, FSDPPrecision, FSDPShardingStrategy
 
-__all__ = ["FSDP", "FSDPDebugConfig", "FSDPPrecision"]
+__all__ = ["FSDP", "FSDPDebugConfig", "FSDPPrecision", "FSDPShardingStrategy"]

src/olmo_core/distributed/fsdp/flat_param_handle.py

@@ -15,7 +15,11 @@
     ShardedFlatTensor,
     ShardingSpec,
 )
-from olmo_core.distributed.utils import get_rank, get_world_size
+from olmo_core.distributed.utils import (
+    get_gradient_divide_factor,
+    get_rank,
+    get_world_size,
+)
 from olmo_core.stream import Stream
 from olmo_core.utils import get_default_device
 
@@ -62,21 +66,41 @@ class FlatParamHandle:
     """
 
     process_group: Optional[dist.ProcessGroup] = None
+    """
+    Process group containing all shards.
+    """
+
+    inter_group_process_group: Optional[dist.ProcessGroup] = None
+    """
+    Process group for between-group reductions with hybrid sharding.
+    """
 
     device: Optional[torch.device] = None
 
     requires_grad: bool = True
 
+    pre_reduce_grad_divide_factor: float = 1.0
+
+    post_reduce_grad_divide_factor: float = 1.0
+
     _ran_pre_unshard: bool = False
 
     _ran_pre_reduce_scatter_grads: bool = False
 
+    def __post_init__(self):
+        data_parallel_world_size = get_world_size(self.process_group)
+        if self.inter_group_process_group is not None:
+            data_parallel_world_size *= self.inter_group_process_group.size()
+        self.pre_reduce_grad_divide_factor = get_gradient_divide_factor(data_parallel_world_size)
+        self.post_reduce_grad_divide_factor = data_parallel_world_size / self.pre_reduce_grad_divide_factor
+
     @classmethod
     def shard_params(
         cls,
         params: Iterable[nn.Parameter],
         param_fqns: Iterable[str],
         process_group: Optional[dist.ProcessGroup] = None,
+        inter_group_process_group: Optional[dist.ProcessGroup] = None,
         device: Optional[torch.device] = None,
     ) -> FlatParamHandle:
         """
@@ -183,6 +207,7 @@ def shard_params(
                     )
             else:
                 flat_param = ShardedFlatParameter(torch.empty(0, device=device))
+            flat_param.requires_grad = param.requires_grad
             flat_param.mark_as_sharded(sharding_spec, process_group=process_group)
 
             flat_params.append(flat_param)
@@ -224,6 +249,7 @@ def shard_params(
             param_fqns=list(param_fqns),
             params_data=params_data,
             process_group=process_group,
+            inter_group_process_group=inter_group_process_group,
             device=device,
             requires_grad=requires_grad,
         )
@@ -253,7 +279,7 @@ def pre_unshard_(self, dtype: Optional[torch.dtype] = None, rank0_only: bool = F
             self.params_sharded_data_lp.copy_(self.params_data.sharded_data)
 
         # Initialize unsharded, padded gradient.
-        if set_grads and self.params_unsharded_grad is None:
+        if set_grads and self.requires_grad and self.params_unsharded_grad is None:
             self.params_unsharded_grad = torch.zeros_like(all_params_unsharded_data)
 
     def unshard_(
@@ -288,7 +314,7 @@ def unshard_(
         if rank0_only or dist.get_backend() == dist.Backend.GLOO:
             assert self.params_data.is_sharded
             self.params_data.unshard_(dtype=dtype, rank0_only=rank0_only)
-            if set_grads:
+            if set_grads and self.requires_grad:
                 self.params_unsharded_grad = torch.zeros_like(self.params_data)
         else:
             assert not self.params_data.is_sharded
@@ -318,7 +344,7 @@ def unshard_(
 
             param.unshard_(unsharded_data, dtype=dtype, rank0_only=rank0_only)
 
-            if set_grads:
+            if set_grads and self.requires_grad:
                 if param.grad is None and self.params_sharded_grad is not None:
                     self.params_sharded_grad = None
                 assert self.params_unsharded_grad is not None
@@ -360,6 +386,9 @@ def pre_reduce_scatter_grads_(
             Stream.current(self.device).record_for(self.params_unsharded_grad)
             self.params_unsharded_grad = self.params_unsharded_grad.to(dtype=grad_reduce_dtype)
 
+        if self.pre_reduce_grad_divide_factor > 1.0:
+            self.params_unsharded_grad.div_(self.pre_reduce_grad_divide_factor)
+
     def reduce_scatter_grads_(
         self, grad_dtype: Optional[torch.dtype] = None, grad_reduce_dtype: Optional[torch.dtype] = None
     ):
@@ -368,6 +397,7 @@ def reduce_scatter_grads_(
         parameter as a view into the new sharded grad.
         """
         if not self.requires_grad or self.params_unsharded_grad is None:
+            self._ran_pre_reduce_scatter_grads = False
             return
 
         if not self._ran_pre_reduce_scatter_grads:
@@ -398,12 +428,20 @@ def post_reduce_scatter_grads_(
         """
         Finalize sharded gradients after the reduce-scatter.
         """
+        if not self.requires_grad or self.params_unsharded_grad is None:
+            return
+
         grad_dtype = grad_dtype or self.params_data.dtype
         grad_reduce_dtype = grad_reduce_dtype or grad_dtype
 
-        assert self.params_unsharded_grad is not None
         new_sharded_grad = self.params_data.sharded_chunk(self.params_unsharded_grad)
 
+        if self.inter_group_process_group is not None:
+            dist.all_reduce(new_sharded_grad, group=self.inter_group_process_group)
+
+        if self.post_reduce_grad_divide_factor > 1.0:
+            new_sharded_grad.div_(self.post_reduce_grad_divide_factor)
+
         # Cast the new sharded gradient to the right dtype, potentially accumulating it into
         # the existing sharded gradient.
         if self.params_sharded_grad is None: