handler.py

"""
Hugging Face Inference Endpoint — dots.ocr CPU handler.

Bu dosyayi Hub repo kokune yukle (handler.py).
HF modeli /repository altina mount eder; GPU aramaz.
"""

from __future__ import annotations

import base64
import io
import os
from typing import Any

os.environ.setdefault("LOCAL_RANK", "0")
os.environ.setdefault("CUDA_VISIBLE_DEVICES", "")

import torch
from PIL import Image
from qwen_vl_utils import process_vision_info
from transformers import AutoModelForCausalLM, AutoProcessor

PROMPT_OCR = "Extract the text content from this image."


class EndpointHandler:
    def __init__(self, path: str = "") -> None:
        model_path = path or os.environ.get("MODEL_DIR") or "/repository"
        self.processor = AutoProcessor.from_pretrained(model_path, trust_remote_code=True)
        try:
            self.model = AutoModelForCausalLM.from_pretrained(
                model_path,
                trust_remote_code=True,
                torch_dtype=torch.float32,
                device_map="cpu",
                low_cpu_mem_usage=True,
                attn_implementation="sdpa",
            )
        except Exception:
            self.model = AutoModelForCausalLM.from_pretrained(
                model_path,
                trust_remote_code=True,
                torch_dtype=torch.float32,
                device_map="cpu",
                low_cpu_mem_usage=True,
                attn_implementation="eager",
            )
        self.model.eval()

    def _load_image(self, raw: Any) -> Image.Image:
        if isinstance(raw, Image.Image):
            img = raw
        elif isinstance(raw, str):
            data = raw.split(",", 1)[1] if raw.startswith("data:") else raw
            img = Image.open(io.BytesIO(base64.b64decode(data)))
        elif isinstance(raw, (bytes, bytearray)):
            img = Image.open(io.BytesIO(raw))
        else:
            raise ValueError("inputs: base64 string veya data:image/... beklenir")
        img = img.convert("RGB")
        w, h = img.size
        m = max(w, h)
        if m > 1024:
            s = 1024 / float(m)
            img = img.resize((max(32, int(w * s)), max(32, int(h * s))), Image.Resampling.LANCZOS)
        return img

    def __call__(self, data: dict[str, Any]) -> dict[str, Any]:
        inputs = data.get("inputs", data)
        params = data.get("parameters") or {}
        if isinstance(inputs, dict):
            raw = inputs.get("image") or inputs.get("image_url") or inputs.get("data")
            prompt = inputs.get("prompt") or params.get("prompt") or PROMPT_OCR
        else:
            raw = inputs
            prompt = params.get("prompt") or PROMPT_OCR
        image = self._load_image(raw)
        messages = [
            {
                "role": "user",
                "content": [
                    {"type": "image", "image": image},
                    {"type": "text", "text": str(prompt)},
                ],
            }
        ]
        text = self.processor.apply_chat_template(
            messages, tokenize=False, add_generation_prompt=True
        )
        image_inputs, video_inputs = process_vision_info(messages)
        kwargs = {
            "text": [text],
            "images": image_inputs,
            "padding": True,
            "return_tensors": "pt",
        }
        if video_inputs:
            kwargs["videos"] = video_inputs
        batch = self.processor(**kwargs)
        max_new = int(params.get("max_new_tokens") or 2048)
        with torch.inference_mode():
            out_ids = self.model.generate(**batch, max_new_tokens=max_new)
        trimmed = [o[len(i) :] for i, o in zip(batch.input_ids, out_ids)]
        text_out = self.processor.batch_decode(
            trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
        )[0]
        return {"generated_text": text_out, "device": "cpu"}