app.py

import os
import uvicorn
import struct
import logging
import asyncio
import resource
from fastapi import FastAPI, HTTPException, Security, Request, status
from fastapi.security import APIKeyHeader
from starlette.concurrency import run_in_threadpool
from llama_cpp import Llama
from huggingface_hub import hf_hub_download, login
from cryptography.hazmat.primitives.ciphers.aead import AESGCM

# --- ⚙️ LOGGING CONFIGURATION ---
logging.basicConfig(
    level=logging.INFO,
    format="%(asctime)s | %(levelname)s | %(message)s",
    datefmt="%H:%M:%S"
)
logger = logging.getLogger("MetanthropicNode")

# --- 🧱 HARDWARE LIMITS (14GB CEILING) ---
def set_memory_limit():
    # We set a hard ceiling to prevent OOM kills.
    limit_bytes = 14 * 1024 * 1024 * 1024 
    resource.setrlimit(resource.RLIMIT_AS, (limit_bytes, limit_bytes))
    logger.info(f"🧱 MEMORY HARD LIMIT SET: 14.0 GB")

set_memory_limit()

# --- 🔧 ENGINE CONFIGURATION ---
SOURCE_REPO = "metanthropic/metanthropic-phi3-encrypted"
SOURCE_FILE = "metanthropic-phi3-v1.mguf"
TEMP_DECRYPTED = "/dev/shm/metanthropic_v3_16.gguf"

# --- 🔑 SECURITY SECRETS ---
HF_TOKEN = os.environ.get("HF_TOKEN")
DECRYPTION_KEY = os.environ.get("DECRYPTION_KEY")
API_KEY_NAME = "x-metanthropic-key"
api_key_header = APIKeyHeader(name=API_KEY_NAME, auto_error=False)

# --- 🚦 BATCH CONTROLLER (SEMAPHORE) ---
# Allows 4 concurrent users. User #5 waits.
MAX_CONCURRENT_USERS = 4
BATCH_SEMAPHORE = asyncio.Semaphore(MAX_CONCURRENT_USERS)

app = FastAPI(title="Metanthropic Sovereign Node")

# --- 🔒 SECURITY HANDSHAKE ---
async def get_api_key(api_key: str = Security(api_key_header)):
    if not DECRYPTION_KEY:
        raise HTTPException(status_code=500, detail="Node Config Error")
    if api_key == DECRYPTION_KEY:
        return api_key
    raise HTTPException(status_code=403, detail="Access Denied")

# --- 🛠️ BOOTLOADER (DECRYPTION) ---
def initialize_engine():
    if os.path.exists(TEMP_DECRYPTED):
        logger.info("⚡ RAM-Disk Hit: Model resident.")
        return True

    if not HF_TOKEN or not DECRYPTION_KEY:
        logger.error("❌ CRITICAL: Missing Secrets")
        return False

    try:
        login(token=HF_TOKEN)
        logger.info(f"⏳ Downloading & Decrypting...")
        path = hf_hub_download(repo_id=SOURCE_REPO, filename=SOURCE_FILE, local_dir=".")
        
        key = bytes.fromhex(DECRYPTION_KEY)
        aes = AESGCM(key)
        
        with open(path, "rb") as f_in, open(TEMP_DECRYPTED, "wb") as f_out:
            nonce = f_in.read(12)
            h_len = struct.unpack("<I", f_in.read(4))[0]
            header = f_in.read(h_len)
            f_out.write(aes.decrypt(nonce, header, None))
            while chunk := f_in.read(128 * 1024 * 1024):
                f_out.write(chunk)
        
        if os.path.exists(path): os.remove(path)
        logger.info("✅ Decryption Complete.")
        return True
    except Exception as e:
        logger.error(f"❌ BOOT FAILURE: {e}")
        return False

# --- 🚀 INITIALIZE ENGINE GLOBAL ---
llm = None
if initialize_engine():
    logger.info("🚀 STARTING ENGINE: PARALLEL BATCH MODE")
    
    # 🔥 THE STABLE CONFIGURATION 🔥
    # FIXED: Reduced to 4096 to match Phi-3 Mini's hard limit.
    # 4096 Total / 4 Users = 1024 Tokens per User (Context Window).
    TOTAL_CONTEXT = 4096  
    
    llm = Llama(
        model_path=TEMP_DECRYPTED,
        
        # 1. PARALLEL SLOTS
        n_parallel=MAX_CONCURRENT_USERS, 
        
        # 2. TOTAL CONTEXT POOL
        n_ctx=TOTAL_CONTEXT, 
        
        # 3. COMPUTE DENSITY
        n_batch=512,        # Process chunks
        f16_kv=True,        # Memory precision
        
        # 4. CPU STRATEGY
        # 2 threads is optimal for Hugging Face Free Tier vCPU
        n_threads=2,
        
        use_mlock=True,     # Pin to RAM
        verbose=False       # Keep logs clean
    )
    logger.info(f"✅ ENGINE READY. Capacity: {MAX_CONCURRENT_USERS} Simultaneous Streams.")

# --- 🏥 HEALTH CHECK ---
@app.get("/")
def health_check():
    free_slots = BATCH_SEMAPHORE._value 
    return {"status": "active", "free_slots": free_slots}

# --- 📡 API ENDPOINT ---
@app.post("/v1/chat/completions")
async def chat_completion(request: Request, api_key: str = Security(get_api_key)):
    if not llm:
        raise HTTPException(status_code=503, detail="System Booting...")
    
    try:
        data = await request.json()
        messages = data.get("messages", [])
        max_tokens = data.get("max_tokens", 1024)
        temperature = data.get("temperature", 0.7)
        
        prompt = ""
        for msg in messages:
            prompt += f"<|{msg['role']}|>\n{msg['content']}<|end|>\n"
        prompt += "<|assistant|>\n"
    except Exception as e:
        raise HTTPException(status_code=400, detail=f"Bad Request: {str(e)}")

    # ⚡ THE BATCH GATE ⚡
    async with BATCH_SEMAPHORE:
        try:
            output = await run_in_threadpool(
                llm,
                prompt,
                max_tokens=max_tokens,
                temperature=temperature,
                stop=["<|end|>"],
                echo=False
            )
            
            return {
                "id": output["id"],
                "object": "chat.completion",
                "created": output["created"],
                "model": "metanthropic-phi3-parallel",
                "choices": [
                    {
                        "index": 0,
                        "message": {
                            "role": "assistant",
                            "content": output['choices'][0]['text'].strip()
                        },
                        "finish_reason": output['choices'][0]['finish_reason']
                    }
                ],
                "usage": output["usage"]
            }
        except Exception as e:
            logger.error(f"Inference Error: {e}")
            raise HTTPException(status_code=500, detail="Neural Core Error")

if __name__ == "__main__":
    uvicorn.run(app, host="0.0.0.0", port=7860, log_level="warning")