【Llama2】ファインチューニング | 7b・13b・70b

git clone https://github.com/facebookresearch/llama-recipes.git

cd llama-recipes

nano requirements.txt

pip install -r requirements.txt

jupyter lab --ip='*' --port=8888 --NotebookApp.token='' --NotebookApp.password='' --no-browser

localhost:8888

import torch
from torch import cuda,bfloat16
import transformers
from transformers import AutoTokenizer,AutoModelForCausalLM
from llama_recipes.utils.dataset_utils import get_preprocessed_dataset
from llama_recipes.configs.datasets import samsum_dataset
from transformers import TrainerCallback
from contextlib import nullcontext
from transformers import default_data_collator, Trainer, TrainingArguments

torch.cuda.is_available()

pip install -U "huggingface_hub[cli]"

from huggingface_hub import login
token = 'hf_***********************'
login(token)

model_id="meta-llama/Llama-2-7b-hf"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id, 
    trust_remote_code=True,
    token=token,
    load_in_8bit=True, 
    device_map='auto',
    torch_dtype=torch.bfloat16
)

model_id="meta-llama/Llama-2-13b-hf"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id, 
    trust_remote_code=True,
    token=token,
    load_in_8bit=True, 
    device_map='auto',
    torch_dtype=torch.bfloat16
)

quant_config = transformers.BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type='nf4',
    bnb_4bit_use_double_quant=True,
    bnb_4bit_compute_dtype=bfloat16
)
model_id="meta-llama/Llama-2-70b-hf"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id, 
    trust_remote_code=True,
    token=token,
    quantization_config=quant_config,
    device_map='auto'
)

train_dataset = get_preprocessed_dataset(tokenizer, samsum_dataset, 'train')

eval_prompt = """
Summarize this dialog:
A: Hi Tom, are you busy tomorrow’s afternoon?
B: I’m pretty sure I am. What’s up?
A: Can you go with me to the animal shelter?.
B: What do you want to do?
A: I want to get a puppy for my son.
B: That will make him so happy.
A: Yeah, we’ve discussed it many times. I think he’s ready now.
B: That’s good. Raising a dog is a tough issue. Like having a baby ;-) 
A: I'll get him one of those little dogs.
B: One that won't grow up too big;-)
A: And eat too much;-))
B: Do you know which one he would like?
A: Oh, yes, I took him there last Monday. He showed me one that he really liked.
B: I bet you had to drag him away.
A: He wanted to take it home right away ;-).
B: I wonder what he'll name it.
A: He said he’d name it after his dead hamster – Lemmy  - he's  a great Motorhead fan :-)))

Summary:
"""
model_input = tokenizer(eval_prompt, return_tensors="pt").to("cuda")
model.eval()
with torch.no_grad():
    print(tokenizer.decode(model.generate(**model_input, max_new_tokens=100)[0], skip_special_tokens=True))

---
Summary:
A: Hi Tom, are you busy tomorrow's afternoon?
B: I'm pretty sure I am. What's up?
A: Can you go with me to the animal shelter?
B: What do you want to do?
A: I want to get a puppy for my son.
B: That will make him so happy.
A: Yeah, we've discussed it many times. I think he's ready now.
B

---
Summary:
A: Hi Tom, are you busy tomorrow's afternoon?
B: I'm pretty sure I am. What's up?
A: Can you go with me to the animal shelter?
B: What do you want to do?
A: I want to get a puppy for my son.
B: That will make him so happy.
A: Yeah, we've discussed it many times. I think he's ready now.
B

---
Summary:
A: Hi Tom, are you busy tomorrow’s afternoon?
B: I’m pretty sure I am. What’s up?
A: Can you go with me to the animal shelter?.
B: What do you want to do?
A: I want to get a puppy for my son.
B: That will make him so happy.
A: Yeah, we’ve discussed it many times. I think he’s ready now.
B

model.train()
def create_peft_config(model):
    from peft import (
        get_peft_model,
        LoraConfig,
        TaskType,
        prepare_model_for_kbit_training,
    )
    peft_config = LoraConfig(
        task_type=TaskType.CAUSAL_LM,
        inference_mode=False,
        r=8,
        lora_alpha=32,
        lora_dropout=0.05,
        target_modules = ["q_proj", "v_proj"]
    )
    # prepare kbit-model for training
    model = prepare_model_for_kbit_training(model)
    model = get_peft_model(model, peft_config)
    model.print_trainable_parameters()
    return model, peft_config
# create peft config
model, lora_config = create_peft_config(model)

enable_profiler = False
output_dir = "tmp/llama-output"

config = {
    'lora_config': lora_config,
    'learning_rate': 1e-4,
    'num_train_epochs': 1,
    'gradient_accumulation_steps': 2,
    'per_device_train_batch_size': 2,
    'gradient_checkpointing': False,
}

# Set up profiler
if enable_profiler:
    wait, warmup, active, repeat = 1, 1, 2, 1
    total_steps = (wait + warmup + active) * (1 + repeat)
    schedule =  torch.profiler.schedule(wait=wait, warmup=warmup, active=active, repeat=repeat)
    profiler = torch.profiler.profile(
        schedule=schedule,
        on_trace_ready=torch.profiler.tensorboard_trace_handler(f"{output_dir}/logs/tensorboard"),
        record_shapes=True,
        profile_memory=True,
        with_stack=True)
    
    class ProfilerCallback(TrainerCallback):
        def __init__(self, profiler):
            self.profiler = profiler
            
        def on_step_end(self, *args, **kwargs):
            self.profiler.step()

    profiler_callback = ProfilerCallback(profiler)
else:
    profiler = nullcontext()

# Define training args
training_args = TrainingArguments(
    output_dir=output_dir,
    overwrite_output_dir=True,
    bf16=True,
    logging_dir=f"{output_dir}/logs",
    logging_strategy="steps",
    logging_steps=10,
    save_strategy="no",
    optim="adamw_torch_fused",
    max_steps=total_steps if enable_profiler else -1,
    **{k:v for k,v in config.items() if k != 'lora_config'}
)

with profiler:
    trainer = Trainer(
        model=model,
        args=training_args,
        train_dataset=train_dataset,
        data_collator=default_data_collator,
        callbacks=[profiler_callback] if enable_profiler else [],
    )

    # Start training
    trainer.train()

model.save_pretrained(output_dir)

model.eval()
with torch.no_grad():
    print(tokenizer.decode(model.generate(**model_input, max_new_tokens=100)[0], skip_special_tokens=True))

Summarize this dialog:
A: Hi Tom, are you busy tomorrow’s afternoon?
B: I’m pretty sure I am. What’s up?
A: Can you go with me to the animal shelter?.
B: What do you want to do?
A: I want to get a puppy for my son.
B: That will make him so happy.
A: Yeah, we’ve discussed it many times. I think he’s ready now.
B: That’s good. Raising a dog is a tough issue. Like having a baby ;-) 
A: I'll get him one of those little dogs.
B: One that won't grow up too big;-)
A: And eat too much;-))
B: Do you know which one he would like?
A: Oh, yes, I took him there last Monday. He showed me one that he really liked.
B: I bet you had to drag him away.
A: He wanted to take it home right away ;-).
B: I wonder what he'll name it.
A: He said he’d name it after his dead hamster – Lemmy  - he's  a great Motorhead fan :-)))

---
Summary:
A wants to take her son to the animal shelter to get a puppy. A took her son there last Monday. He liked a little dog. He wants to name it Lemmy after his dead hamster. 

---
Summary:
A wants to take a puppy to his son. B will go with him to the animal shelter tomorrow's afternoon.

---
Summary:
Tom will go with A to the animal shelter tomorrow to get a puppy for A's son.