Move optimizer to create_optimizer (modelscope#2851)

swift/llm/argument/merge_args.py

@@ -2,7 +2,7 @@
 from dataclasses import dataclass
 from typing import Optional
 
-from swift.utils import get_logger, is_merge_kit_available
+from swift.utils import get_logger
 
 logger = get_logger()
 
@@ -16,44 +16,7 @@ class MergeArguments:
         merge_lora (bool): Flag to indicate if LoRA merging is enabled. Default is False.
         safe_serialization(bool): Use safetensors or not, default `True`.
         max_shard_size(str): The max size of single shard file.
-        use_merge_kit (bool): Flag to indicate merge with `mergekit`. Default is False.
-        instruct_model (Optional[str]): Path or ID of the instruct model. Use when `use_merge_kit` is True.
-        instruct_model_revision (Optional[str]): Revision of the instruct model. Use when `use_merge_kit` is True.
     """
     merge_lora: bool = False
     safe_serialization: bool = True
     max_shard_size: str = '5GB'
-
-    use_merge_kit: bool = False
-    instruct_model: Optional[str] = None
-    instruct_model_revision: Optional[str] = None
-
-    def __post_init__(self):
-        if self.use_merge_kit:
-            assert is_merge_kit_available(), ('please install mergekit by pip install '
-                                              'git+https://github.com/arcee-ai/mergekit.git')
-            logger.info('Important: You are using mergekit, please remember '
-                        'the LoRA should be trained against the base model,'
-                        'and pass its instruct model by --instruct_model xxx when merging')
-            assert self.instruct_model, 'Please pass in the instruct model'
-
-            self.merge_yaml = """
-models:
-  - model: {merged_model}
-    parameters:
-      weight: 1
-      density: 1
-  - model: {instruct_model}
-    parameters:
-      weight: 1
-      density: 1
-merge_method: ties
-base_model: {base_model}
-parameters:
-  weight: 1
-  density: 1
-  normalize: true
-  int8_mask: true
-tokenizer_source: {merged_model}
-dtype: bfloat16
-"""

swift/llm/argument/train_args.py

@@ -136,6 +136,13 @@ def __post_init__(self) -> None:
         TunerArguments.__post_init__(self)
         TorchAccArguments.__post_init__(self)
 
+        if self.lorap_lr_ratio:
+            self.optimizer = 'lorap'
+        elif self.use_galore:
+            self.optimizer = 'galore'
+        elif self.optimizer is None:
+            self.optimizer = 'default'
+
         if len(self.dataset) == 0:
             raise ValueError(f'self.dataset: {self.dataset}, Please input the training dataset.')
 

swift/llm/export/merge_lora.py

@@ -25,14 +25,6 @@ def merge_lora(args: ExportArguments, device_map=None, replace_if_exists=False)
         origin_device_map = args.device_map
         args.device_map = device_map or args.device_map
         logger.info(f'merge_device_map: {device_map}')
-        if args.use_merge_kit:
-            base_model = args.model
-            if not os.path.exists(base_model):
-                base_model = args.hub.download_model(base_model, revision=args.model_revision)
-            if not os.path.exists(args.instruct_model):
-                args.instruct_model = args.hub.download_model(
-                    args.instruct_model, revision=args.instruct_model_revision)
-            args.model = args.instruct_model
         model, template = prepare_model_template(args)
         logger.info('Merge LoRA...')
         Swift.merge_and_unload(model)
@@ -52,19 +44,3 @@ def merge_lora(args: ExportArguments, device_map=None, replace_if_exists=False)
 
     args.model = output_dir
     args.adapters = []
-    if args.use_merge_kit:
-        tempdir = tempfile.gettempdir()
-        mergekit_path = os.path.join(output_dir, 'mergekit')
-        merge_yaml = args.merge_yaml.replace('{merged_model}', output_dir).replace('{instruct_model}',
-                                                                                   args.instruct_model).replace(
-                                                                                       '{base_model}', base_model)
-        try:
-            yamlfile = os.path.join(tempdir, 'mergekit.yaml')
-            with open(yamlfile, 'w', encoding='utf-8') as f:
-                f.write(merge_yaml)
-            logger.info(f'Merging with config: {merge_yaml}')
-            os.system(f'mergekit-yaml {yamlfile} {mergekit_path}')
-            logger.info(f'Merge complete with path: {mergekit_path}')
-        finally:
-            if tempdir:
-                shutil.rmtree(tempdir, ignore_errors=True)

swift/llm/train/sft.py

@@ -122,8 +122,6 @@ def run(self):
         self.train_msg['model_parameter_info'] = model_parameter_info
         logger.info(f'model_parameter_info: {model_parameter_info}')
 
-        optimizers = self._get_optimizers(train_dataset)
-
         trainer_cls = TrainerFactory.get_trainer_cls(args)
         trainer = trainer_cls(
             model=self.model,
@@ -132,7 +130,6 @@ def run(self):
             train_dataset=train_dataset,
             eval_dataset=val_dataset,
             callbacks=self.callbacks,
-            optimizers=optimizers,
             template=self.template,
             **self._get_trainer_kwargs(),
         )
@@ -192,18 +189,6 @@ def train(self, trainer):
 
         return self._save_trainer_state(trainer)
 
-    def _get_optimizers(self, train_dataset):
-        args = self.args
-        if args.lorap_lr_ratio:
-            optimizer_callback = optimizers_map['lorap']
-        elif args.use_galore:
-            optimizer_callback = optimizers_map['galore']
-        elif args.optimizer is not None:
-            optimizer_callback = optimizers_map[args.optimizer]
-        else:
-            optimizer_callback = optimizers_map['default']
-        return optimizer_callback(args, self.model, train_dataset)
-
     def _prepare_callbacks(self):
         from .callback import DynamicLayerActivationCallback, TrainerAdapterCallback
         args = self.args

swift/llm/train/tuner.py

@@ -403,6 +403,7 @@ def prepare_model(
                 gamma_proj=args.galore_gamma_proj,
                 queue_size=args.galore_queue_size,
             )
+            args.training_args.galore_config = args.galore_config
 
         if args.sequence_parallel_size > 1:
             from swift.trainers.xtuner import dispatch_module_xtuner

swift/plugin/optimizer.py

@@ -11,7 +11,6 @@ def calculate_max_steps(args: 'TrainArguments', dataset) -> int:
     if args.max_steps and args.max_steps > 0:
         max_steps = args.max_steps
     else:
-        assert not args.streaming
         len_dataset = len(dataset)
         _, _, world_size, _ = get_dist_setting()
         total_train_batch_size = args.per_device_train_batch_size * args.gradient_accumulation_steps * world_size
@@ -23,17 +22,14 @@ def calculate_max_steps(args: 'TrainArguments', dataset) -> int:
 
 def create_galore_optimizers(args, model, dataset):
     training_steps = calculate_max_steps(args, dataset)
-    return create_optimizer_and_scheduler(
-        model,
-        args.training_args,
-        args.galore_config,
-        training_steps,
-        lr=args.learning_rate,
-        weight_decay=args.weight_decay)
+    optimizer, lr_scheduler = create_optimizer_and_scheduler(
+        model, args, args.galore_config, training_steps, lr=args.learning_rate, weight_decay=args.weight_decay)
+    # trainer cannot serialize galore_config
+    args.galore_config = None
+    return optimizer, lr_scheduler
 
 
 def create_lorap_optimizers(args, model, dataset):
-    args = args.training_args
     optimizer_grouped_parameters = None
     if hasattr(model, 'create_optimizer_param_groups'):
         # Lora+ parameter groups

swift/trainers/arguments.py

@@ -10,6 +10,7 @@
 from transformers.training_args_seq2seq import Seq2SeqTrainingArguments as HfSeq2SeqTrainingArguments
 
 from swift.utils import use_torchacc
+from .optimizers.galore import GaLoreConfig
 
 
 @dataclass
@@ -28,8 +29,10 @@ class SwiftArgumentsMixin:
     fsdp_num: int = 1
     acc_steps: int = 1
 
-    # Value copied from TrainArguments, Used for external tuners.
+    # Value copied from TrainArguments
     train_type: Optional[str] = None
+    optimizer: Optional[str] = None
+    galore_config: Optional[GaLoreConfig] = None
 
     def _fix_gradient_checkpointing(self):
         # fix use_reentrant

swift/trainers/mixin.py

@@ -21,7 +21,7 @@
 from transformers.data.data_collator import DataCollator
 from transformers.integrations import is_deepspeed_zero3_enabled
 from transformers.modeling_utils import unwrap_model
-from transformers.trainer import Trainer, TrainerCallback
+from transformers.trainer import TrainerCallback
 from transformers.trainer_utils import EvalPrediction
 from transformers.utils import is_torch_npu_available
 
@@ -32,7 +32,6 @@
 from swift.utils import get_logger, is_mp_ddp, use_torchacc
 from swift.utils.torchacc_utils import ta_trim_graph
 from .arguments import TrainingArguments
-from .optimizers.galore import create_optimizer_and_scheduler
 from .utils import can_return_loss, find_labels, get_function, is_instance_of_ms_model
 
 try:
@@ -316,32 +315,17 @@ def _maybe_log_save_evaluate(self, tr_loss, *args, **kwargs):
         super()._maybe_log_save_evaluate(tr_loss, *args, **kwargs)
 
     def create_optimizer_and_scheduler(self, num_training_steps: int):
-        if hasattr(self.args, 'galore_config'):
-            optimizer, lr_scheduler = create_optimizer_and_scheduler(
-                self.model,
-                self.args,
-                self.args.galore_config,
-                num_training_steps,
-                lr=self.args.learning_rate,
-                weight_decay=self.args.weight_decay)
-            self.optimizer = optimizer
-            self.lr_scheduler = lr_scheduler
+        if self.args.optimizer is not None:
+            from swift.plugin import optimizers_map
+            optimizer_callback = optimizers_map[self.args.optimizer]
+            self.optimizer, self.lr_scheduler = optimizer_callback(self.args, self.model, self.train_dataset)
+            if self.optimizer is None:
+                self.create_optimizer()
+            if self.lr_scheduler is None:
+                self.create_scheduler(num_training_steps=num_training_steps, optimizer=self.optimizer)
         else:
             super().create_optimizer_and_scheduler(num_training_steps=num_training_steps)
 
-    def create_optimizer(self):
-
-        if self.optimizer is None and hasattr(self.model, 'create_optimizer_param_groups'):
-            # Lora+ parameter groups
-            optimizer_grouped_parameters = self.model.create_optimizer_param_groups(
-                lr=self.args.learning_rate, weight_decay=self.args.weight_decay)
-            if optimizer_grouped_parameters is not None:
-                optimizer_cls, optimizer_kwargs = Trainer.get_optimizer_cls_and_kwargs(self.args)
-                self.optimizer = optimizer_cls(optimizer_grouped_parameters, **optimizer_kwargs)
-                return self.optimizer
-
-        return super().create_optimizer()
-
     def _get_train_sampler(self) -> Optional[torch.utils.data.Sampler]:
         if self.args.train_sampler_random:
             return super()._get_train_sampler()

swift/ui/llm_infer/llm_infer.py

@@ -29,8 +29,6 @@ class LLMInfer(BaseUI):
 
     is_multimodal = True
 
-    deployed = False
-
     sub_ui = [Model, Runtime]
 
     locale_dict = {
@@ -279,7 +277,6 @@ def deploy_model(cls, *args):
         os.system(run_command)
         gr.Info(cls.locale('load_alert', cls.lang)['value'])
         time.sleep(2)
-        cls.deployed = True
         running_task = Runtime.refresh_tasks(log_file)
         return gr.update(open=True), running_task
 

swift/utils/__init__.py

@@ -2,8 +2,8 @@
 
 from .env import (get_dist_setting, get_pai_tensorboard_dir, is_deepspeed_enabled, is_dist, is_dist_ta, is_local_master,
                   is_master, is_mp, is_mp_ddp, is_pai_training_job, torchacc_trim_graph, use_hf_hub, use_torchacc)
-from .import_utils import (is_liger_available, is_lmdeploy_available, is_megatron_available, is_merge_kit_available,
-                           is_unsloth_available, is_vllm_available, is_xtuner_available)
+from .import_utils import (is_liger_available, is_lmdeploy_available, is_megatron_available, is_unsloth_available,
+                           is_vllm_available, is_xtuner_available)
 from .io_utils import (JsonlWriter, append_to_jsonl, download_ms_file, get_file_mm_type, open_jsonl_writer,
                        read_from_jsonl, write_to_jsonl)
 from .logger import get_logger

swift/utils/import_utils.py

@@ -16,10 +16,6 @@ def is_vllm_available():
     return importlib.util.find_spec('vllm') is not None
 
 
-def is_merge_kit_available():
-    return importlib.util.find_spec('mergekit') is not None
-
-
 def is_lmdeploy_available():
     return importlib.util.find_spec('lmdeploy') is not None