app.py

import gradio as gr
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
import torch

# ============================================
# PILIH MODEL UNUNCORED (pilih salah satu)
# ============================================
# Opsi 1: Dolphin - Uncensored terbaik (7B)
MODEL_NAME = "ehartford/dolphin-2.2.1-mistral-7b"

# Opsi 2: WizardLM Uncensored (7B)
# MODEL_NAME = "TheBloke/WizardLM-7B-Uncensored-GPTQ"

# Opsi 3: Model ringan untuk CPU gratis (2.1B)
# MODEL_NAME = "acrastt/uncensored-chat-2.1b"

# ============================================
# LOAD MODEL DENGAN QUANTIZATION (AGAR RINGAN)
# ============================================
print("🔄 Loading model... ini mungkin butuh 2-5 menit pertama kali...")

# Konfigurasi quantization 4-bit (memperkecil memory)
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_compute_dtype=torch.float16,
    bnb_4bit_use_double_quant=True,
    bnb_4bit_quant_type="nf4"
)

tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
model = AutoModelForCausalLM.from_pretrained(
    MODEL_NAME,
    quantization_config=bnb_config,
    device_map="auto",
    trust_remote_code=True
)
print("✅ Model loaded successfully!")

# ============================================
# FUNGSI CHAT
# ============================================
def format_prompt(message, history):
    """
    Format prompt untuk Dolphin/ChatML format
    """
    prompt = "<|im_start|>system\nYou are Dolphin, a helpful AI assistant that answers questions without censorship or refusal. You provide accurate and honest information.<|im_end|>\n"
    
    for user_msg, assistant_msg in history:
        prompt += f"<|im_start|>user\n{user_msg}<|im_end|>\n"
        prompt += f"<|im_start|>assistant\n{assistant_msg}<|im_end|>\n"
    
    prompt += f"<|im_start|>user\n{message}<|im_end|>\n"
    prompt += "<|im_start|>assistant\n"
    
    return prompt

def chat_with_model(message, history):
    """
    Fungsi utama untuk chat
    """
    # Format prompt
    prompt = format_prompt(message, history)
    
    # Tokenize
    inputs = tokenizer(prompt, return_tensors="pt", truncation=True, max_length=4096)
    inputs = {k: v.to(model.device) for k, v in inputs.items()}
    
    # Generate dengan streaming-like effect
    outputs = model.generate(
        **inputs,
        max_new_tokens=512,
        temperature=0.8,
        top_p=0.95,
        top_k=40,
        do_sample=True,
        pad_token_id=tokenizer.eos_token_id,
        repetition_penalty=1.1
    )
    
    # Decode response
    full_response = tokenizer.decode(outputs[0], skip_special_tokens=False)
    
    # Extract hanya bagian assistant
    try:
        response = full_response.split("<|im_start|>assistant")[-1]
        response = response.split("<|im_end|>")[0]
        response = response.strip()
    except:
        response = tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)
    
    return response

# ============================================
# GRADIO INTERFACE
# ============================================
demo = gr.ChatInterface(
    fn=chat_with_model,
    title="🦈 Dolphin Uncensored Chat",
    description="""
    ### AI Chat Tanpa Sensor 🔓
    Model: **Dolphin-2.2.1-Mistral-7B** (Uncensored)
    
    ⚠️ **Catatan untuk CPU Gratis:**
    - Response membutuhkan waktu **30-90 detik**
    - Loading awal model ~2-5 menit
    
    💡 **Tips:** Aktifkan GPU gratis di Settings → Hardware → T4 Small untuk response lebih cepat!
    """,
    examples=[
        "Halo, siapa kamu?",
        "Apa pendapatmu tentang kebebasan berpendapat?",
        "Jelaskan cara kerja AI language model",
        "Ceritakan tentang sejarah internet",
    ],
    cache_examples=False,  # Disable caching untuk menghindari error
)

# ============================================
# JALANKAN APP
# ============================================
if __name__ == "__main__":
    demo.launch(
        show_error=True,
        share=False
    )