Register NVLink-domain AG groups before IB-domain HSDP group in UBR

Signed-off-by: jeffnvidia <jmahou@nvidia.com>

Author: jeffnvidia

megatron/core/distributed/fsdp/src/megatron_fsdp/param_and_grad_buffer.py

@@ -1691,9 +1691,6 @@ def __init__(
             if self.dist_index.get_fsdp_group(is_expert_parallel=True) is not None:
                 # Expert-DP group when using EP
                 self.ubr_groups.append(self.dist_index.get_fsdp_group(is_expert_parallel=True))
-            if self.dist_index.get_outer_fsdp_group() is not None:
-                # Outer/Inter-FSDP group when using hybrid FSDP
-                self.ubr_groups.append(self.dist_index.get_outer_fsdp_group())
             if (
                 self.dist_index.get_fsdp_group(
                     is_expert_parallel=False, independent_all_gather=True
@@ -1716,6 +1713,9 @@ def __init__(
                         is_expert_parallel=True, independent_all_gather=True
                     )
                 )
+            if self.dist_index.get_outer_fsdp_group() is not None:
+                # Outer/Inter-FSDP group when using hybrid FSDP (IB domain, registered last).
+                self.ubr_groups.append(self.dist_index.get_outer_fsdp_group())
 
             log_single_rank(
                 logger,