train_lora.py

import os
import torch

from datasets import load_dataset
from transformers import (
    AutoTokenizer,
    AutoModelForCausalLM,
    TrainingArguments,
    Trainer,
    DataCollatorForLanguageModeling,
)

from peft import (
    LoraConfig,
    get_peft_model,
)


# ============================================================
# CONFIGURATION
# ============================================================

MODEL_NAME = "Qwen/Qwen2.5-3B-Instruct"

TRAIN_FILE = "/root/ai-tuning/train.jsonl"
VAL_FILE = "/root/ai-tuning/validation.jsonl"

OUTPUT_DIR = "/root/ai-tuning/lora-output"

# 256 is faster for CPU training.
# Use 512 later if your dataset contains long logs/configs.
MAX_LENGTH = 256


# ============================================================
# CPU CONFIGURATION
# ============================================================

CPU_COUNT = os.cpu_count() or 4

# Use available CPU threads
torch.set_num_threads(CPU_COUNT)

# Keep inter-op threads smaller to avoid CPU overhead
torch.set_num_interop_threads(2)

print("\n==========================================")
print("SYSTEM INFORMATION")
print("==========================================")

print("CPU threads       :", CPU_COUNT)
print("PyTorch threads   :", torch.get_num_threads())
print("CUDA available    :", torch.cuda.is_available())
print("PyTorch version   :", torch.__version__)


# ============================================================
# LOAD TOKENIZER
# ============================================================

print("\n==========================================")
print("LOADING TOKENIZER")
print("==========================================")

tokenizer = AutoTokenizer.from_pretrained(
    MODEL_NAME,
    use_fast=True,
)

if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.eos_token

print("Tokenizer loaded")


# ============================================================
# LOAD DATASET
# ============================================================

print("\n==========================================")
print("LOADING DATASET")
print("==========================================")

dataset = load_dataset(
    "json",
    data_files={
        "train": TRAIN_FILE,
        "validation": VAL_FILE,
    },
)

print(dataset)

print(
    "Training examples   :",
    len(dataset["train"])
)

print(
    "Validation examples :",
    len(dataset["validation"])
)


# ============================================================
# FORMAT DATASET
# ============================================================

def format_prompt(example):

    instruction = example.get(
        "instruction",
        ""
    )

    user_input = example.get(
        "input",
        ""
    )

    output = example.get(
        "output",
        ""
    )

    if user_input:

        text = (
            "### Instruction:\n"
            + instruction
            + "\n\n"
            "### Input:\n"
            + user_input
            + "\n\n"
            "### Response:\n"
            + output
        )

    else:

        text = (
            "### Instruction:\n"
            + instruction
            + "\n\n"
            "### Response:\n"
            + output
        )

    return text


# ============================================================
# TOKENIZATION
# ============================================================

def tokenize_function(example):

    text = format_prompt(example)

    tokens = tokenizer(
        text,
        truncation=True,
        max_length=MAX_LENGTH,
        padding=False,
    )

    return tokens


print("\n==========================================")
print("TOKENIZING DATASET")
print("==========================================")

tokenized_dataset = dataset.map(
    tokenize_function,
    remove_columns=dataset["train"].column_names,
    desc="Tokenizing",
)

print(tokenized_dataset)


# ============================================================
# LOAD BASE MODEL
# ============================================================

print("\n==========================================")
print("LOADING QWEN2.5-3B-INSTRUCT")
print("==========================================")

model = AutoModelForCausalLM.from_pretrained(
    MODEL_NAME,
    torch_dtype=torch.float32,
)

model.config.pad_token_id = tokenizer.pad_token_id

print("Base model loaded")


# ============================================================
# LOADING MODEL INFORMATION
# ============================================================

print("\n==========================================")
print("MODEL INFORMATION")
print("==========================================")

total_parameters = sum(
    p.numel()
    for p in model.parameters()
)

print(
    "Total parameters:",
    f"{total_parameters:,}"
)


# ============================================================
# LORA CONFIGURATION
# ============================================================

print("\n==========================================")
print("APPLYING LoRA")
print("==========================================")

lora_config = LoraConfig(

    # LoRA rank
    r=8,

    # LoRA scaling
    lora_alpha=16,

    # Dropout
    lora_dropout=0.05,

    # Transformer modules to train
    target_modules=[
        "q_proj",
        "k_proj",
        "v_proj",
        "o_proj",
    ],

    # Do not train bias
    bias="none",

    # Causal language model
    task_type="CAUSAL_LM",
)


model = get_peft_model(
    model,
    lora_config,
)


print("\nTrainable parameters:")

model.print_trainable_parameters()


# ============================================================
# TRAINING ARGUMENTS
# ============================================================

print("\n==========================================")
print("TRAINING CONFIGURATION")
print("==========================================")

training_args = TrainingArguments(

    # Output
    output_dir=OUTPUT_DIR,

    # --------------------------------------------------------
    # CPU
    # --------------------------------------------------------

    use_cpu=True,

    # --------------------------------------------------------
    # Training
    # --------------------------------------------------------

    num_train_epochs=1,

    per_device_train_batch_size=1,

    per_device_eval_batch_size=1,

    # CPU speed optimization
    gradient_accumulation_steps=1,

    # LoRA learning rate
    learning_rate=2e-4,

    # --------------------------------------------------------
    # Logging
    # --------------------------------------------------------

    logging_strategy="steps",

    logging_steps=1,

    # --------------------------------------------------------
    # Evaluation
    # --------------------------------------------------------

    # Disabled for first fast training run
    eval_strategy="no",

    # --------------------------------------------------------
    # Checkpoints
    # --------------------------------------------------------

    # Disabled to reduce disk I/O
    save_strategy="no",

    # --------------------------------------------------------
    # Optimizer
    # --------------------------------------------------------

    optim="adamw_torch",

    # --------------------------------------------------------
    # Precision
    # --------------------------------------------------------

    fp16=False,

    bf16=False,

    # --------------------------------------------------------
    # DataLoader
    # --------------------------------------------------------

    dataloader_num_workers=0,

    # --------------------------------------------------------
    # Reporting
    # --------------------------------------------------------

    report_to="none",

    # --------------------------------------------------------
    # Other
    # --------------------------------------------------------

    remove_unused_columns=False,
)


print("Epochs                 :", 1)
print("Train batch size       :", 1)
print("Gradient accumulation  :", 1)
print("Learning rate          :", 2e-4)
print("Max sequence length    :", MAX_LENGTH)
print("Evaluation             : Disabled")
print("Checkpoint saving      : Disabled")


# ============================================================
# DATA COLLATOR
# ============================================================

print("\n==========================================")
print("CREATING DATA COLLATOR")
print("==========================================")

data_collator = DataCollatorForLanguageModeling(

    tokenizer=tokenizer,

    mlm=False,
)


# ============================================================
# TRAINER
# ============================================================

print("\n==========================================")
print("CREATING TRAINER")
print("==========================================")

trainer = Trainer(

    model=model,

    args=training_args,

    train_dataset=tokenized_dataset["train"],

    # No evaluation during this fast first run
    eval_dataset=None,

    tokenizer=tokenizer,

    data_collator=data_collator,
)


# ============================================================
# START TRAINING
# ============================================================

print("\n")
print("==========================================")
print("STARTING LoRA TRAINING")
print("==========================================")
print("")
print("Model       :", MODEL_NAME)
print("Dataset     :", TRAIN_FILE)
print("Output      :", OUTPUT_DIR)
print("CPU threads :", CPU_COUNT)
print("Max length  :", MAX_LENGTH)
print("")
print("Training started...")
print("")


train_result = trainer.train()


# ============================================================
# TRAINING RESULTS
# ============================================================

print("\n==========================================")
print("TRAINING RESULTS")
print("==========================================")

print(
    "Training loss:",
    train_result.training_loss
)

print(
    "Training time:",
    train_result.metrics.get(
        "train_runtime",
        "N/A"
    ),
    "seconds"
)


# ============================================================
# SAVE LoRA ADAPTER
# ============================================================

print("\n==========================================")
print("SAVING LoRA ADAPTER")
print("==========================================")

os.makedirs(
    OUTPUT_DIR,
    exist_ok=True
)

trainer.save_model(
    OUTPUT_DIR
)

tokenizer.save_pretrained(
    OUTPUT_DIR
)


# ============================================================
# VERIFY OUTPUT
# ============================================================

print("\n==========================================")
print("VERIFYING OUTPUT")
print("==========================================")

for root, dirs, files in os.walk(OUTPUT_DIR):

    for file in files:

        path = os.path.join(
            root,
            file
        )

        size = os.path.getsize(path)

        print(
            f"{path} "
            f"({size / 1024 / 1024:.2f} MB)"
        )


# ============================================================
# COMPLETE
# ============================================================

print("\n")
print("==========================================")
print("TRAINING COMPLETE")
print("==========================================")

print(
    "LoRA adapter:",
    OUTPUT_DIR
)

print("")
print("Next step:")
print("1. Test LoRA adapter")
print("2. Compare Base vs LoRA")
print("3. Merge adapter")
print("4. Convert to GGUF")
print("5. Deploy in Ollama")
print("")
print("==========================================")