Fine-Tune Llama 3 on a Cloud GPU for Under $5

import os
import torch
from datasets import Dataset
from transformers import (
    AutoTokenizer,
    AutoModelForCausalLM,
    TrainingArguments,
    BitsAndBytesConfig,
)
from peft import LoraConfig, get_peft_model, TaskType
from trl import SFTTrainer

# ── Config ─────────────────────────────────────────────────────────────
MODEL_ID  = "meta-llama/Meta-Llama-3-8B-Instruct"
HF_TOKEN  = os.environ["HF_TOKEN"]        # set via GhostNexus env vars
OUTPUT    = "./llama3-finetuned-adapter"
MAX_LEN   = 512
EPOCHS    = 1
BATCH     = 4

# ── Quantization (loads model in 4-bit — fits in 12 GB VRAM) ───────────
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_use_double_quant=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16,
)

# ── Load model & tokenizer ─────────────────────────────────────────────
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, token=HF_TOKEN)
tokenizer.pad_token = tokenizer.eos_token

model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID,
    quantization_config=bnb_config,
    device_map="auto",
    token=HF_TOKEN,
)
model.config.use_cache = False

# ── LoRA adapter config ────────────────────────────────────────────────
lora_config = LoraConfig(
    r=16,                        # rank — higher = more capacity, more VRAM
    lora_alpha=32,
    target_modules=["q_proj", "v_proj", "k_proj", "o_proj"],
    lora_dropout=0.05,
    bias="none",
    task_type=TaskType.CAUSAL_LM,
)
model = get_peft_model(model, lora_config)
model.print_trainable_parameters()
# → trainable params: ~10M (0.12% of 8B total)

# ── Dataset ────────────────────────────────────────────────────────────
# Replace with your own instruction-response pairs
examples = [
    {"text": "<|system|>You are a helpful assistant.<|user|>What is gradient descent?<|assistant|>Gradient descent is an optimization algorithm that minimizes a loss function by iteratively moving in the direction of steepest descent. Each step size is controlled by the learning rate."},
    {"text": "<|system|>You are a helpful assistant.<|user|>Explain overfitting in one sentence.<|assistant|>Overfitting happens when a model memorizes training data so well that it fails to generalize to new, unseen examples."},
    # ... add your 1000+ examples here
]
dataset = Dataset.from_list(examples)

# ── Training ───────────────────────────────────────────────────────────
training_args = TrainingArguments(
    output_dir=OUTPUT,
    num_train_epochs=EPOCHS,
    per_device_train_batch_size=BATCH,
    gradient_accumulation_steps=2,
    warmup_ratio=0.03,
    learning_rate=2e-4,
    fp16=True,
    logging_steps=10,
    save_strategy="epoch",
    report_to="none",
)

trainer = SFTTrainer(
    model=model,
    train_dataset=dataset,
    args=training_args,
    dataset_text_field="text",
    max_seq_length=MAX_LEN,
)

print("Starting fine-tuning...")
trainer.train()
trainer.save_model(OUTPUT)
print(f"Adapter saved to {OUTPUT}")

torch>=2.2.0
transformers>=4.40.0
peft>=0.10.0
trl>=0.8.6
bitsandbytes>=0.43.0
datasets>=2.18.0
accelerate>=0.28.0

import ghostnexus

client = ghostnexus.Client(api_key="gn_live_YOUR_KEY")

job = client.run(
    "finetune_llama3.py",
    task_name="llama3-qlora-run1",
    # Pass your HuggingFace token as an env var (never hardcode it)
    env={"HF_TOKEN": "hf_YOUR_TOKEN"},
)

print(f"Job dispatched: {job.job_id}")
result = job.wait(timeout=3600)  # wait up to 1 hour

print(f"Status:   {result.status}")
print(f"Duration: {result.duration_seconds:.0f}s")
print(f"Cost:     ${result.cost_credits:.4f}")
print("\n--- Training output ---")
print(result.output)

# Upload adapter to HuggingFace Hub after training
from huggingface_hub import HfApi

api = HfApi(token=HF_TOKEN)
api.upload_folder(
    folder_path=OUTPUT,
    repo_id="YOUR_HF_USERNAME/llama3-my-adapter",
    repo_type="model",
)
print("Adapter uploaded to HuggingFace Hub")