Run in kaggle

# =========================================================
# Install dependencies (Kaggle usually already has some)
# =========================================================
!pip install -q transformers peft accelerate bitsandbytes

# =========================================================
# Imports
# =========================================================
import torch
from transformers import AutoProcessor, AutoModelForCausalLM
from peft import PeftModel

# =========================================================
# Config
# =========================================================
BASE_MODEL = "google/gemma-4-E2B-it"
LORA_MODEL = "rahul7star/gemma_4_lora"

# =========================================================
# Load processor
# =========================================================
processor = AutoProcessor.from_pretrained(BASE_MODEL)

# =========================================================
# Load base model
# =========================================================
model = AutoModelForCausalLM.from_pretrained(
    BASE_MODEL,
    torch_dtype=torch.float16,   # safer for Kaggle GPU
    device_map="auto"
)

# =========================================================
# Load LoRA adapter on top of base model
# =========================================================
model = PeftModel.from_pretrained(model, LORA_MODEL)

# optional: merge LoRA for faster inference
model = model.merge_and_unload()

print("Model + LoRA loaded successfully 🚀")

# =========================================================
# Inference function
# =========================================================
def generate_response(user_input):
    messages = [
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": user_input},
    ]

    text = processor.apply_chat_template(
        messages,
        tokenize=False,
        add_generation_prompt=True,
        enable_thinking=False
    )

    inputs = processor(text=text, return_tensors="pt").to(model.device)
    input_len = inputs["input_ids"].shape[-1]

    with torch.no_grad():
        outputs = model.generate(
            **inputs,
            max_new_tokens=512,
            temperature=0.7,
            top_p=0.9
        )

    response = processor.decode(
        outputs[0][input_len:],
        skip_special_tokens=True
    )

    return response


# =========================================================
# Test
# =========================================================
print(generate_response("Write a short joke about saving RAM."))