From 92957809e473e73f12c5eb9622936d177cfa3f93 Mon Sep 17 00:00:00 2001
From: Linxiushen <1787979356@qq.com>
Date: Mon, 21 Sep 2026 14:49:07 +0800
Subject: [PATCH] fix: import GLM5 fused experts under transformers>=5.13

transformers 5.13 renamed GlmMoeDsaNaiveMoe to GlmMoeDsaExperts (same class
body). angelslim/models/llm/glm5_1.py imported the old name unconditionally
and angelslim/models/llm/__init__.py imports GLM5_1 eagerly, so with any
transformers release from 5.13 up to the declared <6.0 ceiling every CLI and
`from angelslim import Engine` failed at import time, even for non-GLM
models.

Fall back to the new name when the old one is missing, and add a test that
the adapter imports and that GlmExpertsWithLinear reproduces the fused
experts module's forward on the installed transformers.

Fixes #382
---
 angelslim/models/llm/glm5_1.py    | 10 ++++-
 tests/test_glm5_experts_compat.py | 71 +++++++++++++++++++++++++++++++
 2 files changed, 80 insertions(+), 1 deletion(-)
 create mode 100644 tests/test_glm5_experts_compat.py

diff --git a/angelslim/models/llm/glm5_1.py b/angelslim/models/llm/glm5_1.py
index d8eca745..29c4ef95 100644
--- a/angelslim/models/llm/glm5_1.py
+++ b/angelslim/models/llm/glm5_1.py
@@ -38,7 +38,15 @@
 import torch.distributed as dist
 import torch.nn as nn
 from transformers import AutoConfig, AutoModelForCausalLM, AutoTokenizer
-from transformers.models.glm_moe_dsa.modeling_glm_moe_dsa import GlmMoeDsaNaiveMoe
+
+try:
+    from transformers.models.glm_moe_dsa.modeling_glm_moe_dsa import GlmMoeDsaNaiveMoe
+except ImportError:
+    # transformers>=5.13 renamed the fused-experts module to GlmMoeDsaExperts;
+    # the class body is unchanged (3-D gate_up_proj / down_proj, same forward).
+    from transformers.models.glm_moe_dsa.modeling_glm_moe_dsa import (
+        GlmMoeDsaExperts as GlmMoeDsaNaiveMoe,
+    )
 
 from ...compressor.quant.core import PTQSaveVllmHF
 from ...utils.utils import find_parent_layer_and_sub_name, print_info
diff --git a/tests/test_glm5_experts_compat.py b/tests/test_glm5_experts_compat.py
new file mode 100644
index 00000000..1f2c0cab
--- /dev/null
+++ b/tests/test_glm5_experts_compat.py
@@ -0,0 +1,71 @@
+# Copyright 2025 Tencent Inc.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""GLM5 adapter must import and wrap the fused experts module on every
+transformers release in the supported range: the class was renamed from
+``GlmMoeDsaNaiveMoe`` to ``GlmMoeDsaExperts`` in transformers 5.13."""
+
+import importlib
+
+import pytest
+import torch
+from transformers.models.glm_moe_dsa import modeling_glm_moe_dsa
+
+from angelslim.models.llm import glm5_1
+
+
+def _fused_experts_class():
+    return getattr(
+        modeling_glm_moe_dsa,
+        "GlmMoeDsaNaiveMoe",
+        getattr(modeling_glm_moe_dsa, "GlmMoeDsaExperts", None),
+    )
+
+
+def test_glm5_module_imports_and_resolves_fused_experts_class():
+    assert importlib.import_module("angelslim.models.llm") is not None
+    assert glm5_1.GlmMoeDsaNaiveMoe is _fused_experts_class()
+
+
+def test_experts_with_linear_matches_fused_experts_forward():
+    from transformers.models.glm_moe_dsa.configuration_glm_moe_dsa import (
+        GlmMoeDsaConfig,
+    )
+
+    torch.manual_seed(0)
+    config = GlmMoeDsaConfig(
+        hidden_size=8, moe_intermediate_size=6, num_local_experts=4, hidden_act="silu"
+    )
+    fused = _fused_experts_class()(config)
+    with torch.no_grad():
+        fused.gate_up_proj.normal_()
+        fused.down_proj.normal_()
+    assert glm5_1._is_glm_naive_moe(fused)
+
+    linearized = glm5_1.GlmExpertsWithLinear(fused)
+
+    tokens, top_k = 5, 2
+    hidden_states = torch.randn(tokens, config.hidden_size)
+    top_k_index = torch.stack(
+        [torch.randperm(config.num_local_experts)[:top_k] for _ in range(tokens)]
+    )
+    top_k_weights = torch.rand(tokens, top_k)
+    with torch.no_grad():
+        expected = fused(hidden_states, top_k_index, top_k_weights)
+        actual = linearized(hidden_states, top_k_index, top_k_weights)
+    torch.testing.assert_close(actual, expected)
+
+
+if __name__ == "__main__":
+    pytest.main([__file__])