[LLM] add decay steps option for finetuning (#8251)

llm/finetune_generation.py

@@ -14,6 +14,7 @@
 import json
 import os
 import sys
+from dataclasses import dataclass, field
 from functools import partial
 
 import paddle
@@ -49,6 +50,23 @@ from paddlenlp.transformers import (
 from paddlenlp.utils.log import logger
 
 
+def add_start_docstrings(*docstr):
+    def docstring_decorator(fn):
+        fn.__doc__ = "".join(docstr) + (fn.__doc__ if fn.__doc__ is not None else "")
+        return fn
+
+    return docstring_decorator
+
+
+@dataclass
+@add_start_docstrings(TrainingArguments.__doc__)
+class FinetuneArguments(TrainingArguments):
+    decay_steps: int = field(
+        default=0,
+        metadata={"help": "The steps use to control the learing rate."},
+    )
+
+
 def read_local_dataset(path):
     with open(path, "r", encoding="utf-8") as fp:
         for line in fp:
@@ -57,7 +75,7 @@ def read_local_dataset(path):
 
 def main():
     # Arguments
-    parser = PdArgumentParser((GenerateArgument, QuantArgument, ModelArgument, DataArgument, TrainingArguments))
+    parser = PdArgumentParser((GenerateArgument, QuantArgument, ModelArgument, DataArgument, FinetuneArguments))
     # Support format as "args.json --arg1 value1 --arg2 value2.”
     # In case of conflict, command line arguments take precedence.
     if len(sys.argv) >= 2 and sys.argv[1].endswith(".json"):

paddlenlp/trainer/trainer.py

@@ -1650,13 +1650,16 @@ class Trainer:
         warmup = (
             self.args.warmup_steps if self.args.warmup_steps > 0 else int(self.args.warmup_ratio * num_training_steps)
         )
+        decay_steps = num_training_steps
+        if hasattr(self.args, "decay_steps") and self.args.decay_steps > 0:
+            decay_steps = self.args.decay_steps
 
         if self.lr_scheduler is None:
             self.lr_scheduler = get_scheduler(
                 self.args.lr_scheduler_type,
                 learning_rate=self.args.learning_rate,
                 num_warmup_steps=warmup,
-                num_training_steps=num_training_steps,
+                num_training_steps=decay_steps,
                 num_cycles=self.args.num_cycles,
                 lr_end=self.args.lr_end,
                 power=self.args.power,