fix

inference/generate.py

@@ -6,9 +6,8 @@
 
 # %%
 parser = ArgumentParser()
-parser.add_argument('--model_path', '-m', type=str, default='/mnt/h/models/llama-7b')
+parser.add_argument('--model_path', '-m', type=str, default='')
 parser.add_argument('--lora_path', '-l', type=str, default='')
-# parser.add_argument('--precision', '-p', type=str, default='bf16')
 parser.add_argument('--prompt', '-t', type=str, default='你好')
 
 args = parser.parse_args()
@@ -27,7 +26,7 @@
 print(output)
 while True:
     prompt = input('Human: ')
-    output = pipeline(prompt, max_new_tokens=32, do_sample=False, top_p=1, num_return_sequences=1, return_full_text=False)
+    output = pipeline(prompt, max_new_tokens=32, num_beams=10, num_return_sequences=5, return_full_text=False)
     for o in output:
         print('Bot:', o['generated_text'])
     print('-' * 100)

requirements.txt

@@ -21,4 +21,6 @@ pandas
 scipy
 
 wandb
-tensorboard
+tensorboard
+
+opencc

training/config/config.yml

@@ -15,6 +15,12 @@ do_eval: true
 do_predict: false
 do_train: true
 
+# lora
+lora: true
+load_in_4bit: true
+load_in_8bit: false
+target_modules: [ q_proj, v_proj, k_proj, o_proj, gate_proj, down_proj, up_proj ]
+
 # hyperparameters
 per_device_eval_batch_size: 8
 per_device_train_batch_size: 8
@@ -31,9 +37,11 @@ optim: adamw_torch
 gradient_accumulation_steps: 1
 gradient_checkpointing: true
 
+# backend
 torch_compile: true
 #torch_compile_backend: None
 #torch_compile_mode: None
+deepspeed: None
 
 # precision
 bf16: true
@@ -45,13 +53,12 @@ fp16_opt_level: O1
 tf32: true
 
 # data
+seed: 42
 data_seed: None
 dataloader_drop_last: false
 dataloader_num_workers: 0
 dataloader_pin_memory: true
-
-deepspeed: None
-disable_tqdm: false
+remove_unused_columns: true
 
 # eval
 eval_delay: 0
@@ -62,9 +69,6 @@ prediction_loss_only: True
 #greater_is_better: None
 #metric_for_best_model: None
 
-label_smoothing_factor: 0.0
-
-
 # log
 log_level: passive
 log_level_replica: warning
@@ -75,10 +79,7 @@ logging_steps: 10
 logging_strategy: steps
 report_to:
   - wandb
-
-remove_unused_columns: true
-
-#resume_from_checkpoint: None
+disable_tqdm: false
 
 # save
 save_safetensors: false
@@ -87,9 +88,6 @@ save_strategy: steps
 save_total_limit: 2
 load_best_model_at_end: true
 
-seed: 42
-skip_memory_metrics: true
-
-
+#resume_from_checkpoint: None
 
 

training/train_gpt_hf.py

@@ -51,13 +51,14 @@ def __post_init__(self):
 class MyTrainingArguments(TrainingArguments):
     project_name: str = ''
     # lora
+    lora: bool = False
     target_modules: list[str] = None
 
     def __post_init__(self):
         super().__post_init__()
         # wandb
         os.environ['WANDB_PROJECT'] = self.project_name
-        self.run_name += f'-{time.time()}'
+        self.run_name += f'-{int(time.time())}'
 
 
 # %% config
@@ -88,18 +89,20 @@ def __post_init__(self):
 
 # %% model
 model = load_model(**model_args.__dict__)
-model.enable_input_require_grads()
-peft_config = LoraConfig(
-    task_type=TaskType.CAUSAL_LM,
-    inference_mode=False,
-    r=8,
-    lora_alpha=32,
-    lora_dropout=0.05,
-    # target_modules=training_args.target_modules
-    # bias="all"
-)
-model = get_peft_model(model, peft_config)
-model.print_trainable_parameters()
+
+if training_args.lora:
+    model.enable_input_require_grads()
+    peft_config = LoraConfig(
+        task_type=TaskType.CAUSAL_LM,
+        inference_mode=False,
+        r=8,
+        lora_alpha=32,
+        lora_dropout=0.05,
+        # target_modules=training_args.target_modules
+        # bias="all"
+    )
+    model = get_peft_model(model, peft_config)
+    model.print_trainable_parameters()
 
 # %% optimizer
 # if data_args.optim == 'adam_bf16':
@@ -122,6 +125,7 @@ def __post_init__(self):
 tokenizer.save_pretrained(f'{training_args.output_dir}/best')
 
 """
-CONFIG=''
-torchrun --nnodes 1 --nproc-per-node 1 training/train_gpt_hf.py -c $CONFIG
+config=''
+python training/train_gpt_hf.py -c $config
+torchrun --nnodes 1 --nproc-per-node 1 training/train_gpt_hf.py -c $config
 """

utils/data/load_data.py

@@ -64,9 +64,9 @@ def load_data(path: str, tokenizer: PreTrainedTokenizer, max_length: int = 1024,
         data = processor.load_json(path=path)
     # else:
     #     data = processor.load_dataset_dict(path=path)
-    print(d)
-    print(d['train'][0]['input_ids'])
-    print(tokenizer.decode(d['train'][0]['input_ids']))
+    print(data)
+    print(data['train'][0]['input_ids'])
+    print(tokenizer.decode(data['train'][0]['input_ids']))
     return data
 
 

utils/model/load_model.py

@@ -5,7 +5,7 @@
 
 
 def load_model(model_name_or_path, model_type, **kwargs) -> PreTrainedModel:
-    args = {'device_map': 'auto'}
+    args = {'device_map': 'auto', 'torch_dtype': 'auto'}
     if kwargs.get('load_in_8bit'):
         args['load_in_8bit'] = True
         quantization_config = BitsAndBytesConfig(