fix

tests/monkeypatch/test_moe_grouped.py

@@ -100,6 +100,36 @@ def test_grouped_uses_per_expert_nested_modules(monkeypatch):
     assert not torch.equal(first_call[0], first_call[1])
 
 
+def test_grouped_accepts_module_list_experts(monkeypatch):
+    hidden = 4
+    intermediate = 2
+    experts = nn.ModuleList(
+        [DummyQwenExpert(i, hidden, intermediate) for i in range(2)]
+    )
+    gate = nn.Linear(hidden, len(experts), bias=False)
+    nn.init.zeros_(gate.weight)
+
+    calls = {"count": 0}
+
+    def fake_grouped_mm(As, Bs):
+        calls["count"] += 1
+        return [
+            torch.zeros(a.shape[0], b.shape[-1], device=a.device, dtype=a.dtype)
+            for a, b in zip(As, Bs, strict=False)
+        ]
+
+    monkeypatch.setattr(torch_grouped_module, "_call_grouped_mm", fake_grouped_mm)
+
+    hidden_states = torch.randn(1, 2, hidden)
+    y, router_logits = torch_grouped_module.moe_ffn_forward_grouped(
+        hidden_states, gate, experts, top_k=2
+    )
+
+    assert y is not None
+    assert router_logits is not None
+    assert calls["count"] > 0
+
+
 class _DummyCfg:
     moe_backend = "torch_grouped"