from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

# Load base model
base_model = AutoModelForCausalLM.from_pretrained(
    "google/gemma-3-270m-it",
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

# Load adapter
model = PeftModel.from_pretrained(
    base_model,
    "bhismaperkasa/gemma-3-1B-it-form-generator-adapter_unslothed2048"
)
model.eval()

tokenizer = AutoTokenizer.from_pretrained("bhismaperkasa/gemma-3-1B-it-form-generator-adapter_unslothed2048")

Generate Form

python

prompt = "<start_of_turn>user\nbuatkan form login<end_of_turn>\n<start_of_turn>model\n"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

outputs = model.generate(
    **inputs,
    max_new_tokens=256,
    temperature=0.7,
    top_p=0.95,
    do_sample=True
)

result = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(result.split("<start_of_turn>model\n")[-1])

📊 Training Details

Framework: Unsloth (2x faster, 60% less VRAM)
Precision: BF16 (pure, no quantization)
Batch Size: 8
Learning Rate: 5e-5
Optimizer: AdamW 8-bit
Final Loss: ~0.23-0.25

🎓 Why LoRA Adapter?

Using adapter instead of merged model because:

✅ Smaller size (~50 MB vs ~540 MB)
✅ Easy to switch base models
✅ Better for experimentation
✅ Can combine multiple adapters

🔗 Related Models

Merged Version: bhismaperkasa/gemma-3-270m-form-generator-bf16 (if available)

⚖️ License

Apache 2.0 (following Gemma license)