rinna-3bを試してみる

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("rinna/japanese-gpt-neox-3.6b", use_fast=False)
model = AutoModelForCausalLM.from_pretrained("rinna/japanese-gpt-neox-3.6b")

if torch.cuda.is_available():
    model = model.to("cuda")

text = "ユーザー: こんにちは <NL> システム: こんにちは！ <NL>ユーザー: システムちゃんの趣味は何？ <NL> システム:"
token_ids = tokenizer.encode(text, add_special_tokens=False, return_tensors="pt")

with torch.no_grad():
    output_ids = model.generate(
        token_ids.to(model.device),
        max_new_tokens=100,
        min_new_tokens=100,
        do_sample=True,
        temperature=0.8,
        pad_token_id=tokenizer.pad_token_id,
        bos_token_id=tokenizer.bos_token_id,
        eos_token_id=tokenizer.eos_token_id
    )

output = tokenizer.decode(output_ids.tolist()[0])
print(output)

# モデルの読み込み
import os

from peft.utils.config import TaskType
os.environ["CUDA_VISIBLE_DEVICES"]="0"
from transformers import AutoTokenizer, AutoConfig, AutoModelForCausalLM,T5Tokenizer
from peft import LoraConfig, get_peft_model 

model = AutoModelForCausalLM.from_pretrained(
    "rinna/japanese-gpt-neox-3.6b", 
    device_map='auto'
)
tokenizer = T5Tokenizer.from_pretrained("rinna/japanese-gpt-neox-3.6b")
config = LoraConfig(
    r=8,
    lora_alpha=32,
    lora_dropout=0.05,
    bias="none",
    task_type=TaskType.CAUSAL_LM
)

model = get_peft_model(model, config)

import transformers
from datasets import load_dataset
data = load_dataset("kunishou/databricks-dolly-15k-ja")


train_args = transformers.TrainingArguments(
        per_device_train_batch_size=4,
        gradient_accumulation_steps=4,
        gradient_checkpointing=True,
        warmup_steps=100,
        max_steps=200,
        learning_rate=2e-4,
        logging_steps=1,
        output_dir='result',
        num_train_epochs=1,
    )


trainer = transformers.Trainer(
    model=model, 
    train_dataset=data["train"],
    args=train_args,
    data_collator=transformers.DataCollatorForLanguageModeling(tokenizer, mlm=False)
)
model.config.use_cache = False  # 警告を黙らせます。 推論のために再度有効にしてください。
trainer.train()

config = LoraConfig(
    r=8,
    lora_alpha=32,
    lora_dropout=0.05,
    bias="none",
    task_type="CAUSAL_LM"
)

# モデルの読み込み
import os
os.environ["CUDA_VISIBLE_DEVICES"]="0"
import torch
import torch.nn as nn
import bitsandbytes as bnb
from transformers import AutoTokenizer, AutoConfig, AutoModelForCausalLM

model = AutoModelForCausalLM.from_pretrained(
    "rinna/japanese-gpt-neox-3.6b", 
    load_in_8bit=True, 
    device_map='auto',
)

tokenizer = AutoTokenizer.from_pretrained("rinna/japanese-gpt-neox-3.6b")


for param in model.parameters():
  param.requires_grad = False  # モデルをフリーズ
  if param.ndim == 1:
    # 安定のためにレイヤーノルムをfp32にキャスト
    param.data = param.data.to(torch.float32)

model.gradient_checkpointing_enable()
model.enable_input_require_grads()

class CastOutputToFloat(nn.Sequential):
  def forward(self, x): return super().forward(x).to(torch.float32)
def print_trainable_parameters(model):
    """
    モデル内の学習可能なパラメータ数を出力
    """
    trainable_params = 0
    all_param = 0
    for _, param in model.named_parameters():
        all_param += param.numel()
        if param.requires_grad:
            trainable_params += param.numel()
    print(
        f"trainable params: {trainable_params} || all params: {all_param} || trainable%: {100 * trainable_params / all_param}"
    )


from peft import LoraConfig, get_peft_model 

config = LoraConfig(
    r=16,
    lora_alpha=32,
    lora_dropout=0.05,
    bias="none",
    task_type="CAUSAL_LM"
)

model = get_peft_model(model, config)
print_trainable_parameters(model)

import transformers
from datasets import load_dataset
data = load_dataset("kunishou/databricks-dolly-15k-ja")
data = data.map(lambda samples: tokenizer(samples['output']), batched=True)

trainer = transformers.Trainer(
    model=model, 
    train_dataset=data['train'],
    args=transformers.TrainingArguments(
        per_device_train_batch_size=4, 
        gradient_accumulation_steps=4,
        warmup_steps=100, 
        max_steps=200, 
        learning_rate=2e-4, 
        fp16=True,
        logging_steps=1, 
        output_dir='outputs'
    ),
    data_collator=transformers.DataCollatorForLanguageModeling(tokenizer, mlm=False)
)
model.config.use_cache = False  # 警告を黙らせます。 推論のために再度有効にしてください。
trainer.train()

batch = tokenizer("Two things are infinite: ", return_tensors='pt')

with torch.cuda.amp.autocast():
  output_tokens = model.generate(**batch, max_new_tokens=50)

print('\n\n', tokenizer.decode(output_tokens[0], skip_special_tokens=True))

    attn_output, attn_weights = self._attn(query, key, value, attention_mask, head_mask)
  File "D:\dev\AI_model\peft\.venv\lib\site-packages\transformers\models\gpt_neox\modeling_gpt_neox.py", line 229, in _attn
    attn_scores = torch.where(causal_mask, attn_scores, mask_value)
RuntimeError: The size of tensor a (2048) must match the size of tensor b (2439) at non-singleton dimension 3

import os
import peft
from peft.utils.config import TaskType
os.environ["CUDA_VISIBLE_DEVICES"]="0"
import torch
import torch.nn as nn
import bitsandbytes as bnb
import transformers
from datasets import load_dataset

# トークナイザ準備
tokenizer = transformers.AutoTokenizer.from_pretrained("rinna/japanese-gpt-neox-3.6b")

# モデルの読み込み
model = transformers.AutoModelForCausalLM.from_pretrained(
    "rinna/japanese-gpt-neox-3.6b", 
    device_map='auto',
    torch_dtype=torch.float16
)
model.enable_input_require_grads()
model.gradient_checkpointing_enable()

config = peft.LoraConfig(
    r=8,
    lora_alpha=32,
    lora_dropout=0.01,
    inference_mode=False,
    task_type=TaskType.CAUSAL_LM,
)

model = peft.get_peft_model(model, config)

def tokenize(batch):
    return tokenizer(
        batch["output"],
        padding="max_length",
        truncation=True
    )

# データセット準備
data = load_dataset("kunishou/databricks-dolly-15k-ja")
# データセットを訓練用と検証用に分割
train_val = data["train"].train_test_split(test_size=0.1,shuffle=True,seed=42)
train_data = train_val["train"]
val_data = train_val["test"]
data = data.map(tokenize, batched=True,batch_size=None)

train_args = transformers.TrainingArguments(
        per_device_train_batch_size=1, 
        gradient_checkpointing=True,
        logging_steps=1, 
        output_dir='outputs',
        num_train_epochs=1
    )

def generate_prompt(data_point):
    # sorry about the formatting disaster gotta move fast
    if data_point["input"]:
        return f'以下は、タスクを説明する指示と、文脈のある入力の組み合わせです。要求を適切に満たす応答を書きなさい。<NL>### 指示:<NL>{data_point["instruction"]}<NL>### 入力:<NL>{data_point["input"]}<NL>### 応答:<NL>{data_point["output"]}'
    else:
        return f'以下は、タスクを説明する指示です。要求を適切に満たす応答を書きなさい。<NL>### 指示:<NL>{data_point["instruction"]}<NL>### 応答:<NL>{data_point["output"]}'

def tokenize(prompt):
    result = tokenizer(
        prompt,
        truncation=True,
        max_length=256 + 1,
        padding="max_length",
    )
    return {
        "input_ids": result["input_ids"][:-1],
        "attention_mask": result["attention_mask"][:-1],
    }


train_data = train_data.shuffle().map(lambda x: tokenize(generate_prompt(x)))
val_data = val_data.shuffle().map(lambda x: tokenize(generate_prompt(x)))

trainer = transformers.Trainer(
    model=model,
    train_dataset=train_data,
    eval_dataset=val_data,
    args=train_args,
    data_collator=transformers.DataCollatorForLanguageModeling(tokenizer, mlm=False),
)
model.config.use_cache = False  # 警告を黙らせます。 推論のために再度有効にしてください。
trainer.train()

import os
import sys
from typing import List

import fire
import torch
import transformers
from datasets import load_dataset

"""
Unused imports:
import torch.nn as nn
import bitsandbytes as bnb
"""

from peft import (
    LoraConfig,
    get_peft_model,
    get_peft_model_state_dict,
    prepare_model_for_int8_training,
    set_peft_model_state_dict,
)
from transformers import LlamaForCausalLM, LlamaTokenizer, GPTNeoXForCausalLM, GPTNeoXTokenizerFast,AutoTokenizer

from prompter import Prompter


def train(
    # model/data params
    base_model: str = "rinna/japanese-gpt-neox-3.6b",  # the only required argument
    data_path: str = "kunishou/databricks-dolly-15k-ja",
    output_dir: str = "./dolly-rinna-3b",
    # training hyperparams
    batch_size: int = 128,
    micro_batch_size: int = 4,
    num_epochs: int = 3,
    learning_rate: float = 3e-4,
    cutoff_len: int = 256,
    val_set_size: int = 2000,
    # lora hyperparams
    lora_r: int = 8,
    lora_alpha: int = 16,
    lora_dropout: float = 0.05,
    lora_target_modules: List[str] = [
        "query_key_value",
    ],
    # llm hyperparams
    train_on_inputs: bool = True,  # if False, masks out inputs in loss
    group_by_length: bool = False,  # faster, but produces an odd training loss curve
    # wandb params
    wandb_project: str = "",
    wandb_run_name: str = "",
    wandb_watch: str = "",  # options: false | gradients | all
    wandb_log_model: str = "",  # options: false | true
    resume_from_checkpoint: str = None,  # either training checkpoint or final adapter
    prompt_template_name: str = "japanese",  # The prompt template to use, will default to alpaca.
):
    if int(os.environ.get("LOCAL_RANK", 0)) == 0:
        print(
            f"Training Alpaca-LoRA model with params:\n"
            f"base_model: {base_model}\n"
            f"data_path: {data_path}\n"
            f"output_dir: {output_dir}\n"
            f"batch_size: {batch_size}\n"
            f"micro_batch_size: {micro_batch_size}\n"
            f"num_epochs: {num_epochs}\n"
            f"learning_rate: {learning_rate}\n"
            f"cutoff_len: {cutoff_len}\n"
            f"val_set_size: {val_set_size}\n"
            f"lora_r: {lora_r}\n"
            f"lora_alpha: {lora_alpha}\n"
            f"lora_dropout: {lora_dropout}\n"
            f"lora_target_modules: {lora_target_modules}\n"
            f"train_on_inputs: {train_on_inputs}\n"
            f"group_by_length: {group_by_length}\n"
            f"wandb_project: {wandb_project}\n"
            f"wandb_run_name: {wandb_run_name}\n"
            f"wandb_watch: {wandb_watch}\n"
            f"wandb_log_model: {wandb_log_model}\n"
            f"resume_from_checkpoint: {resume_from_checkpoint or False}\n"
            f"prompt template: {prompt_template_name}\n"
        )
    assert (
        base_model
    ), "Please specify a --base_model, e.g. --base_model='huggyllama/llama-7b'"
    gradient_accumulation_steps = batch_size // micro_batch_size

    prompter = Prompter(prompt_template_name)

    device_map = "auto"
    world_size = int(os.environ.get("WORLD_SIZE", 1))
    ddp = world_size != 1
    if ddp:
        device_map = {"": int(os.environ.get("LOCAL_RANK") or 0)}
        gradient_accumulation_steps = gradient_accumulation_steps // world_size

    # Check if parameter passed or if set within environ
    use_wandb = len(wandb_project) > 0 or (
        "WANDB_PROJECT" in os.environ and len(os.environ["WANDB_PROJECT"]) > 0
    )
    # Only overwrite environ if wandb param passed
    if len(wandb_project) > 0:
        os.environ["WANDB_PROJECT"] = wandb_project
    if len(wandb_watch) > 0:
        os.environ["WANDB_WATCH"] = wandb_watch
    if len(wandb_log_model) > 0:
        os.environ["WANDB_LOG_MODEL"] = wandb_log_model

    model = GPTNeoXForCausalLM.from_pretrained(
        base_model,
        load_in_8bit=True,
        torch_dtype=torch.float16,
        device_map=device_map,
    )

    tokenizer = AutoTokenizer.from_pretrained(base_model)

    tokenizer.pad_token_id = (
        0  # unk. we want this to be different from the eos token
    )
    tokenizer.padding_side = "left"  # Allow batched inference

    def tokenize(prompt, add_eos_token=True):
        # there's probably a way to do this with the tokenizer settings
        # but again, gotta move fast
        result = tokenizer(
            prompt,
            truncation=True,
            max_length=cutoff_len,
            padding=False,
            return_tensors=None,
        )
        if (
            result["input_ids"][-1] != tokenizer.eos_token_id
            and len(result["input_ids"]) < cutoff_len
            and add_eos_token
        ):
            result["input_ids"].append(tokenizer.eos_token_id)
            result["attention_mask"].append(1)

        result["labels"] = result["input_ids"].copy()

        return result

    def generate_and_tokenize_prompt(data_point):
        full_prompt = prompter.generate_prompt(
            data_point["instruction"],
            data_point["input"],
            data_point["output"],
        )
        tokenized_full_prompt = tokenize(full_prompt)
        if not train_on_inputs:
            user_prompt = prompter.generate_prompt(
                data_point["instruction"], data_point["input"]
            )
            tokenized_user_prompt = tokenize(user_prompt, add_eos_token=False)
            user_prompt_len = len(tokenized_user_prompt["input_ids"])

            tokenized_full_prompt["labels"] = [
                -100
            ] * user_prompt_len + tokenized_full_prompt["labels"][
                user_prompt_len:
            ]  # could be sped up, probably
        return tokenized_full_prompt

    model = prepare_model_for_int8_training(model)

    config = LoraConfig(
        r=lora_r,
        lora_alpha=lora_alpha,
        target_modules=lora_target_modules,
        lora_dropout=lora_dropout,
        bias="none",
        task_type="CAUSAL_LM",
    )
    model = get_peft_model(model, config)

    if data_path.endswith(".json") or data_path.endswith(".jsonl"):
        data = load_dataset("json", data_files=data_path)
    else:
        data = load_dataset(data_path)

    if resume_from_checkpoint:
        # Check the available weights and load them
        checkpoint_name = os.path.join(
            resume_from_checkpoint, "pytorch_model.bin"
        )  # Full checkpoint
        if not os.path.exists(checkpoint_name):
            checkpoint_name = os.path.join(
                resume_from_checkpoint, "adapter_model.bin"
            )  # only LoRA model - LoRA config above has to fit
            resume_from_checkpoint = (
                False  # So the trainer won't try loading its state
            )
        # The two files above have a different name depending on how they were saved, but are actually the same.
        if os.path.exists(checkpoint_name):
            print(f"Restarting from {checkpoint_name}")
            adapters_weights = torch.load(checkpoint_name)
            model = set_peft_model_state_dict(model, adapters_weights)
        else:
            print(f"Checkpoint {checkpoint_name} not found")

    model.print_trainable_parameters()  # Be more transparent about the % of trainable params.

    if val_set_size > 0:
        train_val = data["train"].train_test_split(
            test_size=val_set_size, shuffle=True, seed=42
        )
        train_data = (
            train_val["train"].shuffle().map(generate_and_tokenize_prompt)
        )
        val_data = (
            train_val["test"].shuffle().map(generate_and_tokenize_prompt)
        )
    else:
        train_data = data["train"].shuffle().map(generate_and_tokenize_prompt)
        val_data = None

    if not ddp and torch.cuda.device_count() > 1:
        # keeps Trainer from trying its own DataParallelism when more than 1 gpu is available
        model.is_parallelizable = True
        model.model_parallel = True

    trainer = transformers.Trainer(
        model=model,
        train_dataset=train_data,
        eval_dataset=val_data,
        args=transformers.TrainingArguments(
            per_device_train_batch_size=micro_batch_size,
            gradient_accumulation_steps=gradient_accumulation_steps,
            warmup_steps=100,
            num_train_epochs=num_epochs,
            learning_rate=learning_rate,
            fp16=True,
            logging_steps=10,
            optim="adamw_torch",
            evaluation_strategy="steps" if val_set_size > 0 else "no",
            save_strategy="steps",
            eval_steps=200 if val_set_size > 0 else None,
            save_steps=200,
            output_dir=output_dir,
            save_total_limit=3,
            load_best_model_at_end=True if val_set_size > 0 else False,
            ddp_find_unused_parameters=False if ddp else None,
            group_by_length=group_by_length,
            report_to="wandb" if use_wandb else None,
            run_name=wandb_run_name if use_wandb else None,
        ),
        data_collator=transformers.DataCollatorForSeq2Seq(
            tokenizer, pad_to_multiple_of=8, return_tensors="pt", padding=True
        ),
    )
    model.config.use_cache = False

    old_state_dict = model.state_dict
    model.state_dict = (
        lambda self, *_, **__: get_peft_model_state_dict(
            self, old_state_dict()
        )
    ).__get__(model, type(model))

    if torch.__version__ >= "2" and sys.platform != "win32":
        model = torch.compile(model)

    trainer.train(resume_from_checkpoint=resume_from_checkpoint)

    model.save_pretrained(output_dir)

    print(
        "\n If there's a warning about missing keys above, please disregard :)"
    )


if __name__ == "__main__":
    fire.Fire(train)

from transformers import T5Tokenizer, AutoModelForCausalLM,TrainingArguments,Trainer
from torch.utils.data import Dataset
import os

os.environ["WANDB_DISABLED"] = "true"
model_name = "rinna/japanese-gpt-neox-3.6b"
class TextDataset(Dataset):
    def __init__(self, txt_file, tokenizer):
        self.tokenizer = tokenizer
        with open(txt_file, 'r', encoding='utf-8') as f:
            lines = f.read().split('\n')
        self.lines = lines

    def __getitem__(self, idx):
        line = self.lines[idx]
        encoding = self.tokenizer.encode_plus(
            line,
            truncation=True,
            max_length=256,
            padding='max_length',
            return_tensors='pt'
        )
        encoding['input_ids'] = encoding['input_ids'].squeeze()  # バッチ次元を除去
        encoding['labels'] = encoding['input_ids'].clone()  # ラベルとして入力IDを複製
        return encoding

    def __len__(self):
        return len(self.lines)

pre_trained_model = AutoModelForCausalLM.from_pretrained(model_name)

tokenizer = T5Tokenizer.from_pretrained(model_name)

dataset = TextDataset("dataset_20230525.txt", tokenizer)

training_args = TrainingArguments(
    output_dir="./results",  # 出力先のディレクトリ
    overwrite_output_dir=True,  # 上書きするか
    num_train_epochs=3,  # 学習エポック数
    per_device_train_batch_size=1,  # 訓練のバッチサイズ
    per_device_eval_batch_size=64,   # 評価のバッチサイズ
    warmup_steps=500,                # 学習率スケジューラのウォームアップステップ数
    weight_decay=0.005,               # 重み減衰の強さ
    logging_dir='./logs',            # ログ保存フォルダ
    fp16=True,                       # 半精度演算を有効にするか
    gradient_accumulation_steps=2048,   # 勾配の累積ステップ数
    gradient_checkpointing=True,        # チェックポイントを有効にするか
    optim="adafactor",                      # 最適化手法 (adamw, adafactor)
)

trainer = Trainer(
    model=pre_trained_model,                         # 学習させるモデル
    args=training_args,                              # 学習設定
    train_dataset=dataset                           # 訓練データ
)

trainer.train()
print("学習が終了しました")
trainer.save_model("./trained_model")

# モデルの読み込み
import os

from peft.utils.config import TaskType
os.environ["CUDA_VISIBLE_DEVICES"]="0"
import torch
import torch.nn as nn
import bitsandbytes as bnb
from peft import LoraConfig, get_peft_model 
import transformers
from datasets import load_dataset

# 基本パラメータ
model_name = "rinna/japanese-gpt-neox-3.6b"
# dataset = "kunishou/databricks-dolly-15k-ja"
is_dataset_local = False
dataset_file_path = "./20230528_sakura.json"
peft_name = "lora-sakura-3.6b"
output_dir = "lora-sakura-3.6b-results"

# トレーニング用パラメータ
eval_steps = 50 #200
save_steps = 100 #200
logging_steps = 50 #20
max_steps = 100 # dollyだと 4881

# データセットの準備
# data = load_dataset(dataset)
data = load_dataset("json",data_files=dataset_file_path)
CUTOFF_LEN = 512  # コンテキスト長の上限

# Rinnaのトークナイザーでは、「use_fast=False」も必要になる
tokenizer = transformers.AutoTokenizer.from_pretrained(model_name, use_fast=False)

def print_trainable_parameters(model):
    """
    モデル内の学習可能なパラメータ数を出力
    """
    trainable_params = 0
    all_param = 0
    for _, param in model.named_parameters():
        all_param += param.numel()
        if param.requires_grad:
            trainable_params += param.numel()
    print(
        f"trainable params: {trainable_params} || all params: {all_param} || trainable%: {100 * trainable_params / all_param}"
    )


model = transformers.AutoModelForCausalLM.from_pretrained(
    model_name, 
    device_map='auto',
    load_in_8bit=True,
)
model.enable_input_require_grads()
model.gradient_checkpointing_enable()

config = LoraConfig(
    r=8,
    lora_alpha=32,
    lora_dropout=0.01,
    inference_mode=False,
    task_type=TaskType.CAUSAL_LM,
)

model = get_peft_model(model, config)
print_trainable_parameters(model)



# トークナイズ
def tokenize(prompt, tokenizer):
    result = tokenizer(
        prompt,
        truncation=True,
        max_length=CUTOFF_LEN,
        padding=False,
    )
    return {
        "input_ids": result["input_ids"],
        "attention_mask": result["attention_mask"],
    }

# プロンプトテンプレートの準備
def generate_prompt(data_point):
    if data_point["input"]:
        result = f"""### 指示:
{data_point["instruction"]}

### 入力:
{data_point["input"]}

### 回答:
{data_point["output"]}"""
    else:
        result = f"""### 指示:
{data_point["instruction"]}

### 回答:
{data_point["output"]}"""

    # 改行→<NL>
    result = result.replace('\n', '<NL>')
    return result

VAL_SET_SIZE = 10
# 学習データと検証データの準備
train_val = data["train"].train_test_split(
    test_size=VAL_SET_SIZE, shuffle=True, seed=42
)
train_data = train_val["train"]
val_data = train_val["test"]
train_data = train_data.shuffle().map(lambda x: tokenize(generate_prompt(x), tokenizer))
val_data = val_data.shuffle().map(lambda x: tokenize(generate_prompt(x), tokenizer))


trainer = transformers.Trainer(
    model=model, 
    train_dataset=train_data,
    eval_dataset=val_data,
    args=transformers.TrainingArguments(
        num_train_epochs=3,
        learning_rate=3e-4,
        logging_steps=logging_steps,
        evaluation_strategy="steps",
        save_strategy="steps",
        max_steps=max_steps,
        eval_steps=eval_steps,
        save_steps=save_steps,
        output_dir=output_dir,
        report_to="none",
        save_total_limit=3,
        push_to_hub=False,
        auto_find_batch_size=True
    ),
    data_collator=transformers.DataCollatorForLanguageModeling(tokenizer, mlm=False),
)
model.config.use_cache = False  # 警告を黙らせます。 推論のために再度有効にしてください。
trainer.train()
model.config.use_cache = True
# LoRAモデルの保存
trainer.model.save_pretrained(peft_name)
print("Done!")