app.py

import os
import torch
import gradio as gr
from transformers import (
    AutoConfig,
    AutoTokenizer,
    AutoModelForCausalLM
)

# ==================================================
# GOSHAWK AI — Hugging Face Space
# ==================================================

APP_NAME = "Goshawk AI"
MODEL_PATH = os.getenv("MODEL_PATH", "./")

MAX_NEW_TOKENS = 256
MAX_CONTEXT = 2048

device = "cuda" if torch.cuda.is_available() else "cpu"
dtype = torch.float16 if device == "cuda" else torch.float32

print(f"[{APP_NAME}] Device: {device}")
print(f"[{APP_NAME}] Model path: {MODEL_PATH}")

tokenizer = None
model = None
load_error = None

try:
    config = AutoConfig.from_pretrained(
        MODEL_PATH,
        local_files_only=True,
        trust_remote_code=False
    )

    print(f"Model architecture: {config.model_type}")

    tokenizer = AutoTokenizer.from_pretrained(
        MODEL_PATH,
        local_files_only=True,
        trust_remote_code=False
    )

    model = AutoModelForCausalLM.from_pretrained(
        MODEL_PATH,
        config=config,
        torch_dtype=dtype,
        low_cpu_mem_usage=True,
        local_files_only=True,
        trust_remote_code=False
    )

    model.to(device)
    model.eval()

    if tokenizer.pad_token_id is None:
        tokenizer.pad_token = tokenizer.eos_token

    print(f"[{APP_NAME}] Model loaded successfully.")

except Exception as exc:
    load_error = f"{type(exc).__name__}: {exc}"
    print(f"[{APP_NAME}] Loading failed: {load_error}")


SYSTEM_PROMPT = (
    "You are Goshawk AI, a helpful and precise AI assistant. "
    "Answer in the user's language. Be transparent about uncertainty. "
    "Never invent facts, live market data, or sources."
)


def build_prompt(message, history):
    messages = [
        {"role": "system", "content": SYSTEM_PROMPT}
    ]

    for item in (history or [])[-8:]:
        if isinstance(item, dict):
            role = item.get("role")
            content = item.get("content", "")

            if role in ("user", "assistant") and isinstance(content, str):
                messages.append({
                    "role": role,
                    "content": content
                })

        elif isinstance(item, (list, tuple)) and len(item) == 2:
            if item[0]:
                messages.append({
                    "role": "user",
                    "content": str(item[0])
                })
            if item[1]:
                messages.append({
                    "role": "assistant",
                    "content": str(item[1])
                })

    messages.append({"role": "user", "content": message})

    if hasattr(tokenizer, "apply_chat_template"):
        try:
            return tokenizer.apply_chat_template(
                messages,
                tokenize=False,
                add_generation_prompt=True
            )
        except Exception:
            pass

    # Fallback for models without a chat template.
    prompt = f"System: {SYSTEM_PROMPT}\n"
    for msg in messages[1:]:
        label = "User" if msg["role"] == "user" else "Assistant"
        prompt += f"{label}: {msg['content']}\n"

    return prompt + "Assistant:"


def respond(message, history, temperature, max_tokens):
    if not message or not message.strip():
        yield "Lütfen bir mesaj yaz."
        return

    if model is None or tokenizer is None:
        yield (
            "Model yüklenemedi.\n\n"
            f"Hata: {load_error}\n\n"
            "config.json, model.safetensors ve tokenizer "
            "dosyalarını kontrol et. Model mimarisi metin "
            "üretimini desteklemiyor olabilir."
        )
        return

    try:
        prompt = build_prompt(message.strip(), history)

        inputs = tokenizer(
            prompt,
            return_tensors="pt",
            truncation=True,
            max_length=MAX_CONTEXT
        )
        inputs = {k: v.to(device) for k, v in inputs.items()}

        input_length = inputs["input_ids"].shape[1]

        if input_length >= MAX_CONTEXT:
            yield "Girdi bağlam sınırına ulaştı. Daha kısa bir mesaj dene."
            return

        with torch.inference_mode():
            output = model.generate(
                **inputs,
                max_new_tokens=int(max_tokens),
                do_sample=float(temperature) > 0,
                temperature=max(float(temperature), 0.01),
                top_p=0.9,
                repetition_penalty=1.08,
                pad_token_id=tokenizer.pad_token_id,
                eos_token_id=tokenizer.eos_token_id
            )

        new_tokens = output[0][input_length:]
        answer = tokenizer.decode(
            new_tokens,
            skip_special_tokens=True
        ).strip()

        yield answer or "Model boş yanıt üretti."

    except Exception as exc:
        yield f"Üretim hatası: {type(exc).__name__}: {exc}"


with gr.Blocks(title=APP_NAME) as demo:
    gr.Markdown(
        "# 🦅 Goshawk AI\n"
        "### Yerel model tabanlı yapay zekâ asistanı\n"
        f"**Cihaz:** `{device}`"
    )

    if load_error:
        gr.Markdown(
            "⚠️ Model yüklenemedi. Ayrıntılar sohbet alanında görünür."
        )

    chatbot = gr.ChatInterface(
        fn=respond,
        chatbot=gr.Chatbot(height=480),
        textbox=gr.Textbox(
            placeholder="Goshawk AI'ye bir soru sor...",
            lines=2
        ),
        additional_inputs=[
            gr.Slider(
                minimum=0.1,
                maximum=1.2,
                value=0.7,
                step=0.1,
                label="Yaratıcılık"
            ),
            gr.Slider(
                minimum=32,
                maximum=512,
                value=MAX_NEW_TOKENS,
                step=32,
                label="Maksimum yeni token"
            )
        ]
    )

    gr.Markdown(
        "Not: Yanıt kalitesi ve hızı kullanılan modelin "
        "mimarisine ve donanıma bağlıdır."
    )

if __name__ == "__main__":
    demo.queue().launch()