Azure Machine learning Fine tuning LLama2 7b

Balamurugan Balakreshnan

·

Follow

3 min read

·

Aug 19, 2023

--

2

Listen

Share

Introduction

• Fine tune LLama2 model in Azure ML
• Using Azure ML
• Using NVdia A100 GPU
• SKU NCADSA100v4
• I had to request quota increase using Azure ML to achieve this experiment
• using open source data set
• Following this experiment from here

Code

• First install necesary packages

!pip install -U pip
!pip install accelerate==0.18.0
!pip install appdirs==1.4.4
!pip install bitsandbytes==0.37.2
!pip install datasets==2.10.1
!pip install fire==0.5.0
!pip install git+https://github.com/huggingface/peft.git
!pip install git+https://github.com/huggingface/transformers.git
!pip install torch==2.0.0
!pip install sentencepiece==0.1.97
!pip install tensorboardX==2.6
!pip install gradio==3.23.0

pip install fire

pip install transformers

pip install peft

pip install datasets

pip install seaborn

pip install accelerate

pip install bitsandbytes

!pip install SentencePiece

pip install tensorboardX

pip install tensorboard

• Now import the necessary packages
• Set device to cuda

import transformers
import textwrap
from transformers import LlamaTokenizer, LlamaForCausalLM
import os
import sys
from typing import List
 
from peft import (
    LoraConfig,
    get_peft_model,
    get_peft_model_state_dict,
    prepare_model_for_int8_training,
)
 
import fire
import torch
from datasets import load_dataset
import pandas as pd
 
import matplotlib.pyplot as plt
import matplotlib as mpl
import seaborn as sns
from pylab import rcParams
 
%matplotlib inline
sns.set(rc={'figure.figsize':(10, 7)})
sns.set(rc={'figure.dpi':100})
sns.set(style='white', palette='muted', font_scale=1.2)
 
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
DEVICE

• Load the model

BASE_MODEL = "decapoda-research/llama-7b-hf"
 
model = LlamaForCausalLM.from_pretrained(
    BASE_MODEL,
    load_in_8bit=True,
    torch_dtype=torch.float16,
    device_map="auto",
)
 
tokenizer = LlamaTokenizer.from_pretrained(BASE_MODEL)
 
tokenizer.pad_token_id = (
    0  # unk. we want this to be different from the eos token
)
tokenizer.padding_side = "left"

• Now load the datesets

data = load_dataset("json", data_files="alpaca-bitcoin-sentiment-dataset.json")
data["train"]

• Create functions

CUTOFF_LEN = 10
def generate_prompt(data_point):
    return f"""Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.  # noqa: E501
### Instruction:
{data_point["instruction"]}
### Input:
{data_point["input"]}
### Response:
{data_point["output"]}"""
 
 
def tokenize(prompt, add_eos_token=True):
    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 = generate_prompt(data_point)
    tokenized_full_prompt = tokenize(full_prompt)
    return tokenized_full_prompt

• Split training and test

train_val = data["train"].train_test_split(
    test_size=200, shuffle=True, seed=42
)
train_data = (
    train_val["train"].map(generate_and_tokenize_prompt)
)
val_data = (
    train_val["test"].map(generate_and_tokenize_prompt)
)

• LORA configuration

LORA_R = 8
LORA_ALPHA = 16
LORA_DROPOUT= 0.05
LORA_TARGET_MODULES = [
    "q_proj",
    "v_proj",
]
 
BATCH_SIZE = 128
MICRO_BATCH_SIZE = 4
GRADIENT_ACCUMULATION_STEPS = BATCH_SIZE // MICRO_BATCH_SIZE
LEARNING_RATE = 3e-4
TRAIN_STEPS = 300
OUTPUT_DIR = "experiments"

• Prep for int 8 training

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)
model.print_trainable_parameters()

• Now set the training experiments

training_arguments = transformers.TrainingArguments(
    per_device_train_batch_size=MICRO_BATCH_SIZE,
    gradient_accumulation_steps=GRADIENT_ACCUMULATION_STEPS,
    warmup_steps=100,
    max_steps=TRAIN_STEPS,
    learning_rate=LEARNING_RATE,
    fp16=True,
    logging_steps=10,
    optim="adamw_torch",
    evaluation_strategy="steps",
    save_strategy="steps",
    eval_steps=50,
    save_steps=50,
    output_dir=OUTPUT_DIR,
    save_total_limit=3,
    load_best_model_at_end=True,
    report_to="tensorboard"
)

data_collator = transformers.DataCollatorForSeq2Seq(
    tokenizer, pad_to_multiple_of=8, return_tensors="pt", padding=True
)

• Now train the model

trainer = transformers.Trainer(
    model=model,
    train_dataset=train_data,
    eval_dataset=val_data,
    args=training_arguments,
    data_collator=data_collator
)
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))
 
model = torch.compile(model)
 
trainer.train()
model.save_pretrained(OUTPUT_DIR)

original article — Samples2023/AzureML/finetuneaphaca.md at main · balakreshnan/Samples2023 (github.com)

WRITER at MLearning.ai / 800+ AI plugins / AI Searching 2024