Skip to content

PTQ-Level Efficiency, QAT-Level Accuracy | 训练后量化的效率,量化感知训练的精度

fix: update get_predefined_ignore_layers to include shared_expert_gat… (#2397)

Xin He6db9435fbff6
diff --git a/auto_round/special_model_handler.py b/auto_round/special_model_handler.py
index 40a6c1565..dd9202c60 100644
--- a/auto_round/special_model_handler.py
+++ b/auto_round/special_model_handler.py
@@ -1183,7 +1183,7 @@ def get_predefined_ignore_layers(model: torch.nn.Module) -> list[str]:
config = getattr(model, "config", None)
if not layers and is_moe_model_via_config(config):
for name, _ in model.named_modules():
- if name.endswith(".gate"):
+ if name.endswith((".gate", ".shared_expert_gate")):
layers.append(name)
return list(dict.fromkeys(layers))
diff --git a/test/unit/common/models/test_omni_model.py b/test/unit/common/models/test_omni_model.py
index 13f349471..ddf417f66 100644
--- a/test/unit/common/models/test_omni_model.py
+++ b/test/unit/common/models/test_omni_model.py
@@ -353,6 +353,7 @@ class TestQwen3OmniMoeUtils:
assert ignore_layers == [
"thinker.model.layers.0.mlp.gate",
"talker.model.layers.0.mlp.gate",
+ "talker.model.layers.0.mlp.shared_expert_gate",
], f"Expected mlp.gate in ignore_layers for qwen3_omni_moe, got: {ignore_layers}"
diff --git a/test/unit/common/models/test_special_model_handler.py b/test/unit/common/models/test_special_model_handler.py
index 159daa28b..5bfcebe97 100644
--- a/test/unit/common/models/test_special_model_handler.py
+++ b/test/unit/common/models/test_special_model_handler.py
@@ -888,8 +888,11 @@ class TestRegisterIgnoreLayers:
initial_count = len(_PRE_DEFINED_IGNORE_LAYERS)
matcher = MagicMock(return_value=True)
- register_ignore_layers(matchers=[matcher], ignore_layers=["layer.0"])
- assert len(_PRE_DEFINED_IGNORE_LAYERS) == initial_count + 1
+ try:
+ register_ignore_layers(matchers=[matcher], ignore_layers=["layer.0"])
+ assert len(_PRE_DEFINED_IGNORE_LAYERS) == initial_count + 1
+ finally:
+ del _PRE_DEFINED_IGNORE_LAYERS[initial_count:]
class TestGetPredefinedIgnoreLayers:
@@ -977,6 +980,27 @@ class TestGetPredefinedIgnoreLayers:
mock_model.named_modules.return_value = iter([])
layers = get_predefined_ignore_layers(mock_model)
# Should not add any layers without matching rules
+ assert layers == []
+
+ def test_generic_moe_ignores_router_and_shared_expert_gates(self):
+ from types import SimpleNamespace
+
+ from auto_round.special_model_handler import get_predefined_ignore_layers
+
+ mock_model = MagicMock()
+ mock_model.config = SimpleNamespace(model_type="test_moe", architectures=[])
+ mock_model.named_modules.return_value = iter(
+ [
+ ("layers.0.mlp.gate", MagicMock()),
+ ("layers.0.mlp.shared_expert_gate", MagicMock()),
+ ("layers.0.mlp.gate", MagicMock()),
+ ("layers.0.mlp.up_proj", MagicMock()),
+ ]
+ )
+
+ layers = get_predefined_ignore_layers(mock_model)
+
+ assert layers == ["layers.0.mlp.gate", "layers.0.mlp.shared_expert_gate"]
class TestTorchCompileOff: