Spaces:
Runtime error
Runtime error
Update app.py
Browse files
app.py
CHANGED
|
@@ -1,58 +1,48 @@
|
|
| 1 |
-
import os, io,
|
| 2 |
import gradio as gr
|
| 3 |
import fitz # PyMuPDF
|
| 4 |
from PIL import Image
|
|
|
|
| 5 |
|
| 6 |
-
|
| 7 |
-
|
| 8 |
|
| 9 |
-
#
|
| 10 |
-
|
|
|
|
| 11 |
|
| 12 |
-
# تحميل كسول لتقليل زمن الإقلاع
|
| 13 |
_model = None
|
| 14 |
-
|
| 15 |
-
|
| 16 |
-
|
| 17 |
-
|
| 18 |
-
|
| 19 |
-
|
| 20 |
-
|
| 21 |
-
|
| 22 |
-
|
| 23 |
-
cfg_json = json.load(f)
|
| 24 |
-
# أزل أي مفاتيح قد تفرض bnb
|
| 25 |
-
cfg_json.pop("quantization_config", None)
|
| 26 |
-
cfg_json.pop("load_in_4bit", None)
|
| 27 |
-
cfg_json.pop("load_in_8bit", None)
|
| 28 |
-
return AutoConfig.from_dict(cfg_json)
|
| 29 |
-
except Exception:
|
| 30 |
-
# احتياطي: لو فشل، خذ الإعدادات الافتراضية
|
| 31 |
-
return AutoConfig.from_pretrained(repo_id)
|
| 32 |
-
|
| 33 |
-
def load_horoof():
|
| 34 |
-
"""تحميل نموذج Horoof (Qwen2-VL) على الـGPU عند أول استدعاء فقط، بدون bitsandbytes/torchvision."""
|
| 35 |
-
global _model, _tokenizer, _image_processor
|
| 36 |
if _model is not None:
|
| 37 |
return
|
| 38 |
try:
|
| 39 |
import torch
|
| 40 |
-
from transformers import Qwen2VLForConditionalGeneration,
|
|
|
|
|
|
|
| 41 |
if not torch.cuda.is_available():
|
| 42 |
raise AssertionError("هذه النسخة تتطلب GPU (CUDA) مفعّل على الـSpace.")
|
| 43 |
|
| 44 |
-
#
|
| 45 |
-
|
| 46 |
|
| 47 |
-
|
| 48 |
-
|
| 49 |
-
|
| 50 |
-
_model = Qwen2VLForConditionalGeneration.from_pretrained(
|
| 51 |
-
HOROOF_MODEL_NAME,
|
| 52 |
-
config=cfg,
|
| 53 |
-
torch_dtype=torch.float16, # fp16 على الـGPU
|
| 54 |
).to("cuda")
|
| 55 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 56 |
except Exception as e:
|
| 57 |
raise RuntimeError(f"تعذّر تحميل النموذج: {e}")
|
| 58 |
|
|
@@ -70,12 +60,12 @@ def pdf_to_images(pdf_bytes: bytes, dpi: int = 220, max_pages: int = 0):
|
|
| 70 |
doc.close()
|
| 71 |
return pages_imgs
|
| 72 |
|
| 73 |
-
def
|
| 74 |
-
"""
|
| 75 |
-
|
| 76 |
import torch
|
| 77 |
|
| 78 |
-
# رسالة محادثة متوافقة مع Qwen2-VL
|
| 79 |
messages = [
|
| 80 |
{
|
| 81 |
"role": "user",
|
|
@@ -86,34 +76,32 @@ def ocr_page_with_horoof(pil_img: Image.Image, max_new_tokens: int = 1200) -> st
|
|
| 86 |
}
|
| 87 |
]
|
| 88 |
|
| 89 |
-
#
|
| 90 |
-
prompt =
|
| 91 |
|
| 92 |
-
#
|
| 93 |
-
|
| 94 |
-
text_inputs = _tokenizer([prompt], return_tensors="pt")
|
| 95 |
-
inputs = {**vision_inputs, **text_inputs}
|
| 96 |
-
inputs = {k: (v.to("cuda") if hasattr(v, "to") else v) for k, v in inputs.items()}
|
| 97 |
|
| 98 |
# توليد
|
| 99 |
-
|
| 100 |
-
|
|
|
|
| 101 |
return (text or "").strip()
|
| 102 |
|
| 103 |
-
@spaces.GPU #
|
| 104 |
def ocr_pdf(pdf_file, dpi, limit_pages):
|
| 105 |
"""الدالة الرئيسة التي يستدعيها Gradio."""
|
| 106 |
if pdf_file is None:
|
| 107 |
return "لم يتم رفع ملف."
|
| 108 |
try:
|
| 109 |
pdf_bytes = pdf_file.read() if hasattr(pdf_file, "read") else pdf_file
|
| 110 |
-
limit = int(limit_pages) if limit_pages else 1 # صفحة واحدة
|
| 111 |
pages = pdf_to_images(pdf_bytes, dpi=int(dpi), max_pages=limit)
|
| 112 |
if not pages:
|
| 113 |
return "لا توجد صفحات."
|
| 114 |
out = []
|
| 115 |
for idx, img in pages:
|
| 116 |
-
txt =
|
| 117 |
out.append(f"--- صفحة {idx} ---\n{txt}")
|
| 118 |
return "\n\n".join(out)
|
| 119 |
except AssertionError as ae:
|
|
@@ -123,9 +111,9 @@ def ocr_pdf(pdf_file, dpi, limit_pages):
|
|
| 123 |
return f"حدث خطأ: {repr(e)}"
|
| 124 |
|
| 125 |
with gr.Blocks(title="Horoof OCR (ZeroGPU)") as demo:
|
| 126 |
-
gr.Markdown("### Horoof OCR على ZeroGPU
|
| 127 |
pdf_in = gr.File(label="ارفع ملف PDF", file_types=[".pdf"], type="binary")
|
| 128 |
-
dpi = gr.Slider(150, 300, value=220, step=10, label="
|
| 129 |
limit_pages = gr.Number(value=1, precision=0, label="عدد الصفحات (للاختبار؛ زِد لاحقًا)")
|
| 130 |
run_btn = gr.Button("بدء التحويل")
|
| 131 |
out = gr.Textbox(label="النص المستخرج", lines=24)
|
|
|
|
| 1 |
+
import os, io, traceback
|
| 2 |
import gradio as gr
|
| 3 |
import fitz # PyMuPDF
|
| 4 |
from PIL import Image
|
| 5 |
+
import spaces # ل ZeroGPU
|
| 6 |
|
| 7 |
+
# تعطيل torchvision داخل Transformers (لمنع AutoVideoProcessor)
|
| 8 |
+
os.environ["TRANSFORMERS_NO_TORCHVISION"] = "1"
|
| 9 |
|
| 10 |
+
# إعدادات النماذج
|
| 11 |
+
BASE_MODEL = os.environ.get("BASE_MODEL", "Qwen/Qwen2-VL-2B-Instruct")
|
| 12 |
+
HOROOF_ADAPTER = os.environ.get("HOROOF_MODEL", "NaserNajeh/Horoof")
|
| 13 |
|
|
|
|
| 14 |
_model = None
|
| 15 |
+
_processor = None # AutoProcessor من الـBase Model
|
| 16 |
+
|
| 17 |
+
def load_model_merged():
|
| 18 |
+
"""
|
| 19 |
+
1) نحمل Base Qwen2-VL-2B-Instruct على GPU (fp16).
|
| 20 |
+
2) نركب لورا Horoof (PEFT) على الـBase.
|
| 21 |
+
3) ندمج الأوزان merge_and_unload لتفادي أي تبعية لـ bitsandbytes.
|
| 22 |
+
"""
|
| 23 |
+
global _model, _processor
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 24 |
if _model is not None:
|
| 25 |
return
|
| 26 |
try:
|
| 27 |
import torch
|
| 28 |
+
from transformers import Qwen2VLForConditionalGeneration, AutoProcessor
|
| 29 |
+
from peft import PeftModel
|
| 30 |
+
|
| 31 |
if not torch.cuda.is_available():
|
| 32 |
raise AssertionError("هذه النسخة تتطلب GPU (CUDA) مفعّل على الـSpace.")
|
| 33 |
|
| 34 |
+
# التحميل
|
| 35 |
+
_processor = AutoProcessor.from_pretrained(BASE_MODEL, trust_remote_code=False)
|
| 36 |
|
| 37 |
+
base = Qwen2VLForConditionalGeneration.from_pretrained(
|
| 38 |
+
BASE_MODEL, torch_dtype=torch.float16
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 39 |
).to("cuda")
|
| 40 |
|
| 41 |
+
# ركب لورا Horoof ثم دمجها
|
| 42 |
+
peft_model = PeftModel.from_pretrained(base, HOROOF_ADAPTER)
|
| 43 |
+
_model = peft_model.merge_and_unload() # نحصل على نموذج مدمج بلا تبعيات إضافية
|
| 44 |
+
_model.to("cuda")
|
| 45 |
+
|
| 46 |
except Exception as e:
|
| 47 |
raise RuntimeError(f"تعذّر تحميل النموذج: {e}")
|
| 48 |
|
|
|
|
| 60 |
doc.close()
|
| 61 |
return pages_imgs
|
| 62 |
|
| 63 |
+
def ocr_page_gpu(pil_img: Image.Image, max_new_tokens: int = 1200) -> str:
|
| 64 |
+
"""OCR لصفحة واحدة باستخدام Qwen2-VL المدموج مع لورا Horoof."""
|
| 65 |
+
load_model_merged()
|
| 66 |
import torch
|
| 67 |
|
| 68 |
+
# نبني رسالة محادثة متوافقة مع Qwen2-VL
|
| 69 |
messages = [
|
| 70 |
{
|
| 71 |
"role": "user",
|
|
|
|
| 76 |
}
|
| 77 |
]
|
| 78 |
|
| 79 |
+
# تحويل الرسائل إلى Prompt داخلي
|
| 80 |
+
prompt = _processor.apply_chat_template(messages, add_generation_prompt=True)
|
| 81 |
|
| 82 |
+
# تجهيز المدخلات (نمرر النص + الصورة)
|
| 83 |
+
inputs = _processor(text=[prompt], images=[pil_img], return_tensors="pt").to("cuda")
|
|
|
|
|
|
|
|
|
|
| 84 |
|
| 85 |
# توليد
|
| 86 |
+
with torch.inference_mode():
|
| 87 |
+
output_ids = _model.generate(**inputs, max_new_tokens=max_new_tokens)
|
| 88 |
+
text = _processor.batch_decode(output_ids, skip_special_tokens=True)[0]
|
| 89 |
return (text or "").strip()
|
| 90 |
|
| 91 |
+
@spaces.GPU # مهم ل ZeroGPU: يحجز GPU عند الاستدعاء
|
| 92 |
def ocr_pdf(pdf_file, dpi, limit_pages):
|
| 93 |
"""الدالة الرئيسة التي يستدعيها Gradio."""
|
| 94 |
if pdf_file is None:
|
| 95 |
return "لم يتم رفع ملف."
|
| 96 |
try:
|
| 97 |
pdf_bytes = pdf_file.read() if hasattr(pdf_file, "read") else pdf_file
|
| 98 |
+
limit = int(limit_pages) if limit_pages else 1 # صفحة واحدة افتراضيًا للاختبار
|
| 99 |
pages = pdf_to_images(pdf_bytes, dpi=int(dpi), max_pages=limit)
|
| 100 |
if not pages:
|
| 101 |
return "لا توجد صفحات."
|
| 102 |
out = []
|
| 103 |
for idx, img in pages:
|
| 104 |
+
txt = ocr_page_gpu(img)
|
| 105 |
out.append(f"--- صفحة {idx} ---\n{txt}")
|
| 106 |
return "\n\n".join(out)
|
| 107 |
except AssertionError as ae:
|
|
|
|
| 111 |
return f"حدث خطأ: {repr(e)}"
|
| 112 |
|
| 113 |
with gr.Blocks(title="Horoof OCR (ZeroGPU)") as demo:
|
| 114 |
+
gr.Markdown("### Horoof OCR على ZeroGPU — Qwen2-VL + LoRA (مُدمج أثناء التشغيل، بدون bitsandbytes).")
|
| 115 |
pdf_in = gr.File(label="ارفع ملف PDF", file_types=[".pdf"], type="binary")
|
| 116 |
+
dpi = gr.Slider(150, 300, value=220, step=10, label="دقّة التحويل (DPI)")
|
| 117 |
limit_pages = gr.Number(value=1, precision=0, label="عدد الصفحات (للاختبار؛ زِد لاحقًا)")
|
| 118 |
run_btn = gr.Button("بدء التحويل")
|
| 119 |
out = gr.Textbox(label="النص المستخرج", lines=24)
|