[Bugfix][MoE] Keep the gate's hash_indices_table out of RoutedExperts parameters (#61930)

Signed-off-by: khluu <khluu000@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

vllm/model_executor/layers/fused_moe/layer.py

@@ -439,8 +439,6 @@ def FusedMoEFactory(
         hash_indices_table=hash_indices_table,
         **routed_experts_args if routed_experts_args is not None else {},
     )
-    if hash_indices_table is not None:
-        object.__setattr__(routed_experts, "hash_indices_table", hash_indices_table)
 
     if runner_cls is None:
         runner_cls = MoERunner

vllm/model_executor/layers/fused_moe/routed_experts.py

@@ -79,7 +79,7 @@ class RoutedExperts(PluggableLayer):
     - Executing routed experts via quant_method.apply()
     """
 
-    # Hash-routing table, owned by the gate; set by FusedMoEFactory.
+    # Hash-routing table, owned by the gate.
     hash_indices_table: torch.Tensor | None = None
 
     def __init__(
@@ -156,7 +156,8 @@ class RoutedExperts(PluggableLayer):
         # Set by bind_routed_experts_capturer for monolithic kernels.
         self.routing_sink: RoutedExpertsSink | None = None
         self.apply_router_weight_on_input = apply_router_weight_on_input
-        self.hash_indices_table = hash_indices_table
+        # The gate owns this Parameter: keep it out of this module's parameters.
+        object.__setattr__(self, "hash_indices_table", hash_indices_table)
         # End random parameters
         self._loaded_expert_biases: set[str] = set()
         self._fused_shared_expert_quantizer = self._get_fused_shared_expert_quantizer()