diff --git a/megatron/core/distributed/fsdp/src/megatron_fsdp/param_and_grad_buffer.py b/megatron/core/distributed/fsdp/src/megatron_fsdp/param_and_grad_buffer.py
index 6c907bc25ec..a5ba81124f7 100644
--- a/megatron/core/distributed/fsdp/src/megatron_fsdp/param_and_grad_buffer.py
+++ b/megatron/core/distributed/fsdp/src/megatron_fsdp/param_and_grad_buffer.py
@@ -1404,6 +1404,28 @@ def _does_param_require_new_bucket(param):
 
     is_expert_parameter = lambda n, p: ".experts." in n
 
+    def _should_split_from_grouped_expert_bucket(
+        is_expert_param: bool,
+        param: torch.nn.Parameter,
+        param_csf_base: int,
+        chunk_size_factor: int,
+        same_factor_params: List[torch.nn.Parameter],
+    ) -> bool:
+        """
+        Split grouped expert (>=3D) tensors with heterogeneous CSF bases into
+        separate buckets to avoid LCM-inflated bucket alignment padding.
+        """
+        # Non-expert groups keep the original LCM/fragment merge.
+        if not is_expert_param:
+            return False
+        # Param already aligns with bucket CSF (always true for the first param
+        # after sort); no split needed.
+        if param_csf_base == chunk_size_factor:
+            return False
+        return to_local_if_dtensor(param).dim() >= 3 or any(
+            to_local_if_dtensor(p).dim() >= 3 for p in same_factor_params
+        )
+
     # Step 1: Group the parameters according to their execution order and attributes.
     # FSDP unit module parameters are split into multiple parameter sub-groups.
     # All parameters in the module are assigned a parameter group, even non-FSDP modules.
@@ -1503,17 +1525,27 @@ def _does_param_require_new_bucket(param):
             remaining_params = []
             for param in params:
                 param_shape = to_local_if_dtensor(param).shape
+                param_csf_base = param_shape[1:].numel()
+                if _should_split_from_grouped_expert_bucket(
+                    group.is_expert_param,
+                    param,
+                    param_csf_base,
+                    chunk_size_factor,
+                    same_factor_params,
+                ):
+                    remaining_params.append(param)
+                    continue
                 if (
-                    param_shape[1:].numel() == chunk_size_factor
+                    param_csf_base == chunk_size_factor
                     or (
-                        chunk_size_factor % param_shape[1:].numel() == 0
+                        chunk_size_factor % param_csf_base == 0
                         and param_shape.numel() % chunk_size_factor == 0
                     )
                     or (param_shape.numel() < chunk_size_factor)
                 ):
                     same_factor_params.append(param)
                 else:
-                    lcm_chunk_size_factor = math.lcm(chunk_size_factor, param_shape[1:].numel())
+                    lcm_chunk_size_factor = math.lcm(chunk_size_factor, param_csf_base)
                     chunk_size_factor = lcm_chunk_size_factor
                     same_factor_params.append(param)
             # Create a new parameter group with the same chunk size factor.