PTQ-Level Efficiency, QAT-Level Accuracy | 训练后量化的效率,量化感知训练的精度
switch to transformers API in cpu ut (#580)
Wenhua Cheng6ceaa8af9f38
diff --git a/test/test_act_quantization.py b/test/test_act_quantization.pyindex 57fde6396..63544aca5 100644--- a/test/test_act_quantization.py+++ b/test/test_act_quantization.py@@ -5,7 +5,6 @@ import unittestsys.path.insert(0, "..")import torch-import transformersfrom transformers import AutoModelForCausalLM, AutoTokenizerfrom auto_round import AutoRounddiff --git a/test/test_autoround.py b/test/test_autoround.pyindex 0fb7feed1..d8d2efed5 100644--- a/test/test_autoround.py+++ b/test/test_autoround.py@@ -7,8 +7,7 @@ from auto_round.eval.evaluation import simple_evaluate_user_modelsys.path.insert(0, "..")import torch-import transformers-from transformers import AutoModelForCausalLM, AutoTokenizer+from transformers import AutoModelForCausalLM, AutoTokenizer, AutoRoundConfigfrom auto_round import AutoRoundfrom _test_helpers import model_infer@@ -358,7 +357,6 @@ class TestAutoRound(unittest.TestCase):)quantized_model_path = self.save_folderautoround.quantize_and_save(output_dir=quantized_model_path, format="auto_round")- from auto_round import AutoRoundConfigmodel = AutoModelForCausalLM.from_pretrained(self.save_folder,torch_dtype=torch.float16,@@ -391,7 +389,6 @@ class TestAutoRound(unittest.TestCase):quantized_model_path = self.save_folderautoround.save_quantized(output_dir=quantized_model_path, format="auto_round", inplace=True)- from auto_round import AutoRoundConfigquantization_config = AutoRoundConfig(backend="ipex")@@ -408,7 +405,6 @@ class TestAutoRound(unittest.TestCase):import requestsfrom PIL import Imagefrom transformers import Qwen2VLForConditionalGeneration, AutoProcessor- from auto_round import AutoRoundConfigfrom auto_round_extension.ipex.qlinear_ipex_awq import QuantLinearmodel_name = "Qwen/Qwen2-VL-2B-Instruct-AWQ"quantization_config = AutoRoundConfig()@@ -480,7 +476,6 @@ class TestAutoRound(unittest.TestCase):quantized_model_path = self.save_folderautoround.save_quantized(output_dir=quantized_model_path, format="auto_awq", inplace=True)- from auto_round import AutoRoundConfigquantization_config = AutoRoundConfig()model = AutoModelForCausalLM.from_pretrained(quantized_model_path,@@ -515,7 +510,6 @@ class TestAutoRound(unittest.TestCase):quantized_model_path = self.save_folderautoround.save_quantized(output_dir=quantized_model_path, format="auto_gptq", inplace=True)- from auto_round import AutoRoundConfigquantization_config = AutoRoundConfig()model = AutoModelForCausalLM.from_pretrained(quantized_model_path,@@ -550,7 +544,6 @@ class TestAutoRound(unittest.TestCase):quantized_model_path = self.save_folderautoround.save_quantized(output_dir=quantized_model_path, format="auto_round", inplace=True)- from auto_round import AutoRoundConfigquantization_config = AutoRoundConfig()model = AutoModelForCausalLM.from_pretrained(quantized_model_path,diff --git a/test/test_block_names.py b/test/test_block_names.pyindex e1e70742e..fce253cee 100644--- a/test/test_block_names.py+++ b/test/test_block_names.py@@ -163,7 +163,7 @@ class TestQuantizationBlocks(unittest.TestCase):quantized_model_path = "./saved"autoround.save_quantized(quantized_model_path, inplace=False, safe_serialization=False, format="auto_round")- from auto_round import AutoRoundConfig+model = AutoModelForCausalLM.from_pretrained(quantized_model_path, device_map="auto")tokenizer = AutoTokenizer.from_pretrained(quantized_model_path)text = "There is a girl who likes adventure,"diff --git a/test/test_conv1d.py b/test/test_conv1d.pyindex 4f45fcbdb..17503cfd2 100644--- a/test/test_conv1d.py+++ b/test/test_conv1d.py@@ -34,7 +34,6 @@ class TestQuantizationConv1d(unittest.TestCase):def test_quant(self):self.model = AutoModelForCausalLM.from_pretrained(self.model_name, torch_dtype="auto", trust_remote_code=True)bits, group_size, sym = 4, 128, True- from auto_round import AutoRoundConfigautoround = AutoRound(self.model,self.tokenizer,diff --git a/test/test_export.py b/test/test_export.pyindex 55cecd37b..75c75c024 100644--- a/test/test_export.py+++ b/test/test_export.py@@ -4,7 +4,7 @@ import unittestsys.path.insert(0, "..")import torch-from transformers import AutoModelForCausalLM, AutoTokenizer+from transformers import AutoModelForCausalLM, AutoTokenizer, AutoRoundConfigfrom auto_round import AutoRound@@ -75,7 +75,6 @@ class TestAutoRound(unittest.TestCase):device = "auto" ##cpu, hpu, cuda- from auto_round import AutoRoundConfigquantization_config = AutoRoundConfig(backend=device)@@ -104,7 +103,6 @@ class TestAutoRound(unittest.TestCase):autoround.save_quantized(output_dir=quantized_model_path, inplace=False, format="auto_round:awq")- from auto_round import AutoRoundConfigquantization_config = AutoRoundConfig(backend="cpu")diff --git a/test/test_generation.py b/test/test_generation.pyindex e2a140970..d76f66970 100644--- a/test/test_generation.py+++ b/test/test_generation.py@@ -5,8 +5,7 @@ import unittestsys.path.insert(0, "..")import torch-import transformers-from transformers import AutoModelForCausalLM, AutoTokenizer+from transformers import AutoModelForCausalLM, AutoTokenizer, AutoRoundConfigfrom auto_round import AutoRound@@ -52,7 +51,6 @@ class TestAutoRoundFormatGeneration(unittest.TestCase):autoround.quantize_and_save(output_dir=quantized_model_path, format="auto_round", inplace=False)- from auto_round import AutoRoundConfigquantization_config = AutoRoundConfig(backend="ipex")diff --git a/test/test_load_awq_gptq.py b/test/test_load_awq_gptq.pyindex 6db06c03a..9266c8ee9 100644--- a/test/test_load_awq_gptq.py+++ b/test/test_load_awq_gptq.py@@ -6,7 +6,7 @@ sys.path.insert(0, "..")from transformers import AutoModelForCausalLM, AutoTokenizer-from auto_round import AutoRoundConfig+from transformers import AutoRoundConfigclass TestAutoRound(unittest.TestCase):diff --git a/test/test_mllm.py b/test/test_mllm.pyindex f92cdf0d7..2e247796d 100644--- a/test/test_mllm.py+++ b/test/test_mllm.py@@ -131,8 +131,6 @@ class TestAutoRoundMLLM(unittest.TestCase):autoround.quantize()quantized_model_path = "./saved"autoround.save_quantized(quantized_model_path, format="auto_round", inplace=False)-- from auto_round import AutoRoundConfigimport requestsfrom PIL import Image