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 quantization configuration handling in FakeFormat and related tests (#2311)
Xin He777c4ea7bbe4
diff --git a/auto_round/experimental/qmodules/fake.py b/auto_round/experimental/qmodules/fake.pyindex 7f1a56c40..0089a2555 100644--- a/auto_round/experimental/qmodules/fake.py+++ b/auto_round/experimental/qmodules/fake.py@@ -69,10 +69,6 @@ class FakeActQuantLinear(QModuleBase):returndef qdq_input(self, activation: torch.Tensor) -> torch.Tensor:- # No activation quantization configured (e.g. weight-only GGUF-style schemes where- # act_bits defaults to 16 / act_data_type is None) -> skip activation fake-quant.- if self.config.act_data_type is None or (self.config.act_bits is not None and self.config.act_bits >= 16):- return activationquant_func, _ = get_quant_func(dtype=self.config.act_data_type,bits=self.config.act_bits,diff --git a/auto_round/export/formats/backends/fake.py b/auto_round/export/formats/backends/fake.pyindex d29e131f8..bea104ab9 100644--- a/auto_round/export/formats/backends/fake.py+++ b/auto_round/export/formats/backends/fake.py@@ -119,16 +119,18 @@ class FakeFormat(OutputFormat):if config_act_bits is not None and config_act_bits <= 8:has_fake_act_quant = True- quantization_config = _serialize_quantization_config_value(dict(serialization_dict or {}))- quantization_config["quant_method"] = "auto-round"if has_fake_act_quant:+ quantization_config = _serialize_quantization_config_value(dict(serialization_dict or {}))+ quantization_config["quant_method"] = "auto-round"quantization_config["packing_format"] = "auto_round:fake"- quantization_config["block_name_to_quantize"] = quantization_config.pop("to_quant_block_names", None)- from auto_round.export.utils import filter_quantization_config-- filter_quantization_config(quantization_config)- if hasattr(model, "config") and model.config is not None:- model.config.quantization_config = quantization_config+ quantization_config["block_name_to_quantize"] = quantization_config.pop("to_quant_block_names", None)+ from auto_round.export.utils import filter_quantization_config++ filter_quantization_config(quantization_config)+ if hasattr(model, "config") and model.config is not None:+ model.config.quantization_config = quantization_config+ elif hasattr(model, "config") and model.config is not None and hasattr(model.config, "quantization_config"):+ delattr(model.config, "quantization_config")if not has_meta_device:model = model.to("cpu")diff --git a/test/unit/test_cpu/export/test_fake_format.py b/test/unit/test_cpu/export/test_fake_format.pyindex b4d3a3842..798ac9767 100644--- a/test/unit/test_cpu/export/test_fake_format.py+++ b/test/unit/test_cpu/export/test_fake_format.py@@ -58,8 +58,11 @@ class _SaveableModel(torch.nn.Module):def save_pretrained(self, output_dir):os.makedirs(output_dir, exist_ok=True)torch.save(self.state_dict(), os.path.join(output_dir, "pytorch_model.bin"))+ config = {}+ if hasattr(self.config, "quantization_config"):+ config["quantization_config"] = self.config.quantization_configwith open(os.path.join(output_dir, "config.json"), "w") as config_file:- json.dump({"quantization_config": self.config.quantization_config}, config_file)+ json.dump(config, config_file)def test_fake_format_unwraps_quantized_layers_before_save(tmp_path):@@ -211,7 +214,7 @@ def test_transformers_load_replaces_fake_linear(tmp_path):assert not torch.equal(q_proj.qdq_input(activation), activation)-def test_fake_format_keeps_woq_packing_format(tmp_path):+def test_fake_format_omits_woq_quantization_config(tmp_path):model = _SaveableModel()output_dir = str(tmp_path / "woq_model")@@ -232,10 +235,9 @@ def test_fake_format_keeps_woq_packing_format(tmp_path):)with open(os.path.join(output_dir, "config.json")) as config_file:- quantization_config = json.load(config_file)["quantization_config"]+ config = json.load(config_file)- assert quantization_config["packing_format"] == "auto_round:auto_gptq"- assert "act_bits" not in quantization_config+ assert "quantization_config" not in configdef test_fake_format_still_saves_when_env_disabled(tmp_path):