gomain.py

import os
import time
import re
import gc
import logging
from functools import lru_cache
from typing import List

import numpy as np
import psutil
import onnxruntime as ort
from transformers import AutoTokenizer

from fastapi import FastAPI, HTTPException
from pydantic import BaseModel, Field

# =====================================================
# إعدادات عامة (CPU – HuggingFace Spaces)
# =====================================================

MODEL_PATH = os.environ["MODEL_PATH"]
TOKENIZER_PATH = os.environ["TOKENIZER_PATH"]

MAXTEXTLENGTH = int(os.environ.get("MAXTEXTLENGTH", 512))
CACHESIZE = int(os.environ.get("CACHESIZE", 512))
PORT = int(os.environ.get("PORT", 7860))

DEFAULT_DIM = 256   # أفضل توازن سرعة / جودة

# تقليل logging
logging.basicConfig(level=logging.WARNING)
logger = logging.getLogger("embedding-api")

# =====================================================
# تسريع ONNX Runtime (CPU)
# =====================================================

os.environ["OMP_NUM_THREADS"] = "4"
os.environ["OMP_WAIT_POLICY"] = "ACTIVE"

sess_options = ort.SessionOptions()
sess_options.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_ALL
sess_options.intra_op_num_threads = 4
sess_options.inter_op_num_threads = 1
sess_options.enable_cpu_mem_arena = True
sess_options.execution_mode = ort.ExecutionMode.ORT_SEQUENTIAL

session = ort.InferenceSession(
    MODEL_PATH,
    sess_options=sess_options,
    providers=["CPUExecutionProvider"],
)

# =====================================================
# تحميل tokenizer مرة واحدة
# =====================================================

tokenizer = AutoTokenizer.from_pretrained(
    TOKENIZER_PATH,
    local_files_only=True,
    use_fast=True
)

# =====================================================
# تطبيع النص العربي (مع cache)
# =====================================================

@lru_cache(maxsize=1024)
def normalize_arabic(text: str) -> str:
    text = re.sub(r"[ًٌٍَُِّْـ]", "", text)
    text = re.sub(r"[إأآ]", "ا", text)
    text = re.sub(r"ى", "ي", text)
    text = re.sub(r"ؤ", "و", text)
    text = re.sub(r"ئ", "ي", text)
    text = re.sub(r"ة\b", "ه", text)
    text = re.sub(r"[^\w\s]", " ", text)
    text = re.sub(r"\s+", " ", text)
    return text.strip()

# =====================================================
# تحويل النص إلى Embedding (سريع جدًا + cache)
# =====================================================

@lru_cache(maxsize=CACHESIZE)
def text_to_embedding(text: str) -> np.ndarray:
    if not text or not text.strip():
        return None

    text = normalize_arabic(text)

    inputs = tokenizer(
        f"query: {text}",
        return_tensors="np",
        truncation=True,
        max_length=96,      # أقصر = أسرع
        padding=False,
        return_token_type_ids=False,
    )

    outputs = session.run(
        None,
        {
            "input_ids": inputs["input_ids"],
            "attention_mask": inputs["attention_mask"],
        }
    )

    # CLS embedding (الأسرع)
    vector = outputs[1][0].astype(np.float32)

    # L2 Normalize (مهم للبحث)
    np.divide(vector, np.linalg.norm(vector) + 1e-12, out=vector)

    return vector

# =====================================================
# نماذج API
# =====================================================

class TextRequest(BaseModel):
    text: str = Field(..., minlength=1, maxlength=MAXTEXTLENGTH)
    dim: int = Field(
        DEFAULT_DIM,
        ge=32,
        description="Embedding dimension (default=256)"
    )

class EmbeddingResponse(BaseModel):
    embedding: List[float]
    dimension: int
    processing_time: float

class HealthResponse(BaseModel):
    status: str
    memory_usage: str
    memory_available_gb: float
    uptime: float

# =====================================================
# إنشاء التطبيق
# =====================================================

app = FastAPI(
    title="Fast Arabic Embedding API (CPU Optimized)",
    version="3.1.0"
)

# =====================================================
# نقاط النهاية
# =====================================================

@app.get("/")
def root():
    return {
        "message": "✅ Arabic Embedding API is running",
        "docs": "/docs",
        "health": "/health"
    }

@app.get("/health", response_model=HealthResponse)
def health():
    memory = psutil.virtual_memory()
    uptime = time.time() - app.state.start_time

    return HealthResponse(
        status="healthy",
        memory_usage=f"{memory.percent}%",
        memory_available_gb=round(memory.available / (1024 ** 3), 2),
        uptime=uptime,
    )

@app.post("/query", response_model=EmbeddingResponse)
def query_endpoint(request: TextRequest):
    start_time = time.time()

    vector = text_to_embedding(request.text)
    if vector is None:
        raise HTTPException(400, "فشل إنشاء embedding")

    dim = min(request.dim, vector.shape[0])
    vector = vector[:dim]

    return EmbeddingResponse(
        embedding=vector.tolist(),
        dimension=dim,
        processing_time=time.time() - start_time
    )

# =====================================================
# startup / shutdown
# =====================================================

@app.on_event("startup")
def startup():
    app.state.start_time = time.time()

    # warm-up (مهم جدًا)
    text_to_embedding("warm up")

    logger.warning("🚀 Embedding API started")

@app.on_event("shutdown")
def shutdown():
    gc.collect()
    logger.warning("🛑 Embedding API stopped")

# =====================================================
# تشغيل السيرفر
# =====================================================

if __name__ == "__main__":
    import uvicorn

    uvicorn.run(
        "main:app",
        host="0.0.0.0",
        port=PORT,
        workers=1,      # مهم لـ HuggingFace Spaces
        access_log=False
    )