PTQ-Level Efficiency, QAT-Level Accuracy | 训练后量化的效率,量化感知训练的精度
fix: keep model obj evaluation for regular workflow (#2454)
Xin Heea35ee11fe20
diff --git a/auto_round/eval/evaluation.py b/auto_round/eval/evaluation.pyindex 48ff1cfd5..f39ba47a4 100644--- a/auto_round/eval/evaluation.py+++ b/auto_round/eval/evaluation.py@@ -516,39 +516,31 @@ def run_model_evaluation(model, tokenizer, autoround, folders, formats, args):if model is None:returnelse:- # Evaluate the exported artifact for both regular and model-free flows.- # The in-memory regular model still contains quantization wrappers,- # while fake-format loading materializes FakeActQuantLinear modules.- if model is not None:- model_context = getattr(autoround, "model_context", None)- if model_context is not None and getattr(model_context, "model", None) is model:- model_context.model = None- model = None- from auto_round.utils import clear_memory-- clear_memory()-- eval_model_dtype = get_model_dtype(args.eval_model_dtype, "auto")- if getattr(autoround, "mllm", False):- from auto_round.utils.model import mllm_load_model-- model, _, loaded_tokenizer, _ = mllm_load_model(- eval_folder,- device=device_str,- torch_dtype=eval_model_dtype,- trust_remote_code=not args.disable_trust_remote_code,- )- if tokenizer is None:- tokenizer = loaded_tokenizer- else:- from transformers import AutoModelForCausalLM, AutoTokenizer-- model = AutoModelForCausalLM.from_pretrained(- eval_folder, device_map=device_str, torch_dtype=eval_model_dtype- )- model.eval()- if tokenizer is None:- tokenizer = AutoTokenizer.from_pretrained(eval_folder)+ if model is None:+ eval_model_dtype = get_model_dtype(args.eval_model_dtype, "auto")+ if getattr(autoround, "mllm", False):+ from auto_round.utils.model import mllm_load_model++ model, _, loaded_tokenizer, _ = mllm_load_model(+ eval_folder,+ device=device_str,+ torch_dtype=eval_model_dtype,+ trust_remote_code=not args.disable_trust_remote_code,+ )+ if tokenizer is None:+ tokenizer = loaded_tokenizer+ else:+ from auto_round.utils.model import llm_load_model++ model, loaded_tokenizer = llm_load_model(+ eval_folder,+ device=device_str,+ torch_dtype=eval_model_dtype,+ trust_remote_code=not args.disable_trust_remote_code,+ )+ model.eval()+ if tokenizer is None:+ tokenizer = loaded_tokenizer# Evaluate with model instanceevaluate_with_model_instance(model, tokenizer, device_str, args)diff --git a/test/unit/common/advanced/test_evaluation_functions.py b/test/unit/common/advanced/test_evaluation_functions.pyindex 570e66a3b..092681066 100644--- a/test/unit/common/advanced/test_evaluation_functions.py+++ b/test/unit/common/advanced/test_evaluation_functions.py@@ -206,51 +206,42 @@ def _fake_lm_eval_modules(simple_evaluate_impl):class TestEvalArgumentForwarding:- def test_run_model_evaluation_fake_reloads_exported_artifact(self, tmp_path):+ def test_run_model_evaluation_fake_loads_exported_artifact_with_llm_loader(self, tmp_path):from auto_round.eval.evaluation import run_model_evaluationargs = _make_eval_args()- in_memory_model = MagicMock()loaded_model = MagicMock()- tokenizer = MagicMock()- model_context = SimpleNamespace(model=in_memory_model)- autoround = SimpleNamespace(model_context=model_context)+ loaded_tokenizer = MagicMock()+ autoround = SimpleNamespace()with patch("auto_round.utils.model.detect_model_type", return_value="llm"), patch("auto_round.utils.device_manager.get_device_and_parallelism", return_value=("cpu", False)), patch("auto_round.utils.get_library_version", return_value="test"), patch(- "auto_round.utils.clear_memory"- ) as mock_clear_memory, patch(- "transformers.AutoModelForCausalLM.from_pretrained", return_value=loaded_model- ) as mock_from_pretrained, patch(+ "auto_round.utils.model.llm_load_model", return_value=(loaded_model, loaded_tokenizer)+ ) as mock_load_model, patch("auto_round.eval.evaluation.evaluate_with_model_instance") as mock_evaluate:- run_model_evaluation(in_memory_model, tokenizer, autoround, str(tmp_path), ["fake"], args)+ run_model_evaluation(None, None, autoround, str(tmp_path), ["fake"], args)- mock_from_pretrained.assert_called_once_with(str(tmp_path), device_map="cpu", torch_dtype="auto")- mock_clear_memory.assert_called_once()- assert model_context.model is None- mock_evaluate.assert_called_once_with(loaded_model, tokenizer, "cpu", args)+ mock_load_model.assert_called_once_with(str(tmp_path), device="cpu", torch_dtype="auto", trust_remote_code=True)+ mock_evaluate.assert_called_once_with(loaded_model, loaded_tokenizer, "cpu", args)- def test_run_model_evaluation_fake_reloads_mllm_artifact(self, tmp_path):+ def test_run_model_evaluation_fake_loads_mllm_artifact_when_model_missing(self, tmp_path):from auto_round.eval.evaluation import run_model_evaluationargs = _make_eval_args(mllm=True, disable_trust_remote_code=True)- in_memory_model = MagicMock()loaded_model = MagicMock()loaded_tokenizer = MagicMock()- autoround = SimpleNamespace(mllm=True, model_context=SimpleNamespace(model=in_memory_model))+ autoround = SimpleNamespace(mllm=True)with patch("auto_round.utils.model.detect_model_type", return_value="mllm"), patch("auto_round.utils.device_manager.get_device_and_parallelism", return_value=("cpu", False)), patch("auto_round.utils.get_library_version", return_value="test"), patch(- "auto_round.utils.clear_memory"- ), patch("auto_round.utils.model.mllm_load_model", return_value=(loaded_model, None, loaded_tokenizer, None)) as mock_load, patch("auto_round.eval.evaluation.evaluate_with_model_instance") as mock_evaluate:- run_model_evaluation(in_memory_model, None, autoround, str(tmp_path), ["fake"], args)+ run_model_evaluation(None, None, autoround, str(tmp_path), ["fake"], args)mock_load.assert_called_once_with(str(tmp_path), device="cpu", torch_dtype="auto", trust_remote_code=False)mock_evaluate.assert_called_once_with(loaded_model, loaded_tokenizer, "cpu", args)