PTQ-Level Efficiency, QAT-Level Accuracy | 训练后量化的效率,量化感知训练的精度
fix bug of evaluate_accuracy (#1430)
Xin He40253afe46cc
diff --git a/test/test_cuda/advanced/test_multiple_card.py b/test/test_cuda/advanced/test_multiple_card.pyindex 1ea003501..33ac5a38d 100644--- a/test/test_cuda/advanced/test_multiple_card.py+++ b/test/test_cuda/advanced/test_multiple_card.py@@ -39,8 +39,7 @@ class TestAutoRound:autoround = AutoRound(model, tokenizer, device_map=device_map)autoround.quantize()autoround.save_quantized(self.save_dir, format="auto_round", inplace=False)- model_args = f"pretrained={self.save_dir}"- evaluate_accuracy(model="hf", model_args=model_args, threshold=0.45, batch_size="auto")+ evaluate_accuracy(self.save_dir, threshold=0.45, batch_size="auto")shutil.rmtree("./saved", ignore_errors=True)@multi_carddiff --git a/test/test_cuda/core/test_main_func.py b/test/test_cuda/core/test_main_func.pyindex dc8cdc15b..82ccbea44 100644--- a/test/test_cuda/core/test_main_func.py+++ b/test/test_cuda/core/test_main_func.py@@ -41,14 +41,12 @@ class TestMainFunc:##test auto_round formatautoround.save_quantized(self.save_dir, format="auto_round", inplace=False)- model_args = f"pretrained={self.save_dir}"- evaluate_accuracy(model="hf", model_args=model_args, threshold=0.35, batch_size="auto")+ evaluate_accuracy(self.save_dir, threshold=0.35, batch_size="auto")shutil.rmtree("./saved", ignore_errors=True)##test auto_gptq formatautoround.save_quantized(self.save_dir, format="auto_gptq", inplace=False)- model_args = f"pretrained={self.save_dir}"- evaluate_accuracy(model="hf", model_args=model_args, threshold=0.35, batch_size="auto")+ evaluate_accuracy(self.save_dir, threshold=0.35, batch_size="auto")shutil.rmtree("./saved", ignore_errors=True)@require_optimum@@ -63,8 +61,7 @@ class TestMainFunc:##test auto_awq formatautoround.save_quantized(self.save_dir, format="auto_awq", inplace=False)- model_args = f"pretrained={self.save_dir}"- evaluate_accuracy(model="hf", model_args=model_args, threshold=0.35, batch_size="auto")+ evaluate_accuracy(self.save_dir, threshold=0.35, batch_size="auto")shutil.rmtree("./saved", ignore_errors=True)@pytest.mark.skipif(not torch.cuda.is_available(), reason="CUDA not available")@@ -84,8 +81,7 @@ class TestMainFunc:##test auto_round formatautoround.save_quantized(self.save_dir, format="auto_round", inplace=False)- model_args = f"pretrained={self.save_dir}"- evaluate_accuracy(model="hf", model_args=model_args, threshold=0.35, batch_size="auto")+ evaluate_accuracy(self.save_dir, threshold=0.35, batch_size="auto")shutil.rmtree("./saved", ignore_errors=True)@pytest.mark.skipif(not torch.cuda.is_available(), reason="CUDA not available")@@ -106,8 +102,7 @@ class TestMainFunc:##test auto_awq formatautoround.save_quantized(self.save_dir, format="auto_awq", inplace=False)- model_args = f"pretrained={self.save_dir}"- evaluate_accuracy(model="hf", model_args=model_args, threshold=0.35, batch_size="auto")+ evaluate_accuracy(self.save_dir, threshold=0.35, batch_size="auto")shutil.rmtree("./saved", ignore_errors=True)@pytest.mark.skipif(not torch.cuda.is_available(), reason="CUDA not available")@@ -128,8 +123,7 @@ class TestMainFunc:##test auto_round formatautoround.save_quantized(self.save_dir, format="auto_round", inplace=False)- model_args = f"pretrained={self.save_dir}"- evaluate_accuracy(model="hf", model_args=model_args, threshold=0.34, batch_size="auto")+ evaluate_accuracy(self.save_dir, threshold=0.34, batch_size="auto")shutil.rmtree("./saved", ignore_errors=True)def test_autoround_asym(self): ##need to install false@@ -146,8 +140,7 @@ class TestMainFunc:##test auto_round formatautoround.save_quantized(self.save_dir, format="auto_round", inplace=False)- model_args = f"pretrained={self.save_dir}"- evaluate_accuracy(model="hf", model_args=model_args, threshold=0.35, batch_size="auto")+ evaluate_accuracy(self.save_dir, threshold=0.35, batch_size="auto")shutil.rmtree("./saved", ignore_errors=True)def test_attention_mask_lm_head(self, tiny_qwen_moe_model_path):diff --git a/test/test_cuda/quantization/test_2_3bits.py b/test/test_cuda/quantization/test_2_3bits.pyindex 4f68612f8..6caccdbce 100644--- a/test/test_cuda/quantization/test_2_3bits.py+++ b/test/test_cuda/quantization/test_2_3bits.py@@ -51,8 +51,7 @@ class TestAutoRound:bits, sym = 3, Falseautoround = AutoRound(model_name, bits=bits, sym=sym)autoround.quantize_and_save(self.save_dir, format="auto_round", inplace=False)- model_args = f"pretrained={self.save_dir}"- evaluate_accuracy(model="hf", model_args=model_args, threshold=0.32, batch_size="auto")+ evaluate_accuracy(self.save_dir, threshold=0.32, batch_size="auto")shutil.rmtree("./saved", ignore_errors=True)@require_greater_than_050@@ -63,8 +62,7 @@ class TestAutoRound:##test auto_round formatautoround.save_quantized(self.save_dir, format="auto_round", inplace=False)- model_args = f"pretrained={self.save_dir}"- evaluate_accuracy(model="hf", model_args=model_args, threshold=0.18, batch_size="auto")+ evaluate_accuracy(self.save_dir, threshold=0.18, batch_size="auto")shutil.rmtree("./saved", ignore_errors=True)@require_greater_than_050@@ -75,11 +73,9 @@ class TestAutoRound:##test auto_round formatautoround.save_quantized(self.save_dir, format="auto_round", inplace=False)- model_args = f"pretrained={self.save_dir}"- evaluate_accuracy(model="hf", model_args=model_args, threshold=0.17, batch_size="auto")+ evaluate_accuracy(self.save_dir, threshold=0.17, batch_size="auto")shutil.rmtree("./saved", ignore_errors=True)autoround.save_quantized(self.save_dir, format="auto_gptq", inplace=False)- model_args = f"pretrained={self.save_dir}"- evaluate_accuracy(model="hf", model_args=model_args, threshold=0.17, batch_size="auto")+ evaluate_accuracy(self.save_dir, threshold=0.17, batch_size="auto")shutil.rmtree("./saved", ignore_errors=True)