use correct PG when collecting metrics with HYBRID shard (#551)

olmo/optim.py

@@ -39,7 +39,10 @@ def _clean_param_name(self, name: str) -> str:
 
     @torch.no_grad()
     def clip_grads_and_collect_metrics(
-        self, global_step: int, collect_param_metrics: bool = True
+        self,
+        global_step: int,
+        collect_param_metrics: bool = True,
+        process_group: Optional[dist.ProcessGroup] = None,
     ) -> Dict[str, torch.Tensor]:
         """
         Clips gradients for every group that has the field `max_grad_norm`.
@@ -69,6 +72,10 @@ def clip_grads_and_collect_metrics(
         per_param_avg_metric_names: List[str] = []
         per_param_norm_metric_names: List[str] = []
 
+        dst_rank = 0
+        if process_group is not None:
+            dst_rank = dist.get_global_rank(process_group, 0)
+
         # Collect metrics locally.
         for group in self.param_groups:
             if is_distributed():
@@ -144,12 +151,12 @@ def is_grad_norm_metric(metric_name: str) -> bool:
             # Reduce mins.
             if per_param_min_metrics:
                 all_mins = torch.cat(per_param_min_metrics).to(device)
-                dist.reduce(all_mins, 0, op=dist.ReduceOp.MIN)
+                dist.reduce(all_mins, dst_rank, op=dist.ReduceOp.MIN, group=process_group)
                 per_param_min_metrics = all_mins.split(1)
             # Reduce maxs.
             if per_param_max_metrics:
                 all_maxs = torch.cat(per_param_max_metrics).to(device)
-                dist.reduce(all_maxs, 0, op=dist.ReduceOp.MAX)
+                dist.reduce(all_maxs, dst_rank, op=dist.ReduceOp.MAX, group=process_group)
                 per_param_max_metrics = all_maxs.split(1)
             # Reduce sums or just norms.
             all_norms = torch.cat(per_param_norm_metrics).to(device) ** 2.0
@@ -159,13 +166,13 @@ def is_grad_norm_metric(metric_name: str) -> bool:
                 all_sums_norms_numels = torch.cat(
                     [all_sums.unsqueeze(0), all_norms.unsqueeze(0), all_numels.unsqueeze(0)], dim=0
                 )
-                dist.all_reduce(all_sums_norms_numels, op=dist.ReduceOp.SUM)
+                dist.all_reduce(all_sums_norms_numels, op=dist.ReduceOp.SUM, group=process_group)
                 all_sums, all_norms, all_numels = all_sums_norms_numels.split(1)
                 # Get averages.
                 # NOTE: could get infs for non-rank0 processes but that's okay.
                 per_param_avg_metrics = (all_sums / all_numels).squeeze(0).split(1)
             else:
-                dist.all_reduce(all_norms, op=dist.ReduceOp.SUM)
+                dist.all_reduce(all_norms, op=dist.ReduceOp.SUM, group=process_group)
             grad_norm_metric_mask = torch.tensor(
                 [float(is_grad_norm_metric(n)) for n in per_param_norm_metric_names], device=all_norms.device
             )
@@ -325,8 +332,10 @@ def _do_global_fixed_clipping(
                 p.grad.detach().mul_(clip_coef_clamped.to(p.grad.device, p.grad.dtype))
         return num_grads_clipped
 
-    def get_post_step_metrics(self, module: nn.Module) -> Dict[str, torch.Tensor]:
-        del module
+    def get_post_step_metrics(
+        self, module: nn.Module, process_group: Optional[dist.ProcessGroup] = None
+    ) -> Dict[str, torch.Tensor]:
+        del module, process_group
         return {}
 
     def get_state_for_param(self, param: nn.Parameter) -> Dict[str, Optional[torch.Tensor]]:
@@ -356,7 +365,9 @@ def __init__(
         self._update_total_norm: Optional[torch.Tensor] = None
         self._signed_update_total_norm: Optional[torch.Tensor] = None
 
-    def get_post_step_metrics(self, module: nn.Module) -> Dict[str, torch.Tensor]:
+    def get_post_step_metrics(
+        self, module: nn.Module, process_group: Optional[dist.ProcessGroup] = None
+    ) -> Dict[str, torch.Tensor]:
         update_total_dot_prod = self._update_total_dot_prod
         update_total_norm = self._update_total_norm
         signed_update_total_norm = self._signed_update_total_norm
@@ -370,7 +381,11 @@ def get_post_step_metrics(self, module: nn.Module) -> Dict[str, torch.Tensor]:
             # Reduce all together to avoid multiple communication calls.
             all_together = torch.stack([update_total_dot_prod, update_total_norm, signed_update_total_norm])
             # Only need the final result on rank0, since that's where we log from.
-            dist.reduce(all_together, 0)
+            dist.reduce(
+                all_together,
+                0 if process_group is None else dist.get_global_rank(process_group, 0),
+                group=process_group,
+            )
             update_total_dot_prod, update_total_norm, signed_update_total_norm = all_together
             update_total_norm = update_total_norm**0.5
             signed_update_total_norm = signed_update_total_norm**0.5

olmo/train.py

@@ -704,7 +704,11 @@ def train_step(self, batch: Dict[str, Any], reduce_global_loss: bool = True) ->
         # Clip gradient norms and collect param/gradient/optim metrics.
         should_log_optim_metrics_this_step = self.should_log_optim_metrics_this_step()
         optim_metrics = self.optim.clip_grads_and_collect_metrics(
-            self.global_step, collect_param_metrics=should_log_optim_metrics_this_step
+            self.global_step,
+            collect_param_metrics=should_log_optim_metrics_this_step,
+            # passing this process group here ensures metrics are reduced correctly when we're using
+            # HYBRID sharding.
+            process_group=self.fsdp_model.process_group,
         )
 
         # Adjust the learning rate.
@@ -742,7 +746,9 @@ def train_step(self, batch: Dict[str, Any], reduce_global_loss: bool = True) ->
 
         # Maybe collect post-step optimizer-specific metrics.
         if should_log_optim_metrics_this_step:
-            optim_metrics = self.optim.get_post_step_metrics(self.fsdp_model)
+            optim_metrics = self.optim.get_post_step_metrics(
+                self.fsdp_model, process_group=self.fsdp_model.process_group
+            )
             for key, value in optim_metrics.items():
                 metrics[f"optim/{key}"] = value.item()
 

scripts/train.py

@@ -154,7 +154,7 @@ def dummy_init_fn(module: torch.nn.Module) -> None:
             raise OLMoConfigurationError("fsdp.hybrid_sharding_num_model_replicas must be a positive integer")
 
         num_nodes = get_world_size() // get_local_world_size()
-        if num_nodes % num_model_replicas != 0:
+        if num_nodes > 1 and num_nodes % num_model_replicas != 0:
             raise OLMoConfigurationError("fsdp.hybrid_sharding_num_model_replicas must divide number of nodes")
 
         device_mesh = init_device_mesh("cuda", (num_model_replicas, get_world_size() // num_model_replicas))