[Bugfix][Hardware][AMD] Fix list aliasing in fused MoE initialization

tests/kernels/moe/test_rocm_aiter_list_aliasing.py

@@ -0,0 +1,112 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""
+Unit tests for ROCm AITER fused MoE list aliasing fix.
+
+This tests the fix for the list aliasing bug where using [list] * n
+creates n references to the same list object, causing unintended
+modifications when any single element is changed.
+
+The fix uses list comprehension [... for _ in range(n)] to create
+independent list copies.
+
+See: https://github.com/vllm-project/vllm/pull/31121
+"""
+
+import pytest
+
+
+class TestListAliasingFix:
+    """Test that list aliasing bug is fixed in MoE initialization."""
+
+    def test_list_multiplication_creates_aliased_references(self):
+        """Demonstrate the bug: [list] * n creates aliased references."""
+        # This is the BUGGY pattern that was used before the fix
+        max_num_tokens = 5
+        fake_expertid = 99
+
+        # Bug: All elements reference the SAME inner list
+        buggy_list = [[fake_expertid] * 3] * max_num_tokens
+
+        # Verify all elements are initially the same
+        assert all(elem == [99, 99, 99] for elem in buggy_list)
+
+        # Modify one element
+        buggy_list[0] = [1, 2, 3]
+
+        # With the buggy pattern and direct assignment, only index 0 changes
+        # But the original bug was when using in-place modification:
+        buggy_list2 = [[fake_expertid] * 3] * max_num_tokens
+        buggy_list2[2][0] = 42  # Modify element at index 2
+
+        # BUG: ALL elements are modified because they reference the same list!
+        for i, elem in enumerate(buggy_list2):
+            assert elem[0] == 42, f"Element {i} should be 42 due to aliasing bug"
+
+    def test_list_comprehension_creates_independent_copies(self):
+        """Verify the fix: list comprehension creates independent copies."""
+        # This is the FIXED pattern using list comprehension
+        max_num_tokens = 5
+        fake_expertid = 99
+
+        # Fix: Each element is an independent list
+        fixed_list = [[fake_expertid] * 3 for _ in range(max_num_tokens)]
+
+        # Verify all elements are initially the same
+        assert all(elem == [99, 99, 99] for elem in fixed_list)
+
+        # Modify one element in-place
+        fixed_list[2][0] = 42
+
+        # Only the modified element should change
+        assert fixed_list[2] == [42, 99, 99], "Modified element should be [42, 99, 99]"
+
+        # Other elements should remain unchanged
+        for i in [0, 1, 3, 4]:
+            assert fixed_list[i] == [99, 99, 99], (
+                f"Element {i} should remain [99, 99, 99]"
+            )
+
+    def test_moe_shared_expert_ids_pattern(self):
+        """Test the actual pattern used in init_aiter_topK_meta_data."""
+        max_num_tokens = 10
+        n_routed_experts = 8
+        n_shared_experts = 2
+        fake_expertid = n_routed_experts + n_shared_experts  # 10
+        is_EP = True
+        tp_rank = 0
+        tp_size = 2
+
+        # Fixed pattern (from the PR)
+        s_topk_ids_list = [
+            [fake_expertid] * (n_shared_experts + is_EP) for _ in range(max_num_tokens)
+        ]
+
+        # Verify initial state
+        expected_initial = [fake_expertid] * (n_shared_experts + is_EP)  # [10, 10, 10]
+        assert all(elem == expected_initial for elem in s_topk_ids_list)
+
+        # Simulate the EP assignment logic
+        shared_expert_ids = [
+            n_routed_experts + i for i in range(n_shared_experts + is_EP)
+        ]  # [8, 9, 10]
+
+        for i in range(tp_rank, max_num_tokens, tp_size):
+            s_topk_ids_list[i] = shared_expert_ids
+
+        # Verify only specific indices were modified
+        for i in range(max_num_tokens):
+            if i % tp_size == tp_rank:
+                # These should have shared_expert_ids
+                assert s_topk_ids_list[i] == shared_expert_ids, (
+                    f"Index {i} should have shared_expert_ids"
+                )
+            else:
+                # These should remain unchanged with fake_expertid
+                assert s_topk_ids_list[i] == expected_initial, (
+                    f"Index {i} should remain unchanged"
+                )
+
+
+if __name__ == "__main__":
+    pytest.main([__file__, "-v"])

vllm/model_executor/layers/fused_moe/rocm_aiter_fused_moe.py

@@ -72,14 +72,14 @@ def init_aiter_topK_meta_data(
     shared_expert_ids = [n_routed_experts + i for i in range(n_shared_experts + is_EP)]
     if is_EP:
         s_topk_ids_list = [
-            [fake_expertid] * (n_shared_experts + is_EP)
-        ] * max_num_tokens
+            [fake_expertid] * (n_shared_experts + is_EP) for _ in range(max_num_tokens)
+        ]
         for i in range(tp_rank, max_num_tokens, tp_size):
             s_topk_ids_list[i] = shared_expert_ids
     else:
         s_topk_ids_list = [
-            list(range(n_routed_experts, fake_expertid))
-        ] * max_num_tokens
+            list(range(n_routed_experts, fake_expertid)) for _ in range(max_num_tokens)
+        ]
     s_topk_ids[:] = torch.tensor(s_topk_ids_list, dtype=torch.int32, device="cuda")
 
     total_topk_weights = torch.empty(