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 missing extra_config export for unsupported ignore_layers like `mlp.gate` (#1660)

Liang Lvd2a62b89ecff
diff --git a/auto_round/compressors/utils.py b/auto_round/compressors/utils.py
index 58f64f683..5d9b97567 100644
--- a/auto_round/compressors/utils.py
+++ b/auto_round/compressors/utils.py
@@ -347,9 +347,11 @@ def set_layer_config(
extra_scheme_keys = ("scale_dtype",)
scheme_keys = tuple(f.name for f in fields(QuantizationScheme)) + ("scale_dtype",)
layer_config = copy.deepcopy(layer_config) or {}
+ ignore_layer_patterns = set()
if ignore_layers:
ignore_layers = ignore_layers.replace(" ", "").split(",")
ignore_layers = [name + "." if name[-1].isdigit() else name for name in ignore_layers]
+ ignore_layer_patterns = set(ignore_layers)
# 1. ignore_layers -> force 16
for name in get_fp_layer_names(model, ignore_layers):
@@ -434,11 +436,16 @@ def set_layer_config(
if name in all_module_names:
m = get_module(model, name)
if len(list(m.children())) == 0 and type(m) not in supported_types:
- layer_config.pop(name)
- logger.warning(
- f"'{name}' exists in the model but is not a supported quantization target "
- f"in the current scheme, ignoring its setting in `layer_config`"
- )
+ val = layer_config.pop(name)
+ if name in ignore_layer_patterns:
+ # Keep unsupported ignore_layers entries so export can serialize
+ # them into regex-based extra_config for loaders like vLLM INC.
+ regex_config[name] = val
+ else:
+ logger.warning(
+ f"'{name}' exists in the model but is not a supported quantization target "
+ f"in the current scheme, ignoring its setting in `layer_config`"
+ )
continue
regex = re.compile(to_standard_regex(name))
diff --git a/test/test_cpu/utils/test_set_layer_config.py b/test/test_cpu/utils/test_set_layer_config.py
index 4ed85bc94..91a2222e7 100644
--- a/test/test_cpu/utils/test_set_layer_config.py
+++ b/test/test_cpu/utils/test_set_layer_config.py
@@ -56,6 +56,30 @@ class _SimpleModel(nn.Module):
return self.lm_head(x)
+class _Router(nn.Module):
+ """Unsupported leaf module used to emulate MoE routers like mlp.gate."""
+
+ def __init__(self, size: int = 32):
+ super().__init__()
+ self.weight = nn.Parameter(torch.randn(size, size))
+
+
+class _MoeBlock(nn.Module):
+ def __init__(self, size: int = 32):
+ super().__init__()
+ self.gate = _Router(size)
+ self.up_proj = nn.Linear(size, size)
+
+
+class _UnsupportedGateModel(nn.Module):
+ def __init__(self, size: int = 32):
+ super().__init__()
+ self.mlp = _MoeBlock(size)
+
+ def forward(self, x): # pragma: no cover
+ return self.mlp.up_proj(x)
+
+
# Shared helpers
_SUPPORTED_TYPES = (nn.Linear,)
_INNER_SUPPORTED_TYPES = ()
@@ -87,6 +111,11 @@ def model():
return _SimpleModel()
+def unsupported_gate_model():
+ return _UnsupportedGateModel()
+
+
# ---------------------------------------------------------------------------
# Tests
# ---------------------------------------------------------------------------
@@ -131,3 +160,15 @@ class TestUnmatchedLayerConfigWarns:
# The warning should mention the bad name
warning_messages = " ".join(str(call) for call in mock_warn.call_args_list)
assert "nonexistent_layer" in warning_messages
+
+
+class TestIgnoreLayersRegexFallback:
+ """Unsupported ignore_layers should be preserved for export as regex extra_config."""
+
+ def test_unsupported_ignore_layer_is_kept_in_regex_config(self, unsupported_gate_model):
+ layer_config, _, regex_config = _call_set_layer_config(unsupported_gate_model, ignore_layers="mlp.gate")
+
+ assert "mlp.gate" not in layer_config
+ assert "mlp.gate" in regex_config
+ assert regex_config["mlp.gate"]["bits"] == 16
+ assert regex_config["mlp.gate"]["data_type"] == "float"