diff --git a/frontier/profiling/linear_op/profiling_plan.py b/frontier/profiling/linear_op/profiling_plan.py
--- a/frontier/profiling/linear_op/profiling_plan.py
+++ b/frontier/profiling/linear_op/profiling_plan.py
@@ -112,5 +112,11 @@ def build_profiling_plan(
 
         ffn_sharded_enabled = tp_size in ffn_tp_set
+        if is_moe and not _supports_share_expert(model_config):
+            # Routed experts are profiled by the MoE profiler.  Keeping the
+            # dense FFN surrogate enabled executes an unprofiled fake weight
+            # path and can reject otherwise-valid attention TP layouts (for
+            # example Qwen3-235B TP8 with 128x128 block FP8 weights).
+            ffn_sharded_enabled = False
 
     padded_n_embd = model_config.embedding_dim
     padded_n_expanded_embd = model_config.mlp_hidden_dim
diff --git a/tests/unit/test_moe_linear_profiling_plan.py b/tests/unit/test_moe_linear_profiling_plan.py
new file mode 100644
--- /dev/null
+++ b/tests/unit/test_moe_linear_profiling_plan.py
@@ -0,0 +1,52 @@
+from types import SimpleNamespace
+
+from frontier.profiling.linear_op.profiling_plan import build_profiling_plan
+
+
+def test_routed_moe_plan_does_not_execute_dense_ffn_surrogate() -> None:
+    model_config = SimpleNamespace(
+        no_tensor_parallel=False,
+        embedding_dim=4096,
+        mlp_hidden_dim=1536,
+        num_q_heads=64,
+        num_kv_heads=4,
+        model_type="qwen3_moe",
+        post_attn_norm=True,
+        is_moe=True,
+        is_step2_mini=False,
+    )
+
+    plan = build_profiling_plan(
+        model_config=model_config,
+        tp_size=8,
+        attn_tp=[8],
+        ffn_tp=[8],
+        disable_replicated=False,
+        is_moe=True,
+    )
+
+    assert plan["attn_sharded_enabled"] is True
+    assert plan["ffn_sharded_enabled"] is False
+    assert plan["ffn_enabled"] is True
+    assert plan["padded_n_expanded_embd"] == 1536
+    assert "attn_pre_proj" in plan["enabled_ops"]
+    assert "post_attention_layernorm" in plan["enabled_ops"]
+    assert "mlp_up_proj" not in plan["enabled_ops"]
+    assert "mlp_down_proj" not in plan["enabled_ops"]
+
+
+def test_dense_plan_still_executes_ffn_surrogate() -> None:
+    model_config = SimpleNamespace(
+        no_tensor_parallel=False,
+        embedding_dim=4096,
+        mlp_hidden_dim=1536,
+        num_q_heads=64,
+        num_kv_heads=4,
+        model_type="qwen2",
+        post_attn_norm=True,
+        is_moe=False,
+        is_step2_mini=False,
+    )
+    plan = build_profiling_plan(model_config, 8, [8], [8], is_moe=False)
+    assert plan["ffn_sharded_enabled"] is True
+    assert "mlp_up_proj" in plan["enabled_ops"]