fix fatal error in finetuning

caused by utils.py:L135: the model.half() cannot be used in finetuning

arguments.py

@@ -2,7 +2,7 @@
 from dataclasses import dataclass, field
 
 
-CHATGLM_LASTEST_HASH = 'acd41f77311be8584836edc2fc7251d5b6e65840'
+CHATGLM_LASTEST_HASH = 'cde457b39fe0670b10dd293909aab17387ea2c80'
 DATASETS = {
     "alpaca_en": {
         "train": {

finetune.sh

@@ -6,6 +6,9 @@ CUDA_VISIBLE_DEVICES=0 python finetune_chatglm.py \
     --output_dir output \
     --per_device_train_batch_size 4 \
     --per_device_eval_batch_size 1 \
-    --gradient_accumulation_steps 8 \
-    --logging_steps 50 \
+    --gradient_accumulation_steps 4 \
+    --logging_steps 10 \
     --save_steps 1000 \
+    --warmup_steps 100 \
+    --fp16 \
+    --num_train_epochs 1.0

finetune_chatglm.py

@@ -46,6 +46,7 @@ def main():
         model.save_pretrained(training_args.output_dir)
 
     # Testing
+    model = model.half()
     model.eval()
     response, _ = model.chat(tokenizer, query='你好', history=[])
     print(response)

utils.py

@@ -26,6 +26,12 @@
 logger = logging.getLogger(__name__)
 
 
+class CastOutputToFloat(torch.nn.Sequential):
+
+    def forward(self, x):
+        return super().forward(x).to(torch.float32)
+
+
 def prepare_args():
     # Load arguments
     parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments, FinetuningArguments))
@@ -113,14 +119,14 @@ def prepare_model(model_args, finetuning_args):
         **config_kwargs
     )
     model = AutoModel.from_pretrained(model_args.model_name_or_path, config=config, **config_kwargs)
+    model.config.use_cache = False
 
     if model_args.quantization_bit is not None:
         print("Quantized to {} bit".format(model_args.quantization_bit))
         model = model.quantize(model_args.quantization_bit)
+        model.lm_head = CastOutputToFloat(model.lm_head)
 
     if finetuning_args.finetuning_type == 'lora':
-        for param in model.parameters():
-            param.requires_grad_(False)
         peft_config = LoraConfig(
             task_type=TaskType.CAUSAL_LM,
             inference_mode=False,
@@ -132,8 +138,6 @@ def prepare_model(model_args, finetuning_args):
         model = get_peft_model(model, peft_config)
         model.print_trainable_parameters()
 
-    model = model.half()
-
     return tokenizer, model
 
 
@@ -174,6 +178,7 @@ def format_example(examples):
                 yield prompt, answer
 
     def preprocess_function_train(examples):
+        # build inputs with format `X [gMASK] [BOS] Y [EOP]`
         model_inputs = {"input_ids": [], "labels": []}
         for prompt, answer in format_example(examples):
             source_ids = tokenizer.encode(text=prompt, add_special_tokens=False)
@@ -193,6 +198,7 @@ def preprocess_function_train(examples):
         return model_inputs
 
     def preprocess_function_eval(examples):
+        # build inputs with format `X [gMASK] [BOS]`
         model_inputs = {"input_ids": [], "labels": []}
         for prompt, answer in format_example(examples):
             source_ids = tokenizer.encode(text=prompt)
@@ -250,7 +256,7 @@ def print_dataset_example(example):
 """
 Note: The ChatGLM tokenizer assigns False on token to be attended in attention mask. In general settings, it should be True.
 Refer to: https://huggingface.co/THUDM/chatglm-6b/blob/6650ae3a53c28fc176d06762ca80b05d5ab3792b/tokenization_chatglm.py#L401
-Inspired by: https://github.com/tatsu-lab/stanford_alpaca/blob/main/train.py#L166
+Inspired by: https://github.com/tatsu-lab/stanford_alpaca/blob/aa65c492bb788e144712daab42bc5d11c2761591/train.py#L166
 """
 @dataclass
 class DataCollatorForChatGLM(DataCollatorWithPadding):
@@ -263,17 +269,15 @@ def __call__(self, features: Sequence[Dict[str, Sequence]]) -> Dict[str, torch.T
         label_pad_token_id = IGNORE_INDEX if self.data_args.ignore_pad_token_for_loss else self.tokenizer.pad_token_id
         labels = torch.nn.utils.rnn.pad_sequence(labels, batch_first=True, padding_value=label_pad_token_id)
         features = {"input_ids": input_ids, "labels": labels}
-        return super().__call__(features)
+        # return super().__call__(features) # enable generating attention mask and position ids
+        return features
 
 
 """
-Inspired by: https://github.com/mymusise/ChatGLM-Tuning/blob/master/finetune.py#L52
+Inspired by: https://github.com/mymusise/ChatGLM-Tuning/blob/997393046a49510e6cda36962f9a399297959311/finetune.py#L52
 """
 class TrainerForChatGLM(Trainer):
 
-    def compute_loss(self, model, inputs, return_outputs=False):
-        return model(**inputs).loss
-
     def _save(self, output_dir: Optional[str] = None, _internal_call: bool = False):
         from transformers.trainer import TRAINING_ARGS_NAME
         output_dir = output_dir if output_dir is not None else self.args.output_dir