diff --git a/vllm/model_executor/layers/fused_moe/config.py b/vllm/model_executor/layers/fused_moe/config.py
index 1f077ab80be..231c5652e45 100644
--- a/vllm/model_executor/layers/fused_moe/config.py
+++ b/vllm/model_executor/layers/fused_moe/config.py
@@ -990,6 +990,7 @@ class FusedMoEParallelConfig:
 
     @property
     def use_batched_activation_format(self):
+        # TODO(bnell): nixl also uses batched format
         return self.use_deepep_ll_kernels
 
     @property
diff --git a/vllm/model_executor/layers/fused_moe/runner/moe_runner_base.py b/vllm/model_executor/layers/fused_moe/runner/moe_runner_base.py
index 2d2d0bb4ebe..136e1b1f5b2 100644
--- a/vllm/model_executor/layers/fused_moe/runner/moe_runner_base.py
+++ b/vllm/model_executor/layers/fused_moe/runner/moe_runner_base.py
@@ -427,9 +427,6 @@ class MoERunnerBase(MoERunner):
         via the router, and the actual fused MoE computation. Returns
         (shared_expert_output, fused_expert_output).
         """
-        # Run this before quant_method to avoid inplace issues.
-        # TODO(bnell): probably not needed anymore since inplace is
-        # disabled when shared experts are present.
         self._maybe_apply_shared_experts(
             shared_experts_input, SharedExpertsOrder.NO_OVERLAP
         )
@@ -447,7 +444,7 @@ class MoERunnerBase(MoERunner):
             )
 
             # Passing shared_experts_input in case SharedExpertsOrder is
-            # NO_OVERLAP or MK_INTERNAL_OVERLAPPED.
+            # MK_INTERNAL_OVERLAPPED.
             fused_out = self.quant_method.apply(
                 layer=layer,
                 x=hidden_states,
@@ -490,6 +487,7 @@ class MoERunnerBase(MoERunner):
         #        parallel execution of shared experts with the FusedMoE via
         #        separate cuda stream)
         if self._shared_experts is not None:
+            assert shared_experts_input is not None
             self._shared_experts.maybe_sync_shared_experts_stream(shared_experts_input)
 
     def _maybe_add_zero_expert_output(
diff --git a/vllm/model_executor/layers/fused_moe/runner/shared_experts.py b/vllm/model_executor/layers/fused_moe/runner/shared_experts.py
index 3a08d5d0fc2..c105badabcb 100644
--- a/vllm/model_executor/layers/fused_moe/runner/shared_experts.py
+++ b/vllm/model_executor/layers/fused_moe/runner/shared_experts.py
@@ -80,10 +80,24 @@ class SharedExperts:
                     "Enabled separate cuda stream for MoE shared_experts", scope="local"
                 )
 
+    @property
+    def _disable_shared_experts_overlap(self) -> bool:
+        # Disable shared expert overlap if:
+        #   - we are using eplb with non-default backend, because of correctness issues
+        #   - we are using flashinfer with DP, since there nothing to gain
+        parallel_config = self._moe_config.moe_parallel_config
+        return (
+            parallel_config.enable_eplb
+            and parallel_config.all2all_backend != "allgather_reducescatter"
+        ) or parallel_config.use_fi_nvl_two_sided_kernels
+
     def _determine_shared_experts_order(
         self,
         hidden_states: torch.Tensor,
     ) -> SharedExpertsOrder:
+        if self._disable_shared_experts_overlap:
+            return SharedExpertsOrder.NO_OVERLAP
+
         if self._quant_method.mk_owns_shared_expert:
             return SharedExpertsOrder.MK_INTERNAL_OVERLAPPED