PTQ-Level Efficiency, QAT-Level Accuracy | 训练后量化的效率,量化感知训练的精度
Refine exllamav2 ut (#1001)
Weiwei758b239089d9
diff --git a/test/test_cuda/test_exllamav2_backend.py b/test/test_cuda/test_exllamav2_backend.pyindex 5c12e0557..c489b37b2 100644--- a/test/test_cuda/test_exllamav2_backend.py+++ b/test/test_cuda/test_exllamav2_backend.py@@ -12,7 +12,7 @@ from transformers import AutoModelForCausalLM, AutoTokenizerfrom auto_round import AutoRound, AutoRoundConfigfrom auto_round.eval.evaluation import simple_evaluate_user_model-from auto_round.testing_utils import require_autogptq, require_gptqmodel+from auto_round.testing_utils import require_autogptq, require_gptqmodel, require_package_version_utclass LLMDataLoader:@@ -24,7 +24,7 @@ class LLMDataLoader:yield torch.ones([1, 10], dtype=torch.long)-class TestAutoRoundMarlinBackend(unittest.TestCase):+class TestAutoRoundexllamaBackend(unittest.TestCase):@classmethoddef setUpClass(self):@@ -99,6 +99,7 @@ class TestAutoRoundMarlinBackend(unittest.TestCase):shutil.rmtree("./saved", ignore_errors=True)@require_autogptq+ @require_package_version_ut("torch", "<2.6.0")def test_gptq_exllamav2_4bits_sym(self):model = AutoModelForCausalLM.from_pretrained(self.model_name, torch_dtype="auto", trust_remote_code=True)tokenizer = AutoTokenizer.from_pretrained(self.model_name, trust_remote_code=True)@@ -130,6 +131,7 @@ class TestAutoRoundMarlinBackend(unittest.TestCase):shutil.rmtree(self.save_folder, ignore_errors=True)@require_autogptq+ @require_package_version_ut("torch", "<2.6.0")def test_gptq_exllamav2_4bits_sym_group_size(self):for group_size in [-1, 32, 64, 128, 256, 1024]: ## 384, 768 has accuracy issueprint(f"!!!!!!!!!!!!!!!!!{group_size}!!!!!!!!!!!!!!!!!")