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 performance regression bug (#2136)

Heng Guof8ad63f4e210
diff --git a/auto_round/algorithms/quantization/base.py b/auto_round/algorithms/quantization/base.py
index c83d805d5..70b3299e4 100644
--- a/auto_round/algorithms/quantization/base.py
+++ b/auto_round/algorithms/quantization/base.py
@@ -198,19 +198,6 @@ class BaseQuantizer(BaseAlgorithm):
the valid-token loss mask. ``None`` for RTN fallback.
"""
self._quantize_layer_via_rtn(layer, disable_opt_rtn=disable_opt_rtn)
- """Quantize a single layer outside a transformer block using RTN fallback.
- Args:
- layer: The layer module to quantize. Must have a
- ``global_name`` attribute for model re-insertion.
- fp_inputs: Optional FP calibration inputs; unused in base RTN.
- q_inputs: Optional quantized activations; unused in base RTN.
- disable_opt_rtn: ``True`` skips optimized-RTN scale/zp search.
- ``None`` defers to ``self.config.disable_opt_rtn``.
- valid_token_mask: Per-sample masks; unused in base RTN.
- input_ids: Original FP calibration inputs, same as ``fp_inputs``
- when ``q_inputs`` is ``None``; unused in base RTN.
- """
- self._quantize_layer_via_rtn(layer, disable_opt_rtn=disable_opt_rtn)
@torch.no_grad()
def _quantize_layer_via_rtn(self, layer: "torch.nn.Module", disable_opt_rtn: "bool | None" = None) -> None:
diff --git a/auto_round/algorithms/quantization/rtn/quantizer.py b/auto_round/algorithms/quantization/rtn/quantizer.py
index 6fee985fd..e4c6466e2 100644
--- a/auto_round/algorithms/quantization/rtn/quantizer.py
+++ b/auto_round/algorithms/quantization/rtn/quantizer.py
@@ -19,6 +19,7 @@ from auto_round.algorithms.quantization.rtn.config import OptimizedRTNConfig, RT
from auto_round.algorithms.registry import register_pipeline_member
from auto_round.logger import logger
from auto_round.utils import (
+ SUPPORTED_LAYER_TYPES,
check_to_quantized,
)
@@ -99,7 +100,7 @@ class OptimizedRTNQuantizer(RTNQuantizer):
handles = []
for _, module in model.named_modules():
- if check_to_quantized(module):
+ if isinstance(module, SUPPORTED_LAYER_TYPES) and check_to_quantized(module):
handles.append(module.register_forward_hook(collect_imatrix))
return handles
diff --git a/auto_round/algorithms/quantization/sign_roundv2/quantizer.py b/auto_round/algorithms/quantization/sign_roundv2/quantizer.py
index 4c306f5cc..e99494fed 100644
--- a/auto_round/algorithms/quantization/sign_roundv2/quantizer.py
+++ b/auto_round/algorithms/quantization/sign_roundv2/quantizer.py
@@ -43,7 +43,7 @@ from auto_round.data_type.utils import (
search_optimized_init_scale,
)
from auto_round.logger import logger
-from auto_round.utils import check_to_quantized, compile_func, get_reciprocal
+from auto_round.utils import SUPPORTED_LAYER_TYPES, check_to_quantized, compile_func, get_reciprocal
from auto_round.wrapper import WrapperLinear, wrapper_block
@@ -416,6 +416,6 @@ class SignRoundV2Quantizer(SignRoundQuantizer):
handles = []
for _, module in model.named_modules():
- if check_to_quantized(module):
+ if isinstance(module, SUPPORTED_LAYER_TYPES) and check_to_quantized(module):
handles.append(module.register_forward_hook(collect_imatrix))
return handles