Update Training Guidelines

README.md

@@ -245,6 +245,7 @@
 问题4：为什么不对模型做全量预训练而是用LoRA？
 问题5：二代模型支不支持某些支持一代LLaMA的工具？
 问题6：Chinese-Alpaca-2是Llama-2-Chat训练得到的吗？
+问题7：为什么24G显存微调chinese-alpaca-2-7b OOM？
 ```
 
 

README_EN.md

@@ -228,6 +228,7 @@ Question 3: Do you accept third-party Pull Requests?
 Question 4: Why not perform full pre-training but use LoRA instead?
 Question 5: Does Llama-2 series support tools that support the first-gen LLaMA?
 Question 6: Is Chinese-Alpaca-2 trained from Llama-2-Chat?
+Question 7: Why does training with 24GB VRAM lead to an OOM error when fine-tuning chinese-alpaca-2-7b?
 ```
 
 For specific questions and answers, please refer to the project >>> [📚 GitHub Wiki](https://github.com/ymcui/Chinese-LLaMA-Alpaca-2/wiki/faq_en)

scripts/inference/gradio_demo.py

@@ -109,7 +109,8 @@
 parent_dir = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
 sys.path.append(parent_dir)
 from attn_and_long_ctx_patches import apply_attention_patch, apply_ntk_scaling_patch
-apply_attention_patch(use_memory_efficient_attention=True)
+if not args.only_cpu:
+    apply_attention_patch(use_memory_efficient_attention=True)
 apply_ntk_scaling_patch(args.alpha)
 
 # Set CUDA devices if available
@@ -192,7 +193,7 @@ def setup():
                 args.lora_model,
                 torch_dtype=load_type,
                 device_map='auto',
-            )
+            ).half()
         else:
             model = base_model
 

scripts/inference/inference_hf.py

@@ -63,7 +63,8 @@
 parent_dir = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
 sys.path.append(parent_dir)
 from attn_and_long_ctx_patches import apply_attention_patch, apply_ntk_scaling_patch
-apply_attention_patch(use_memory_efficient_attention=True)
+if not args.only_cpu:
+    apply_attention_patch(use_memory_efficient_attention=True)
 apply_ntk_scaling_patch(args.alpha)
 
 if args.use_vllm:
@@ -131,7 +132,7 @@ def generate_prompt(instruction, system_prompt=DEFAULT_SYSTEM_PROMPT):
             base_model.resize_token_embeddings(tokenizer_vocab_size)
         if args.lora_model is not None:
             print("loading peft model")
-            model = PeftModel.from_pretrained(base_model, args.lora_model,torch_dtype=load_type,device_map='auto',)
+            model = PeftModel.from_pretrained(base_model, args.lora_model,torch_dtype=load_type,device_map='auto',).half()
         else:
             model = base_model
 

scripts/training/build_dataset.py

@@ -62,7 +62,7 @@ def tokenization(examples):
 
         if data_cache_dir is None:
             data_cache_dir = str(os.path.dirname(file))
-        cache_path = os.path.join(data_cache_dir,os.path.basename(file).split('.')[0])
+        cache_path = os.path.join(data_cache_dir,os.path.basename(file).split('.')[0]+f"_{max_seq_length}")
         os.makedirs(cache_path, exist_ok=True)
         try:
             processed_dataset = datasets.load_from_disk(cache_path)

scripts/training/run_clm_pt_with_peft.py

@@ -467,13 +467,13 @@ def group_texts(examples):
         for idx, file in enumerate(files):
             data_file = os.path.join(path, file)
             filename = ''.join(file.split(".")[:-1])
-            cache_path = os.path.join(data_args.data_cache_dir, filename)
+            cache_path = os.path.join(data_args.data_cache_dir, filename+f"_{block_size}")
             os.makedirs(cache_path, exist_ok=True)
             try:
                 processed_dataset = datasets.load_from_disk(cache_path, keep_in_memory=False)
                 logger.info(f'training datasets-{filename} has been loaded from disk')
             except Exception:
-                cache_dir = os.path.join(data_args.data_cache_dir, filename+"_text")
+                cache_dir = os.path.join(data_args.data_cache_dir, filename+f"_text_{block_size}")
                 os.makedirs(cache_dir, exist_ok=True)
                 raw_dataset = load_dataset("text", data_files=data_file, cache_dir=cache_dir, keep_in_memory=False)
                 logger.info(f"{file} has been loaded")
@@ -503,7 +503,6 @@ def group_texts(examples):
             else:
                 assert lm_datasets.features.type == processed_dataset["train"].features.type
                 lm_datasets = concatenate_datasets([lm_datasets, processed_dataset["train"]])
-
         lm_datasets = lm_datasets.train_test_split(test_size = data_args.validation_split_percentage)
 
     if training_args.do_train:
@@ -522,26 +521,24 @@ def group_texts(examples):
         logger.info(f"Num eval_samples  {len(eval_dataset)}")
         logger.info("Evaluation example:")
         logger.info(tokenizer.decode(eval_dataset[0]['input_ids']))
-    if model_args.model_name_or_path:
-        torch_dtype = (
-            model_args.torch_dtype
-            if model_args.torch_dtype in ["auto", None]
-            else getattr(torch, model_args.torch_dtype)
-        )
-        model = LlamaForCausalLM.from_pretrained(
-            model_args.model_name_or_path,
-            from_tf=bool(".ckpt" in model_args.model_name_or_path),
-            config=config,
-            cache_dir=model_args.cache_dir,
-            revision=model_args.model_revision,
-            use_auth_token=True if model_args.use_auth_token else None,
-            torch_dtype=torch_dtype,
-            low_cpu_mem_usage=True
-        )
-    else:
-        model = AutoModelForCausalLM.from_config(config)
-        n_params = sum({p.data_ptr(): p.numel() for p in model.parameters()}.values())
-        logger.info(f"Training new model from scratch - Total size={n_params/2**20:.2f}M params")
+    torch_dtype = (
+        model_args.torch_dtype
+        if model_args.torch_dtype in ["auto", None]
+        else getattr(torch, model_args.torch_dtype)
+    )
+    device_map = {"":int(os.environ.get("LOCAL_RANK") or 0)}
+    model = LlamaForCausalLM.from_pretrained(
+        model_args.model_name_or_path,
+        from_tf=bool(".ckpt" in model_args.model_name_or_path),
+        config=config,
+        cache_dir=model_args.cache_dir,
+        revision=model_args.model_revision,
+        use_auth_token=True if model_args.use_auth_token else None,
+        torch_dtype=torch_dtype,
+        low_cpu_mem_usage=True,
+        device_map=device_map
+    )
+    model.config.use_cache = False
 
     model_vocab_size = model.get_output_embeddings().weight.size(0)
     tokenizer_vocab_size = len(tokenizer)
@@ -555,7 +552,7 @@ def group_texts(examples):
 
     if training_args.peft_path is not None:
         logger.info("Peft from pre-trained model")
-        model = PeftModel.from_pretrained(model, training_args.peft_path)
+        model = PeftModel.from_pretrained(model, training_args.peft_path, device_map=device_map)
     else:
         logger.info("Init new peft model")
         target_modules = training_args.trainable.split(',')

scripts/training/run_clm_sft_with_peft.py

@@ -51,7 +51,6 @@
 from peft import LoraConfig, TaskType, get_peft_model, PeftModel, get_peft_model_state_dict
 from transformers.trainer_utils import PREFIX_CHECKPOINT_DIR
 
-IGNORE_INDEX = -100
 
 require_version("datasets>=1.8.0", "To fix: pip install -r examples/pytorch/language-modeling/requirements.txt")
 
@@ -295,7 +294,7 @@ def main():
             "You can do it from another script, save it, and load it from here, using --tokenizer_name."
         )
 
-    if (len(tokenizer))!=55296:
+    if (len(tokenizer)) != 55296:
         raise ValueError(f"The vocab size of the tokenizer should be 55296, but found {len(tokenizer)}.\n"
                          "Please use Chinese-LLaMA-2 tokenizer.")
 
@@ -331,26 +330,24 @@ def main():
         logger.info("Evaluation example:")
         logger.info(tokenizer.decode(eval_dataset[0]['input_ids']))
 
-    if model_args.model_name_or_path:
-        torch_dtype = (
-            model_args.torch_dtype
-            if model_args.torch_dtype in ["auto", None]
-            else getattr(torch, model_args.torch_dtype)
-        )
-        model = LlamaForCausalLM.from_pretrained(
-            model_args.model_name_or_path,
-            from_tf=bool(".ckpt" in model_args.model_name_or_path),
-            config=config,
-            cache_dir=model_args.cache_dir,
-            revision=model_args.model_revision,
-            use_auth_token=True if model_args.use_auth_token else None,
-            torch_dtype=torch_dtype,
-            low_cpu_mem_usage=True
-        )
-    else:
-        model = AutoModelForCausalLM.from_config(config)
-        n_params = sum({p.data_ptr(): p.numel() for p in model.parameters()}.values())
-        logger.info(f"Training new model from scratch - Total size={n_params/2**20:.2f}M params")
+    torch_dtype = (
+        model_args.torch_dtype
+        if model_args.torch_dtype in ["auto", None]
+        else getattr(torch, model_args.torch_dtype)
+    )
+    device_map = {"":int(os.environ.get("LOCAL_RANK") or 0)}
+    model = LlamaForCausalLM.from_pretrained(
+        model_args.model_name_or_path,
+        from_tf=bool(".ckpt" in model_args.model_name_or_path),
+        config=config,
+        cache_dir=model_args.cache_dir,
+        revision=model_args.model_revision,
+        use_auth_token=True if model_args.use_auth_token else None,
+        torch_dtype=torch_dtype,
+        low_cpu_mem_usage=True,
+        device_map=device_map
+    )
+    model.config.use_cache = False
 
     model_vocab_size = model.get_input_embeddings().weight.shape[0]
     logger.info(f"Model vocab size: {model_vocab_size}")
@@ -361,7 +358,7 @@ def main():
 
     if training_args.peft_path is not None:
         logger.info("Peft from pre-trained model")
-        model = PeftModel.from_pretrained(model, training_args.peft_path)
+        model = PeftModel.from_pretrained(model, training_args.peft_path, device_map=device_map)
     else:
         logger.info("Init new peft model")
         target_modules = training_args.trainable.split(',')

scripts/training/run_pt.sh

@@ -1,3 +1,5 @@
+# 运行脚本前请仔细阅读wiki(https://github.com/ymcui/Chinese-LLaMA-Alpaca-2/wiki/pt_scripts_zh)
+# Read the wiki(https://github.com/ymcui/Chinese-LLaMA-Alpaca-2/wiki/pt_scripts_zh) carefully before running the script
 lr=2e-4
 lora_rank=64
 lora_alpha=128
@@ -6,12 +8,12 @@ modules_to_save="embed_tokens,lm_head"
 lora_dropout=0.05
 
 pretrained_model=path/to/hf/llama-2/dir
-chinese_tokenizer_path=path/to/chinese/llama-2/tokenizer/dir
+chinese_tokenizer_path=path/to/chinese-llama-2/tokenizer/dir
 dataset_dir=path/to/pt/data/dir
 data_cache=temp_data_cache_dir
 per_device_train_batch_size=1
-per_device_eval_batch_size=1
 gradient_accumulation_steps=8
+block_size=512
 output_dir=output_dir
 
 deepspeed_config_file=ds_zero2_no_offload.json
@@ -24,7 +26,6 @@ torchrun --nnodes 1 --nproc_per_node 1 run_clm_pt_with_peft.py \
     --data_cache_dir ${data_cache} \
     --validation_split_percentage 0.001 \
     --per_device_train_batch_size ${per_device_train_batch_size} \
-    --per_device_eval_batch_size ${per_device_eval_batch_size} \
     --do_train \
     --seed $RANDOM \
     --fp16 \
@@ -40,16 +41,14 @@ torchrun --nnodes 1 --nproc_per_node 1 run_clm_pt_with_peft.py \
     --save_steps 200 \
     --gradient_accumulation_steps ${gradient_accumulation_steps} \
     --preprocessing_num_workers 8 \
-    --block_size 1024 \
+    --block_size ${block_size} \
     --output_dir ${output_dir} \
     --overwrite_output_dir \
     --ddp_timeout 30000 \
     --logging_first_step True \
     --lora_rank ${lora_rank} \
     --lora_alpha ${lora_alpha} \
     --trainable ${lora_trainable} \
-    --modules_to_save ${modules_to_save} \
     --lora_dropout ${lora_dropout} \
-    --torch_dtype float16 \
-    --gradient_checkpointing \
-    --ddp_find_unused_parameters False
+    --modules_to_save ${modules_to_save} \
+    --torch_dtype float16 

scripts/training/run_sft.sh

@@ -1,18 +1,20 @@
+# 运行脚本前请仔细阅读wiki(https://github.com/ymcui/Chinese-LLaMA-Alpaca-2/wiki/sft_scripts_zh)
+# Read the wiki(https://github.com/ymcui/Chinese-LLaMA-Alpaca-2/wiki/sft_scripts_zh) carefully before running the script
 lr=1e-4
 lora_rank=64
 lora_alpha=128
 lora_trainable="q_proj,v_proj,k_proj,o_proj,gate_proj,down_proj,up_proj"
 modules_to_save="embed_tokens,lm_head"
 lora_dropout=0.05
 
-pretrained_model=path/to/hf/llama-2/or/merged/llama-2/dir/or/model_id
-chinese_tokenizer_path=path/to/chinese/llama-2/tokenizer/dir
+pretrained_model=path/to/hf/llama-2/or/chinese-llama-2/dir/or/model_id
+chinese_tokenizer_path=path/to/chinese-llama-2/tokenizer/dir
 dataset_dir=path/to/sft/data/dir
 per_device_train_batch_size=1
 per_device_eval_batch_size=1
 gradient_accumulation_steps=8
+max_seq_length=512
 output_dir=output_dir
-peft_model=path/to/peft/model/dir
 validation_file=validation_file_name
 
 deepspeed_config_file=ds_zero2_no_offload.json
@@ -22,7 +24,6 @@ torchrun --nnodes 1 --nproc_per_node 1 run_clm_sft_with_peft.py \
     --model_name_or_path ${pretrained_model} \
     --tokenizer_name_or_path ${chinese_tokenizer_path} \
     --dataset_dir ${dataset_dir} \
-    --validation_split_percentage 0.001 \
     --per_device_train_batch_size ${per_device_train_batch_size} \
     --per_device_eval_batch_size ${per_device_eval_batch_size} \
     --do_train \
@@ -43,18 +44,15 @@ torchrun --nnodes 1 --nproc_per_node 1 run_clm_sft_with_peft.py \
     --save_steps 200 \
     --gradient_accumulation_steps ${gradient_accumulation_steps} \
     --preprocessing_num_workers 8 \
-    --max_seq_length 1024 \
+    --max_seq_length ${max_seq_length} \
     --output_dir ${output_dir} \
     --overwrite_output_dir \
     --ddp_timeout 30000 \
     --logging_first_step True \
     --lora_rank ${lora_rank} \
     --lora_alpha ${lora_alpha} \
     --trainable ${lora_trainable} \
-    --modules_to_save ${modules_to_save} \
     --lora_dropout ${lora_dropout} \
+    --modules_to_save ${modules_to_save} \
     --torch_dtype float16 \
-    --validation_file ${validation_file} \
-    --peft_path ${peft_model} \
-    --gradient_checkpointing \
-    --ddp_find_unused_parameters False
+    --validation_file ${validation_file}