更新.gitignore以排除rsync.sh，修改TODO列表，重命名evaluate脚本，删除run.sh，添加持续学习的参数类，更新训练和评估脚本以支持新的数据集逻辑

src/utils/accelerator.py

@@ -0,0 +1,76 @@
+from accelerate import Accelerator, DataLoaderConfiguration
+
+
+def create_accelerator_and_postprocess(args):
+    # We explicitly don't rely on the `Accelerator` to do gradient accumulation
+    grad_acc_kwargs = {}
+    if args.accelerator_config.gradient_accumulation_kwargs is not None:
+        grad_acc_kwargs = args.accelerator_config.gradient_accumulation_kwargs
+
+    # check if num_steps is attempted to be passed in gradient_accumulation_kwargs
+    if "num_steps" in grad_acc_kwargs:
+        if args.gradient_accumulation_steps > 1:
+            # raise because we do not know which setting is intended.
+            raise ValueError(
+                "The `AcceleratorConfig`'s `num_steps` is set but `gradient_accumulation_steps` is greater than 1 in the passed `TrainingArguments`"
+                "If using the passed `AcceleratorConfig` is desired, do not set the `TrainingArguments` `gradient_accumulation_steps`."
+            )
+        else:
+            args.gradient_accumulation_steps = grad_acc_kwargs["num_steps"]
+
+    accelerator_config = args.accelerator_config.to_dict()
+
+    dataloader_config = DataLoaderConfiguration(
+        split_batches=accelerator_config.pop("split_batches"),
+        dispatch_batches=accelerator_config.pop("dispatch_batches"),
+        even_batches=accelerator_config.pop("even_batches"),
+        use_seedable_sampler=accelerator_config.pop("use_seedable_sampler"),
+    )
+    dataloader_config.data_seed = args.data_seed
+
+    non_blocking = accelerator_config.pop("non_blocking")
+    dataloader_config.non_blocking = non_blocking
+    # this would have been updated above, no need for it anymore
+    accelerator_config.pop("gradient_accumulation_kwargs")
+
+    accelerator_args = {
+        "deepspeed_plugin": args.deepspeed_plugin,
+    }
+    accelerator_args["dataloader_config"] = dataloader_config
+    # create accelerator object
+    accelerator = Accelerator(**accelerator_args)
+
+    # deepspeed and accelerate flags covering both trainer args and accelerate launcher
+    is_deepspeed_enabled = (
+        getattr(accelerator.state, "deepspeed_plugin", None) is not None
+    )
+    is_fsdp_enabled = getattr(accelerator.state, "fsdp_plugin", None) is not None
+
+    # post accelerator creation setup
+    if is_fsdp_enabled:
+        fsdp_plugin = accelerator.state.fsdp_plugin
+        fsdp_plugin.limit_all_gathers = args.fsdp_config.get(
+            "limit_all_gathers", fsdp_plugin.limit_all_gathers
+        )
+        fsdp_plugin.activation_checkpointing = args.fsdp_config.get(
+            "activation_checkpointing", fsdp_plugin.activation_checkpointing
+        )
+        if fsdp_plugin.activation_checkpointing and args.gradient_checkpointing:
+            raise ValueError(
+                "The activation_checkpointing in FSDP config and the gradient_checkpointing in training arg "
+                "can't be set to True simultaneously. Please use FSDP's activation_checkpointing logic "
+                "when using FSDP."
+            )
+
+    def propagate_args_to_deepspeed(auto_find_batch_size=False):
+        from transformers.integrations.deepspeed import HfTrainerDeepSpeedConfig
+
+        ds_plugin = accelerator.state.deepspeed_plugin
+
+        ds_plugin.hf_ds_config = HfTrainerDeepSpeedConfig(ds_plugin.hf_ds_config.config)
+        ds_plugin.deepspeed_config = ds_plugin.hf_ds_config.config
+        ds_plugin.hf_ds_config.trainer_config_process(args, auto_find_batch_size)
+
+    if is_deepspeed_enabled and getattr(args, "hf_deepspeed_config", None) is None:
+        propagate_args_to_deepspeed()
+    return accelerator

src/utils/args.py

@@ -0,0 +1,17 @@
+from dataclasses import dataclass, field
+from typing import Optional
+
+
+@dataclass
+class ContinualScriptArguments:
+    """Script arguments for continual learning."""
+
+    dataset_name: list[str] = field(
+        default_factory=lambda: ["cifar10", "cifar100", "imagenet2012"]
+    )
+    dataset_config: Optional[str] = None
+    dataset_train_split: str = "train"
+    dataset_test_split: str = "test"
+    dataset_generation_split: str = "generation"
+    gradient_checkpointing_use_reentrant: bool = False
+    ignore_bias_buffers: bool = False

src/utils/trainer.py

@@ -1,9 +1,9 @@
 # _________________________________________________________
 
-from trl import SFTTrainer
+from transformers import Trainer
 
 
-class ContinualTrainer(SFTTrainer):
+class ContinualTrainer(Trainer):
     def __init__(
         self, model, args, data_collator, train_dataset, eval_dataset, accelerator
     ):
@@ -19,6 +19,7 @@ class ContinualTrainer(SFTTrainer):
             self.is_fsdp_enabled = (
                 getattr(self.accelerator.state, "fsdp_plugin", None) is not None
             )
+            self.gather_function = self.accelerator.gather_for_metrics
             return
         else:
             super().create_accelerator_and_postprocess()