main_qwen_fine_tuning.py

from unsloth import FastLanguageModel, is_bfloat16_supported
from unsloth.chat_templates import get_chat_template, train_on_responses_only
from datasets import load_dataset
from trl import SFTTrainer
from transformers import TrainingArguments, DataCollatorForSeq2Seq
import wandb
import torch
import datetime

print(f'Started the script at {datetime.datetime.now()}', flush=True)
max_seq_length = 32768 # Choose any! We auto support RoPE Scaling internally!
dtype = None # None for auto detection. Float16 for Tesla T4, V100, Bfloat16 for Ampere+
load_in_4bit = True # Use 4bit quantization to reduce memory usage. Can be False.

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "unsloth/Qwen2.5-Coder-7B-Instruct",
    max_seq_length = max_seq_length,
    dtype = dtype,
    load_in_4bit = load_in_4bit,
)
print(f'Model loaded successfully at {datetime.datetime.now()}', flush=True)

model = FastLanguageModel.get_peft_model(
    model,
    r = 64, # Choose any number > 0 ! Suggested 8, 16, 32, 64, 128
    target_modules = ["q_proj", "k_proj", "v_proj", "o_proj",
                      "gate_proj", "up_proj", "down_proj",],
    lora_alpha = 64,
    lora_dropout = 0, # Supports any, but = 0 is optimized
    bias = "none",    # Supports any, but = "none" is optimized
    # [NEW] "unsloth" uses 30% less VRAM, fits 2x larger batch sizes!
    use_gradient_checkpointing = "unsloth", # True or "unsloth" for very long context
    random_state = 3407,
    use_rslora = False,  # We support rank stabilized LoRA
    loftq_config = None, # And LoftQ
)

tokenizer = get_chat_template(
    tokenizer,
    chat_template = "qwen-2.5",
)

dataset = load_dataset("atharva2721/standardized-refined-train-aggregated", split = "train")
validation_dataset = load_dataset("atharva2721/standardized-refined-val-test-aggregated", split = "train")

wandb.init(project="codebud")

trainer = SFTTrainer(
    model = model,
    tokenizer = tokenizer,
    train_dataset = dataset,
    eval_dataset=validation_dataset,
    dataset_text_field = "text",
    max_seq_length = max_seq_length,
    data_collator = DataCollatorForSeq2Seq(tokenizer = tokenizer),
    dataset_num_proc = 4,
    packing = False, # Can make training 5x faster for short sequences.
    args = TrainingArguments(
        per_device_train_batch_size = 1,
        gradient_accumulation_steps = 4, # Fixed major bug in latest Unsloth
        warmup_ratio = 0.1,
        num_train_epochs = 3, # Set this for 1 full training run.
        #max_steps = 60,
        learning_rate = 2e-5,
        fp16 = not is_bfloat16_supported(),
        bf16 = is_bfloat16_supported(),
        eval_strategy="steps",
        eval_steps=410,
        per_device_eval_batch_size = 1,
        fp16_full_eval = not is_bfloat16_supported(),
        bf16_full_eval = is_bfloat16_supported(),
        logging_steps = 10,
        save_steps = 410,
        optim = "paged_adamw_8bit", # Save more memory
        weight_decay = 0.01,
        lr_scheduler_type = "cosine",
        seed = 3407,
        output_dir = "outputs",
        report_to = "wandb", # Use this for WandB etc
        run_name = "run-name"
    ),
)

trainer = train_on_responses_only(
    trainer,
    instruction_part = "<|im_start|>user\n",
    response_part = "<|im_start|>assistant\n",
)


tokenizer.decode(trainer.train_dataset[0]["input_ids"])

space = tokenizer(" ", add_special_tokens = False).input_ids[0]
tokenizer.decode([space if x == -100 else x for x in trainer.train_dataset[0]["labels"]])

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.", flush=True)
print(f"{start_gpu_memory} GB of memory reserved.", flush=True)
print(f'Everything initialized. Starting the training at {datetime.datetime.now()}', flush=True)

trainer_stats = trainer.train()
print(f'Successfully completed training at {datetime.datetime.now()}', flush=True)

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} %.")

print(f'Pushing model and tokenizer at {datetime.datetime.now()}', flush=True)
model.save_pretrained("models/finetuned_model_with_three_epochs_eval")  # Local saving
tokenizer.save_pretrained("models/finetuned_model_with_three_epochs_eval")
model.push_to_hub("finetuned_model_with_three_epochs_eval") # Online saving
tokenizer.push_to_hub("finetuned_model_with_three_epochs_eval") # Online saving

wandb.finish()
print(f'Run complete at {datetime.datetime.now()}', flush=True)