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: remove stale quantizer access in dual-transformer diffusion (#2236)

Wang, Chang68f8aa328193
diff --git a/auto_round/compressors/diffusion_mixin.py b/auto_round/compressors/diffusion_mixin.py
index 44465a244..edfd651a4 100644
--- a/auto_round/compressors/diffusion_mixin.py
+++ b/auto_round/compressors/diffusion_mixin.py
@@ -459,8 +459,8 @@ class DiffusionMixin:
self.layer_config = {}
# Get new block names for caching
- if bool(self.quantizer.quant_block_list):
- all_blocks = self.quantizer.quant_block_list
+ if bool(self.quant_block_list):
+ all_blocks = self.quant_block_list
else:
all_blocks = get_block_names(self.model_context.model)
if len(all_blocks) == 0:
@@ -498,7 +498,7 @@ class DiffusionMixin:
self.compress_context.is_immediate_saving = orig_immediate_saving
self.num_inference_steps = orig_steps
- return self.model_context.model, self.quantizer.layer_config
+ return self.model_context.model, self.layer_config
def save_quantized(
self,
diff --git a/test/unit/common/compressors/test_diffusion_mixin.py b/test/unit/common/compressors/test_diffusion_mixin.py
index 648e8eab0..f9885b7e2 100644
--- a/test/unit/common/compressors/test_diffusion_mixin.py
+++ b/test/unit/common/compressors/test_diffusion_mixin.py
@@ -37,6 +37,9 @@ class TestDiffusionMixinProperties:
comp = MockCompressor()
assert comp._get_calibrator_kind() == "diffusion"
+ def test_quantize_does_not_use_removed_quantizer_attribute(self):
+ assert "self.quantizer" not in inspect.getsource(DiffusionMixin.quantize)
+
def test_pipeline_call_kwargs_extracted_from_kwargs(self):
class MockCompressor(DiffusionMixin):
def __init__(self):