PTQ-Level Efficiency, QAT-Level Accuracy | 训练后量化的效率,量化感知训练的精度
Fix act calibration bug (#880)
Wang, Mengnib693a2a6b0cf
diff --git a/auto_round/compressors/base.py b/auto_round/compressors/base.pyindex 505552cbc..48cf0cd1a 100644--- a/auto_round/compressors/base.py+++ b/auto_round/compressors/base.py@@ -2664,7 +2664,7 @@ class BaseCompressor(object):config = self.layer_config[n]act_dynamic = config.get("act_dynamic", True)act_data_type = config.get("act_data_type", None)- act_bits = config.get("act_data_type", 16)+ act_bits = config.get("act_bits", 16)if (config["bits"] <= 8and check_need_act_calibration(act_dynamic, act_data_type, act_bits)diff --git a/test/test_cpu/test_autoround.py b/test/test_cpu/test_autoround.pyindex 9511f0cf8..de5f7412a 100644--- a/test/test_cpu/test_autoround.py+++ b/test/test_cpu/test_autoround.py@@ -49,6 +49,22 @@ class TestAutoRound(unittest.TestCase):if module.bits != 8:raise ValueError(f"Expected bits to be 8, but got {module.bits}")+ def test_layer_config(self):+ model_name = "/tf_dataset/auto_round/models/facebook/opt-125m"+ layer_config = {"self_attn": {"bits": 4, "data_type": "nv_fp", "act_bits": 16, "group_size": 16}}+ autoround = AutoRound(+ model_name,+ self.tokenizer,+ scheme="NVFP4",+ iters=0,+ seqlen=2,+ dataset=self.llm_dataloader,+ layer_config=layer_config,+ amp=False,+ )+ autoround.quantize_and_save(self.save_folder, inplace=False, format="fake")+ shutil.rmtree(self.save_folder)+def test_remove_whole_block(self):model_name = "/tf_dataset/auto_round/models/facebook/opt-125m"layer_config = {