TARE提交 (#11151)

Co-authored-by: zsy <zsy03260058@163.com>

Author: PatriciaPulec

llm/predict/predictor.py

@@ -40,7 +40,13 @@
     pass
 
 from paddlenlp.generation import GenerationConfig, TextIteratorStreamer
-from paddlenlp.peft import LoRAConfig, LoRAModel, PrefixConfig, PrefixModelForCausalLM
+from paddlenlp.peft import (
+    LoRAConfig,
+    LoRAModel,
+    PrefixConfig,
+    PrefixModelForCausalLM,
+    TAREModel,
+)
 from paddlenlp.taskflow.utils import static_mode_guard
 from paddlenlp.trainer import PdArgumentParser
 from paddlenlp.transformers import (
@@ -85,6 +91,9 @@ class PredictorArgument:
     device: str = field(default="gpu", metadata={"help": "Device"})
     dtype: str = field(default=None, metadata={"help": "Model dtype"})
     lora_path: str = field(default=None, metadata={"help": "The directory of LoRA parameters. Default to None"})
+    tare_path: str = field(default=None, metadata={"help": "The directory of TARE parameters. Default to None"})
+    tare_n: int = field(default=8, metadata={"help": "The num of TARE editors. Default to 8."})
+    tare_k: int = field(default=7, metadata={"help": "The num of TARE selected editors. Default to 7."})
     export_precache: bool = field(default=False, metadata={"help": "whether use prefix weight to do infer"})
     prefix_path: str = field(
         default=None, metadata={"help": "The directory of Prefix Tuning parameters. Default to None"}
@@ -355,6 +364,11 @@ def __init__(
                 prefix_path=config.prefix_path,
                 postprocess_past_key_value=prefix_tuning_params["postprocess_past_key_value"],
             )
+
+        if config.tare_path is not None:
+            self.model = TAREModel(base_model=self.model, n=config.tare_n, k=config.tare_k)
+            self.model.load_model(os.path.join(config.tare_path, "delta_vector.pth"))
+
         self.model.eval()
 
     @paddle.no_grad()

llm/run_finetune.py

@@ -39,6 +39,7 @@
     LoRAModel,
     PrefixConfig,
     PrefixModelForCausalLM,
+    TAREModel,
     VeRAConfig,
     VeRAModel,
 )
@@ -500,6 +501,10 @@ def compute_metrics_do_generation(eval_preds):
         elif last_checkpoint is not None:
             checkpoint = last_checkpoint
         train_result = trainer.train(resume_from_checkpoint=checkpoint)
+
+        if model_args.tare:
+            model.save_model(os.path.join(training_args.output_dir, "delta_vector.pth"))
+
         if model_args.neftune:
             neft_post_hook_handle.remove()
         if training_args.benchmark:
@@ -725,6 +730,10 @@ def create_peft_model(
         model.mark_only_vera_as_trainable(notfreezeB=True)
         model.print_trainable_parameters()
 
+    if model_args.tare:
+        model = TAREModel(base_model=model, n=model_args.tare_n, k=model_args.tare_k)
+        model.print_trainable_parameters()
+
     return model
 
 

paddlenlp/peft/__init__.py

@@ -17,4 +17,5 @@
 from .lora import LoRAAutoConfig, LoRAAutoModel, LoRAConfig, LoRAModel
 from .prefix import PrefixConfig, PrefixModelForCausalLM
 from .reft import ReFTModel
+from .tare import TAREModel
 from .vera import VeRAConfig, VeRAModel

paddlenlp/peft/tare/__init__.py

@@ -0,0 +1,17 @@
+# Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from .tare_model import TAREModel
+
+__all__ = ["TAREModel"]