app.py

import os
import requests
from pathlib import Path
from fastapi import FastAPI, HTTPException
from pydantic import BaseModel
from llama_cpp import Llama
import logging

# ================= CONFIG =================
MODEL_URL = "https://huggingface.co/Neon-AI/Kushina/resolve/main/model.gguf"
MODEL_PATH = "model.gguf"
N_CTX = 16384
N_THREADS = 4
N_BATCH = 256
MAX_TOKENS = 16384
TEMPERATURE = 0.7
TOP_P = 0.9
# ==========================================

# ---------- Logging setup ----------
logging.basicConfig(
    level=logging.INFO,
    format="%(asctime)s [%(levelname)s] %(message)s"
)
logger = logging.getLogger("KushinaAPI")

app = FastAPI(title="Kushina API", version="1.0")
llm = None  # lazy-loaded

# ---------- Download GGUF if not present ----------
if not Path(MODEL_PATH).exists():
    try:
        logger.info("Downloading model.gguf from Hugging Face...")
        r = requests.get(MODEL_URL, stream=True)
        r.raise_for_status()
        with open(MODEL_PATH, "wb") as f:
            for chunk in r.iter_content(chunk_size=8192):
                f.write(chunk)
        logger.info("Download complete ✅")
    except Exception as e:
        logger.exception("Failed to download model.gguf")
        raise RuntimeError(f"Failed to download model.gguf: {e}")

# ---------- Lazy load llama.cpp ----------
def get_llm():
    global llm
    if llm is None:
        try:
            logger.info("Loading GGUF model into llama.cpp…")
            llm = Llama(
                model_path=MODEL_PATH,
                n_ctx=N_CTX,
                n_threads=N_THREADS,
                n_batch=N_BATCH,
                f16_kv=True,
                use_mmap=True,
                verbose=False,
            )
            logger.info("Model loaded ✅")
        except Exception as e:
            logger.exception("Failed to load GGUF model")
            raise RuntimeError(f"Failed to load GGUF model: {e}")
    return llm

# ---------- Request schema ----------
class PromptRequest(BaseModel):
    prompt: str

# ---------- System prompt ----------
SYSTEM_PROMPT = """You are Kushina.
Modes: CHAT or CODE
Rules:
- CHAT: mirror user tone, short responses, no explanations unless asked.
- CODE: output only code when user asks, no commentary.
Switch to CODE if user asks for code, script, function, program, website, api, algorithm, app.
Otherwise use CHAT.
"""

def build_prompt(user_text: str) -> str:
    return f"<|system|>\n{SYSTEM_PROMPT}\n<|user|>\n{user_text}\n<|assistant|>\n"

# ---------- API endpoints ----------
@app.get("/")
def root():
    return {"status": "ok"}

@app.post("/generate")
def generate(req: PromptRequest):
    try:
        llm_instance = get_llm()  # lazy load
        full_prompt = build_prompt(req.prompt)
        output_text = ""

        for chunk in llm_instance(
            full_prompt,
            max_tokens=MAX_TOKENS,
            temperature=TEMPERATURE,
            top_p=TOP_P,
            stream=True,
            stop=["<|user|>", "<|system|>"],
        ):
            if "choices" in chunk:
                output_text += chunk["choices"][0]["text"]

        return {"response": output_text}

    except Exception as e:
        # Instead of raising 500, return JSON with error
        print("❌ Error during generation:", e)
        return {"error": str(e)}