diff --git a/nemo_automodel/components/moe/parallelizer.py b/nemo_automodel/components/moe/parallelizer.py
index de3954337c..2a1c97e14c 100644
--- a/nemo_automodel/components/moe/parallelizer.py
+++ b/nemo_automodel/components/moe/parallelizer.py
@@ -144,6 +144,23 @@ def _get_moe_module(block: nn.Module) -> MoE | None:
             return module
 
 
+def _repeated_mtp_moe_block_ids(model: nn.Module) -> set[int]:
+    """Return weight-tied MTP blocks whose experts cannot be recomputed safely.
+
+    A repeated MTP depth may contain multiple physical sublayers. Only MoE
+    sublayers own the EP-sharded expert parameter group whose second FSDP2
+    checkpoint recompute is unsafe; attention, MLP, and Mamba sublayers remain
+    eligible for activation checkpointing.
+    """
+    mtp_module = getattr(model, "mtp", None)
+    if mtp_module is None or not hasattr(mtp_module, "layers"):
+        return set()
+    mtp_repeated = bool(getattr(getattr(mtp_module, "mtp_config", None), "use_repeated_layer", False))
+    if not mtp_repeated:
+        return set()
+    return {id(block) for block in mtp_module.layers.children() if _get_moe_module(block) is not None}
+
+
 def _preserve_gate_load_during_recompute(
     block: nn.Module,
     context_fn: Callable[[], tuple[AbstractContextManager, AbstractContextManager]] | None = None,
@@ -484,6 +501,12 @@ def apply_ac(
 
     scopes = normalize_activation_checkpointing_scope(activation_checkpointing_scope)
     checkpoint_decoder = "all" in scopes or "language" in scopes
+    repeated_mtp_moe_block_ids = _repeated_mtp_moe_block_ids(model) if checkpoint_decoder else set()
+    if repeated_mtp_moe_block_ids:
+        logger.info(
+            "Skipping activation checkpointing on %d weight-tied MTP MoE block(s)",
+            len(repeated_mtp_moe_block_ids),
+        )
     if checkpoint_decoder and not selective and not ignore_router:
         logger.warning(
             "Activation checkpointing is enabled with ignore_router_for_ac=False. The MoE "
@@ -510,6 +533,8 @@ def apply_ac(
                 selective_context_fn,
             )
             for parent_layers, layer_id, block in iter_transformer_and_mtp_blocks(model):
+                if id(block) in repeated_mtp_moe_block_ids:
+                    continue
                 if bool(getattr(block, "_nemo_disable_activation_checkpointing", False)):
                     logger.info("Skipping activation checkpointing for model-owned eager block %s", layer_id)
                     continue
@@ -595,26 +620,12 @@ def selective_checkpointing_context_fn():
     def _with_attention_backend_snapshot(context_fn=None):
         return functools.partial(transformer_engine_attention_backend_snapshot_context_fn, context_fn)
 
-    # Weight-tied (use_repeated_layer) MTP head blocks must NOT be activation
-    # checkpointed: the single physical block is recomputed once per MTP depth in
-    # backward, and FSDP2 cannot re-unshard the *shared* EP-sharded experts param
-    # group on the 2nd+ recompute (the 1st recompute's post_backward reshards it, and
-    # the 2nd recompute's pre_forward unshard does not re-gather it) -> the experts
-    # weight is read in the resharded Shard(1) state and grouped_gemm raises
-    # "Expected hidden_in == a.size(1)". The MTP head is tiny (1 physical block), so
-    # skipping its recompute costs negligible activation memory. Non-tied MTP heads
-    # (each physical block recomputed exactly once) are unaffected and keep AC.
-    mtp_module = getattr(model, "mtp", None)
-    mtp_block_ids: set[int] = set()
-    mtp_repeated = False
-    if mtp_module is not None and hasattr(mtp_module, "layers"):
-        mtp_block_ids = {id(b) for b in mtp_module.layers.children()}
-        mtp_repeated = bool(getattr(getattr(mtp_module, "mtp_config", None), "use_repeated_layer", False))
-    if mtp_repeated and mtp_block_ids:
-        logger.info("Skipping activation checkpointing on %d weight-tied MTP head block(s)", len(mtp_block_ids))
-
     for parent_layers, layer_id, block in iter_transformer_and_mtp_blocks(model):
-        if mtp_repeated and id(block) in mtp_block_ids:
+        # A weight-tied MoE block is recomputed once per logical MTP depth.
+        # FSDP2 cannot re-unshard its shared EP-sharded experts group after the
+        # first recompute, so leave only those blocks uncheckpointed. Repeated
+        # dense/attention/Mamba blocks have no such expert group and keep AC.
+        if id(block) in repeated_mtp_moe_block_ids:
             continue
         if bool(getattr(block, "_nemo_disable_activation_checkpointing", False)):
             logger.info("Skipping activation checkpointing for model-owned eager block %s", layer_id)
diff --git a/tests/unit_tests/moe/test_parallelizer.py b/tests/unit_tests/moe/test_parallelizer.py
index 039bf2e36d..12bc1126d0 100644
--- a/tests/unit_tests/moe/test_parallelizer.py
+++ b/tests/unit_tests/moe/test_parallelizer.py
@@ -707,16 +707,24 @@ def fake_wrapper(block, preserve_rng_state, determinism_check=None, context_fn=N
     assert len(model.layers.registered) == 2
 
 
-def test_apply_ac_skips_model_owned_eager_block(monkeypatch):
+@pytest.mark.parametrize("selective", [False, True])
+def test_apply_ac_skips_model_owned_eager_block(monkeypatch, selective):
     P = _import_parallelizer_with_stubs(monkeypatch)
+    if selective:
+        dense_stub = types.ModuleType("nemo_automodel.components.distributed.activation_checkpointing")
+        dense_stub.make_selective_checkpoint_context_fn = MagicMock(return_value=object())
+        dense_stub.SELECTIVE_AC_WRAPPER_FLAG = "_nemo_selective_ac"
+        dense_stub.transformer_engine_attention_backend_snapshot_context_fn = lambda context_fn=None: context_fn
+        monkeypatch.setitem(sys.modules, "nemo_automodel.components.distributed.activation_checkpointing", dense_stub)
+
     eager_block = DummyBlock()
     eager_block._nemo_disable_activation_checkpointing = True
-    wrapped_block = object()
+    wrapped_block = types.SimpleNamespace()
     wrapper_mock = MagicMock(return_value=wrapped_block)
     monkeypatch.setattr(P, "ptd_checkpoint_wrapper", wrapper_mock)
 
     model = DummyModel([DummyBlock(), eager_block])
-    P.apply_ac(model, ignore_router=True, hidden_size=7168, num_experts=256)
+    P.apply_ac(model, ignore_router=True, hidden_size=7168, num_experts=256, selective=selective)
 
     wrapper_mock.assert_called_once()
     assert model.layers.registered == {"0": wrapped_block}
@@ -2756,6 +2764,53 @@ def fake_wrapper(block, preserve_rng_state, determinism_check=None, context_fn=N
         assert getattr(w, sentinel_flag, False) is True
 
 
+@pytest.mark.parametrize("selective", [False, True])
+def test_apply_ac_repeated_mtp_checkpoints_dense_blocks_and_skips_moe(monkeypatch, selective):
+    """Weight-tied MTP AC skips only blocks with the unsafe shared experts group."""
+    P = _import_parallelizer_with_stubs(monkeypatch)
+    monkeypatch.setattr(P, "MoE", DummyMoE)
+
+    if selective:
+        sentinel_flag = "_nemo_selective_ac"
+        dense_stub = types.ModuleType("nemo_automodel.components.distributed.activation_checkpointing")
+        dense_stub.make_selective_checkpoint_context_fn = MagicMock(return_value=object())
+        dense_stub.SELECTIVE_AC_WRAPPER_FLAG = sentinel_flag
+        dense_stub.transformer_engine_attention_backend_snapshot_context_fn = lambda context_fn=None: context_fn
+        monkeypatch.setitem(sys.modules, "nemo_automodel.components.distributed.activation_checkpointing", dense_stub)
+
+    wrapped = []
+
+    class _Wrapper:
+        def __init__(self, block):
+            self.block = block
+
+    def fake_wrapper(block, **_kwargs):
+        wrapper = _Wrapper(block)
+        wrapped.append(wrapper)
+        return wrapper
+
+    monkeypatch.setattr(P, "ptd_checkpoint_wrapper", MagicMock(side_effect=fake_wrapper))
+
+    backbone = DummyBlock(mlp=object())
+    mtp_attention = DummyBlock(mlp=object())
+    mtp_moe = DummyBlock(mlp=DummyMoE())
+    model = types.SimpleNamespace(
+        model=DummyModel([backbone]),
+        mtp=types.SimpleNamespace(
+            layers=LayerContainer([mtp_attention, mtp_moe]),
+            mtp_config=types.SimpleNamespace(use_repeated_layer=True),
+        ),
+    )
+
+    P.apply_ac(model, selective=selective, hidden_size=8, num_experts=4)
+
+    assert [wrapper.block for wrapper in wrapped] == [backbone, mtp_attention]
+    assert set(model.model.layers.registered) == {"0"}
+    assert set(model.mtp.layers.registered) == {"0"}
+    if selective:
+        assert all(getattr(wrapper, sentinel_flag, False) for wrapper in wrapped)
+
+
 # ============================================================================
 # Tests for block.moe attribute handling (Step3p5 style models)
 # ============================================================================