vakodiya/Viber-Indian-Law-Unsloth-Llama-3.1-8B

Uploaded model

• Developed by: vakodiya
• License: apache-2.0
• Finetuned from model : unsloth/Meta-Llama-3.1-8B-bnb-4bit

This llama model was trained 2x faster with Unsloth and Huggingface's TRL library.

Code To Train Model on Google collab:

Installing required packages

%%capture
!pip install "unsloth[colab-new] @ git+https://github.com/unslothai/unsloth.git"
from torch import __version__; from packaging.version import Version as V
xformers = "xformers==0.0.27" if V(__version__) < V("2.4.0") else "xformers"
!pip install --no-deps {xformers} trl peft accelerate bitsandbytes triton

importing required modules

import torch
from trl import SFTTrainer
from datasets import load_dataset
from transformers import TrainingArguments, TextStreamer
from unsloth.chat_templates import get_chat_template
from unsloth import FastLanguageModel, is_bfloat16_supported

Login to HuggingFace using edit Access token storing in secrets

from huggingface_hub import login
from google.colab import userdata
hf_token = userdata.get('HF_API_KEY')
login(token = hf_token)

Check if a GPU is available

import torch

if torch.cuda.is_available():
  device = torch.device("cuda")
  print("GPU is available and being used.")
else:
  device = torch.device("cpu")
  print("GPU is not available, using CPU.")

Loading model from Hugging Face

max_seq_length = 1024
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="unsloth/Meta-Llama-3.1-8B-bnb-4bit",
    max_seq_length=max_seq_length,
    load_in_4bit=True,
    dtype=None,
)
model = FastLanguageModel.get_peft_model(
    model,
    r=16,
    lora_alpha=16,
    lora_dropout=0,
    target_modules=["q_proj", "k_proj", "v_proj", "up_proj", "down_proj", "o_proj", "gate_proj"],
    use_rslora=True,
    use_gradient_checkpointing="unsloth"
)

loading and formating Dataset

raw_dataset = load_dataset("viber1/indian-law-dataset", split="train[:1000]")

# Define a simple prompt template using only Instruction and Response

alpaca_prompt = """Below is an instruction that describes a task. Write a response that appropriately completes the request.

### Instruction:
{}

### Response:
{}"""

# EOS token for marking the end of each example
EOS_TOKEN = tokenizer.eos_token

# Function to format prompts with only Instruction and Response
def formatting_prompts_func(examples):
    Instruction = examples["Instruction"]
    Response = examples["Response"]

    # Create a formatted text for each example
    texts = []
    for Instruction, Response in zip(Instruction, Response):
        # Format the text with the prompt template and add the EOS token
        text = alpaca_prompt.format(Instruction, Response) + EOS_TOKEN
        texts.append(text)

    return {"text": texts}

# Apply the formatting function to the dataset
dataset = raw_dataset.map(formatting_prompts_func, batched=True)

Using Trainer with low batch sizes, Gradient Checkpointing, LoRA and Quantization

trainer=SFTTrainer(
    model=model,
    tokenizer=tokenizer,
    train_dataset=dataset,
    dataset_text_field="text",
    max_seq_length=max_seq_length,
    dataset_num_proc=2,
    packing=True,
    args=TrainingArguments(
        learning_rate=3e-4,
        lr_scheduler_type="linear",
        per_device_train_batch_size=1,
        gradient_accumulation_steps=1,
        gradient_checkpointing=True,
        num_train_epochs=1,
        fp16=not is_bfloat16_supported(),
        bf16=is_bfloat16_supported(),
        logging_steps=1,
        optim="adamw_8bit",
        weight_decay=0.01,
        warmup_steps=10,
        output_dir="output",
        seed=0,
    ),
)

Show current memory stats

gpu_stats = torch.cuda.get_device_properties(0)
start_gpu_memory = round(torch.cuda.max_memory_reserved() / 1024 / 1024 / 1024, 3)
max_memory = round(gpu_stats.total_memory / 1024 / 1024 / 1024, 3)
print(f"GPU = {gpu_stats.name}. Max memory = {max_memory} GB.")
print(f"{start_gpu_memory} GB of memory reserved.")

Start Training

trainer_stats = trainer.train()

Show final memory and time stats

used_memory = round(torch.cuda.max_memory_reserved() / 1024 / 1024 / 1024, 3)
used_memory_for_lora = round(used_memory - start_gpu_memory, 3)
used_percentage = round(used_memory         /max_memory*100, 3)
lora_percentage = round(used_memory_for_lora/max_memory*100, 3)
print(f"{trainer_stats.metrics['train_runtime']} seconds used for training.")
print(f"{round(trainer_stats.metrics['train_runtime']/60, 2)} minutes used for training.")
print(f"Peak reserved memory = {used_memory} GB.")
print(f"Peak reserved memory for training = {used_memory_for_lora} GB.")
print(f"Peak reserved memory % of max memory = {used_percentage} %.")
print(f"Peak reserved memory for training % of max memory = {lora_percentage} %.")

Finally Saving Trained model and push to HuggingFace

# Merge to 16bit
model.save_pretrained_merged("Indian-Law-Llama-3.1-8B", tokenizer, save_method = "merged_16bit",)

model.push_to_hub_merged("vakodiya/Viber-Indian-Law-Unsloth-Llama-3.1-8B", tokenizer, save_method="merged_16bit", token = hf_token)

Model usage with streaming response

# alpaca_prompt = Copied from above
FastLanguageModel.for_inference(model) # Enable native 2x faster inference
inputs = tokenizer(
[
   alpaca_prompt.format(
        "What is the difference between a petition and a plaint in Indian law?",''
    )
], return_tensors = "pt").to("cuda")

from transformers import TextStreamer
text_streamer = TextStreamer(tokenizer)
_ = model.generate(**inputs, streamer = text_streamer, max_new_tokens = 128)