Update app.py

app.py

@@ -1,539 +1,599 @@
 import os
-import json
-import copy
-import math
+import gradio_pdf
+import hashlib
+import re
 import time
-import random
-import logging
-import numpy as np
-from typing import Any, Dict, List, Optional, Union
-import torch
-from PIL import Image
+import httpx
+import oss2
+import asyncio
+import json
+from typing import Dict, Any, Optional
+from pathlib import Path
+import click
 import gradio as gr
-import spaces
-from diffusers import (
-    DiffusionPipeline,
-    FlowMatchEulerDiscreteScheduler)
-from huggingface_hub import (
-    hf_hub_download,
-    HfFileSystem,
-    ModelCard,
-    snapshot_download)
-from diffusers.utils import load_image
-import requests
-from urllib.parse import urlparse
-import tempfile
-import shutil
-import uuid
-import zipfile
-
-
-# META: CUDA_CHECK / GPU_INFO
-device = "cuda" if torch.cuda.is_available() else "cpu"
-print("CUDA_VISIBLE_DEVICES=", os.environ.get("CUDA_VISIBLE_DEVICES"))
-print("torch.__version__ =", torch.__version__)
-print("torch.version.cuda =", torch.version.cuda)
-print("cuda available:", torch.cuda.is_available())
-print("cuda device count:", torch.cuda.device_count())
-if torch.cuda.is_available():
-    print("current device:", torch.cuda.current_device())
-    print("device name:", torch.cuda.get_device_name(torch.cuda.current_device()))
-
-print("Using device:", device)
-
-loras = [
-    # Sample Qwen-compatible LoRAs
-    {
-        "image": "https://huggingface.co/prithivMLmods/Qwen-Image-Studio-Realism/resolve/main/images/2.png",
-        "title": "Studio Realism",
-        "repo": "prithivMLmods/Qwen-Image-Studio-Realism",
-        "weights": "qwen-studio-realism.safetensors",
-        "trigger_word": "Studio Realism"
-    },
-    {
-        "image": "https://huggingface.co/prithivMLmods/Qwen-Image-Sketch-Smudge/resolve/main/images/1.png",
-        "title": "Sketch Smudge",
-        "repo": "prithivMLmods/Qwen-Image-Sketch-Smudge",
-        "weights": "qwen-sketch-smudge.safetensors",
-        "trigger_word": "Sketch Smudge"
-    },
-    {
-        "image": "https://huggingface.co/Shakker-Labs/AWPortrait-QW/resolve/main/images/08fdaf6b644b61136340d5c908ca37993e47f34cdbe2e8e8251c4c72.jpg",
-        "title": "AWPortrait QW",
-        "repo": "Shakker-Labs/AWPortrait-QW",
-        "weights": "AWPortrait-QW_1.0.safetensors",
-        "trigger_word": "Portrait"
-    },
-    {
-        "image": "https://huggingface.co/prithivMLmods/Qwen-Image-Anime-LoRA/resolve/main/images/1.png",
-        "title": "Qwen Anime",
-        "repo": "prithivMLmods/Qwen-Image-Anime-LoRA",
-        "weights": "qwen-anime.safetensors",
-        "trigger_word": "Qwen Anime"
-    },
-    {
-        "image": "https://huggingface.co/flymy-ai/qwen-image-realism-lora/resolve/main/assets/flymy_realism.png",
-        "title": "Image Realism",
-        "repo": "flymy-ai/qwen-image-realism-lora",
-        "weights": "flymy_realism.safetensors",
-        "trigger_word": "Super Realism Portrait"
-    },
-    {
-        "image": "https://huggingface.co/prithivMLmods/Qwen-Image-Fragmented-Portraiture/resolve/main/images/3.png",
-        "title": "Fragmented Portraiture",
-        "repo": "prithivMLmods/Qwen-Image-Fragmented-Portraiture",
-        "weights": "qwen-fragmented-portraiture.safetensors",
-        "trigger_word": "Fragmented Portraiture"
-    },
-    {
-        "image": "https://huggingface.co/prithivMLmods/Qwen-Image-Synthetic-Face/resolve/main/images/2.png",
-        "title": "Synthetic Face",
-        "repo": "prithivMLmods/Qwen-Image-Synthetic-Face",
-        "weights": "qwen-synthetic-face.safetensors",
-        "trigger_word": "Synthetic Face"
-    },
-    {
-        "image": "https://huggingface.co/itspoidaman/qwenglitch/resolve/main/images/GyZTwJIbkAAhS4h.jpeg",
-        "title": "Qwen Glitch",
-        "repo": "itspoidaman/qwenglitch",
-        "weights": "qwenglitch1.safetensors",
-        "trigger_word": "qwenglitch"
-    },
-    {
-        "image": "https://huggingface.co/alfredplpl/qwen-image-modern-anime-lora/resolve/main/sample1.jpg",
-        "title": "Modern Anime Lora",
-        "repo": "alfredplpl/qwen-image-modern-anime-lora",
-        "weights": "lora.safetensors",
-        "trigger_word": "Japanese modern anime style"
-    },
+from io import BytesIO
+from PIL import Image
+from gradio_pdf import PDF
+from loguru import logger
+from datetime import datetime
+
+# -- ADDED imports for model + PDF rendering --
+import torch
+from transformers import (
+    Qwen2VLForConditionalGeneration,
+    Qwen2_5_VLForConditionalGeneration,
+    AutoModelForCausalLM,
+    AutoModelForVision2Seq,
+    AutoProcessor,
+    TextIteratorStreamer,
+)
+from transformers.image_utils import load_image
+
+# Optional PDF rendering dependency fallbacks
+try:
+    import fitz  # PyMuPDF
+    _HAS_FITZ = True
+except Exception:
+    _HAS_FITZ = False
+
+try:
+    from pdf2image import convert_from_bytes
+    _HAS_PDF2IMAGE = True
+except Exception:
+    _HAS_PDF2IMAGE = False
+
+# --------- original constants and helpers ----------
+pdf_suffixes = [".pdf"]
+image_suffixes = [".png", ".jpeg", ".jpg"]
+
+latex_delimiters_type_a = [
+    {'left': '$$', 'right': '$$', 'display': True},
+    {'left': '$', 'right': '$', 'display': False},
+]
+latex_delimiters_type_b = [
+    {'left': '\\(', 'right': '\\)', 'display': False},
+    {'left': '\\[', 'right': '\\]', 'display': True},
 ]
+latex_delimiters_type_all = latex_delimiters_type_a + latex_delimiters_type_b
+
+header_path = os.path.join(os.path.dirname(os.path.dirname(__file__)), 'parsing/resources', 'header.html')
+with open(header_path, 'r') as header_file:
+    header = header_file.read()
+
+oss_access_key = os.getenv('oss_access_key')
+oss_secret_key = os.getenv('oss_secret_key')
+oss_endpoint = os.getenv('oss_endpoint')
+oss_bucket_name = os.getenv('oss_bucket_name')
+APP_KEY = os.getenv('APP_KEY')
+
+# Initialize the OSS client
+auth = oss2.Auth(oss_access_key, oss_secret_key)
+oss_bucket = oss2.Bucket(auth, oss_endpoint, oss_bucket_name)
+
+
+def upload_file_to_oss(local_data_path, oss_path):
+    with open(local_data_path, "rb") as f:
+        oss_bucket.put_object(oss_path, f)
+    url = oss_bucket.sign_url('GET', oss_path, 31536000)
+    return url
+
+
+def str_md5(input_string):
+    hasher = hashlib.md5()
+    # In Python 3, strings need to be converted to byte objects to be processed by the hash function
+    input_bytes = input_string.encode('utf-8')
+    hasher.update(input_bytes)
+    return hasher.hexdigest()
+
 
-# Initialize the base model
-dtype = torch.bfloat16
-base_model = "Qwen/Qwen-Image"
-
-# Scheduler configuration from the Qwen-Image-Lightning repository
-scheduler_config = {
-    "base_image_seq_len": 256,
-    "base_shift": math.log(3),
-    "invert_sigmas": False,
-    "max_image_seq_len": 8192,
-    "max_shift": math.log(3),
-    "num_train_timesteps": 1000,
-    "shift": 1.0,
-    "shift_terminal": None,
-    "stochastic_sampling": False,
-    "time_shift_type": "exponential",
-    "use_beta_sigmas": False,
-    "use_dynamic_shifting": True,
-    "use_exponential_sigmas": False,
-    "use_karras_sigmas": False,
-}
-
-scheduler = FlowMatchEulerDiscreteScheduler.from_config(scheduler_config)
-pipe = DiffusionPipeline.from_pretrained(
-    base_model, scheduler=scheduler, torch_dtype=dtype
-).to(device)
-
-# Lightning LoRA info (no global state)
-LIGHTNING_LORA_REPO = "lightx2v/Qwen-Image-Lightning"
-LIGHTNING_LORA_WEIGHT = "Qwen-Image-Lightning-8steps-V1.0.safetensors"
-
-MAX_SEED = np.iinfo(np.int32).max
-
-class Timer:
-    def __init__(self, task_name=""):
-        self.task_name = task_name
-
-    def __enter__(self):
-        self.start_time = time.time()
-        return self
-    
-    def __exit__(self, exc_type, exc_value, traceback):
-        self.end_time = time.time()
-        self.elapsed_time = self.end_time - self.start_time
-        if self.task_name:
-            print(f"Elapsed time for {self.task_name}: {self.elapsed_time:.6f} seconds")
+def images_bytes_to_pdf_bytes(image_bytes):
+    # Memory buffer
+    pdf_buffer = BytesIO()
+
+    # Load and convert all images to RGB mode
+    image = Image.open(BytesIO(image_bytes)).convert("RGB")
+
+    # Save the first image as a PDF and append the rest
+    image.save(pdf_buffer, format="PDF", save_all=True)
+
+    # Get PDF bytes and reset the pointer (optional)
+    pdf_bytes = pdf_buffer.getvalue()
+    pdf_buffer.close()
+    return pdf_bytes
+
+
+def read_fn(path):
+    if not isinstance(path, Path):
+        path = Path(path)
+    with open(str(path), "rb") as input_file:
+        file_bytes = input_file.read()
+        if path.suffix in image_suffixes:
+            return images_bytes_to_pdf_bytes(file_bytes)
+        elif path.suffix in pdf_suffixes:
+            return file_bytes
         else:
-            print(f"Elapsed time: {self.elapsed_time:.6f} seconds")
-
-def compute_image_dimensions(aspect_ratio):
-    """Converts aspect ratio string to width, height tuple."""
-    if aspect_ratio == "1:1":
-        return 1024, 1024
-    elif aspect_ratio == "16:9":
-        return 1152, 640
-    elif aspect_ratio == "9:16":
-        return 640, 1152
-    elif aspect_ratio == "4:3":
-        return 1024, 768
-    elif aspect_ratio == "3:4":
-        return 768, 1024
-    elif aspect_ratio == "3:2":
-        return 1024, 688
-    elif aspect_ratio == "2:3":
-        return 688, 1024
+            raise Exception(f"Unknown file suffix: {path.suffix}")
+
+
+def safe_stem(file_path):
+    stem = Path(file_path).stem
+    # Keep only letters, numbers, underscores, and dots, and replace other characters with underscores
+    return re.sub(r'[^\w.]', '_', stem)
+
+
+def sanitize_filename(filename: str, max_prefix_len: int = 15) -> str:
+    """
+    Sanitize filename: remove illegal characters, truncate, and add a hash to prevent duplicates.
+    """
+    # 1. Extract the extension
+    name, ext = '', ''
+    if '.' in filename:
+        name = filename.rsplit('.', 1)[0]
+        ext = '.' + filename.rsplit('.', 1)[1].lower()
     else:
-        return 1024, 1024
-
-def handle_lora_selection(evt: gr.SelectData, aspect_ratio):
-    selected_lora = loras[evt.index]
-    new_placeholder = f"Type a prompt for {selected_lora['title']}"
-    lora_repo = selected_lora["repo"]
-    updated_text = f"### Selected: [{lora_repo}](https://huggingface.co/{lora_repo}) ✅"
-    
-    # Update aspect ratio if specified in LoRA config
-    if "aspect" in selected_lora:
-        if selected_lora["aspect"] == "portrait":
-            aspect_ratio = "9:16"
-        elif selected_lora["aspect"] == "landscape":
-            aspect_ratio = "16:9"
-        else:
-            aspect_ratio = "1:1"
-    
-    return (
-        gr.update(placeholder=new_placeholder),
-        updated_text,
-        evt.index,
-        aspect_ratio,
-    )
-
-def adjust_generation_mode(speed_mode):
-    """Update UI based on speed/quality toggle."""
-    if speed_mode == "Fast (8 steps)":
-        return gr.update(value="Fast mode selected - 8 steps with Lightning LoRA"), 8, 1.0
-    else: 
-        return gr.update(value="Base mode selected - 50 steps for best quality"), 50, 4.0
-
-@spaces.GPU(duration=100)
-def create_image(prompt_mash, steps, seed, cfg_scale, width, height, lora_scale, negative_prompt=""):
-    pipe.to("cuda")
-    generator = torch.Generator(device="cuda").manual_seed(seed)
-    
-    with Timer("Generating image"):
-        # Generate image
-        image = pipe(
-            prompt=prompt_mash,
-            negative_prompt=negative_prompt,
-            num_inference_steps=steps,
-            true_cfg_scale=cfg_scale,  # Use true_cfg_scale for Qwen-Image
-            width=width,
-            height=height,
-            generator=generator,
-        ).images[0]
-        
-    return image
-
-@spaces.GPU(duration=100)
-def process_adapter_generation(prompt, cfg_scale, steps, selected_index, randomize_seed, seed, aspect_ratio, lora_scale, speed_mode, progress=gr.Progress(track_tqdm=True)):
-    if selected_index is None:
-        raise gr.Error("You must select a LoRA before proceeding.")
-    
-    selected_lora = loras[selected_index]
-    lora_path = selected_lora["repo"]
-    trigger_word = selected_lora["trigger_word"]
-    
-    # Prepare prompt with trigger word
-    if trigger_word:
-        if "trigger_position" in selected_lora:
-            if selected_lora["trigger_position"] == "prepend":
-                prompt_mash = f"{trigger_word} {prompt}"
+        name = filename
+        ext = ''
+
+    # 2. Remove illegal characters (Windows/Linux compatible)
+    # Allowed: letters, numbers, -_.()
+    name = re.sub(r'[\\/:\*\?"<>\|\s]+', '_', name)  # Replace spaces and illegal characters with underscores
+    name = re.sub(r'[\x00-\x1f\x7f-\x9f]', '', name)  # Remove control characters
+
+    # 3. Truncate and reserve space for the hash
+    prefix = name[:max_prefix_len]
+
+    # 4. Add an MD5 prefix hash to ensure uniqueness (based on the original path or content)
+    hash_suffix = hashlib.md5(filename.encode('utf-8')).hexdigest()[:6]
+
+    # 5. Combine
+    safe_name = f"{prefix}_{hash_suffix}{ext}"
+
+    # 6. Prevent starting or ending with a dot (sensitive in some systems)
+    while safe_name.startswith('.'):
+        safe_name = safe_name[1:]
+    if len(safe_name) == 0:
+        safe_name = f"file_{hash_suffix}.bin"
+
+    if len(safe_name.encode('utf-8')) > 250:
+        # Fallback to an absolutely safe name
+        unique_hash = hashlib.md5(filename.encode('utf-8')).hexdigest()[:8]
+        safe_name = f"doc_{unique_hash}.pdf"
+
+    return safe_name
+
+def to_pdf(file_path):
+    if file_path is None:
+        return None
+
+    pdf_bytes = read_fn(file_path)
+
+    # unique_filename = f'{uuid.uuid4()}.pdf'
+    unique_filename = f'{safe_stem(file_path)}.pdf'
+
+    # Construct the full file path
+    tmp_file_path = os.path.join(os.path.dirname(file_path), unique_filename)
+
+    # Write the byte data to the file
+    with open(tmp_file_path, 'wb') as tmp_pdf_file:
+        tmp_pdf_file.write(pdf_bytes)
+
+    return tmp_file_path
+
+
+def arg_parse(ctx: 'click.Context') -> dict:
+    # Parse extra arguments
+    extra_kwargs = {}
+    i = 0
+    while i < len(ctx.args):
+        arg = ctx.args[i]
+        if arg.startswith('--'):
+            param_name = arg[2:].replace('-', '_')  # Convert parameter name format
+            i += 1
+            if i < len(ctx.args) and not ctx.args[i].startswith('--'):
+                # The parameter has a value
+                try:
+                    # Try to convert to the appropriate type
+                    if ctx.args[i].lower() == 'true':
+                        extra_kwargs[param_name] = True
+                    elif ctx.args[i].lower() == 'false':
+                        extra_kwargs[param_name] = False
+                    elif '.' in ctx.args[i]:
+                        try:
+                            extra_kwargs[param_name] = float(ctx.args[i])
+                        except ValueError:
+                            extra_kwargs[param_name] = ctx.args[i]
+                    else:
+                        try:
+                            extra_kwargs[param_name] = int(ctx.args[i])
+                        except ValueError:
+                            extra_kwargs[param_name] = ctx.args[i]
+                except:
+                    extra_kwargs[param_name] = ctx.args[i]
             else:
-                prompt_mash = f"{prompt} {trigger_word}"
-        else:
-            prompt_mash = f"{trigger_word} {prompt}"
+                # Boolean flag parameter
+                extra_kwargs[param_name] = True
+                i -= 1
+        i += 1
+    return extra_kwargs
+
+
+# ----------------- NEW: local model integration -----------------
+# Device detection
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+logger.info(f"Using device: {device}")
+
+# Model ID - change if you want to use a different local model
+MODEL_ID_M = "Logics-MLLM/Logics-Parsing"
+
+# Load processor & model (may take time; expected)
+try:
+    logger.info(f"Loading processor and model {MODEL_ID_M} ... (this can take a while)")
+    processor_m = AutoProcessor.from_pretrained(MODEL_ID_M, trust_remote_code=True)
+    model_m = Qwen2_5_VLForConditionalGeneration.from_pretrained(
+        MODEL_ID_M,
+        trust_remote_code=True,
+        torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32
+    ).to(device).eval()
+    logger.info("Model loaded successfully.")
+except Exception as e:
+    logger.error(f"Failed to load model {MODEL_ID_M}: {e}")
+    # Do not raise here — but subsequent calls will error informatively
+    processor_m = None
+    model_m = None
+
+
+def pdf_bytes_to_images(pdf_bytes: bytes, max_pages: Optional[int] = 20):
+    """
+    Convert PDF bytes to a list of PIL Images (page thumbnails). Uses fitz if available or pdf2image as fallback.
+    Limits pages to max_pages for speed/memory reasons.
+    """
+    images = []
+    if _HAS_FITZ:
+        doc = fitz.open(stream=pdf_bytes, filetype="pdf")
+        for i in range(min(len(doc), max_pages)):
+            page = doc[i]
+            # zoom matrix to increase resolution
+            mat = fitz.Matrix(2, 2)
+            pix = page.get_pixmap(matrix=mat, alpha=False)
+            img = Image.frombytes("RGB", [pix.width, pix.height], pix.samples)
+            images.append(img)
+        doc.close()
+        return images
+    elif _HAS_PDF2IMAGE:
+        pil_images = convert_from_bytes(pdf_bytes, fmt="png")
+        return pil_images[:max_pages]
     else:
-        prompt_mash = prompt
-
-    # Always unload any existing LoRAs first to avoid conflicts
-    with Timer("Unloading existing LoRAs"):
-        pipe.unload_lora_weights()
-
-    # Load LoRAs based on speed mode
-    if speed_mode == "Fast (8 steps)":
-        with Timer("Loading Lightning LoRA and style LoRA"):
-            # Load Lightning LoRA first
-            pipe.load_lora_weights(
-                LIGHTNING_LORA_REPO, 
-                weight_name=LIGHTNING_LORA_WEIGHT,
-                adapter_name="lightning"
-            )
-            
-            # Load the selected style LoRA
-            weight_name = selected_lora.get("weights", None)
-            pipe.load_lora_weights(
-                lora_path, 
-                weight_name=weight_name, 
-                low_cpu_mem_usage=True,
-                adapter_name="style"
-            )
-            
-            # Set both adapters active with their weights
-            pipe.set_adapters(["lightning", "style"], adapter_weights=[1.0, lora_scale])
-    else:
-        # Quality mode - only load the style LoRA
-        with Timer(f"Loading LoRA weights for {selected_lora['title']}"):
-            weight_name = selected_lora.get("weights", None)
-            pipe.load_lora_weights(
-                lora_path, 
-                weight_name=weight_name, 
-                low_cpu_mem_usage=True,
-                adapter_name="style"
-            )
-            pipe.set_adapters(["style"], adapter_weights=[lora_scale])
-                
-    # Set random seed for reproducibility
-    with Timer("Randomizing seed"):
-        if randomize_seed:
-            seed = random.randint(0, MAX_SEED)
-    
-    # Get image dimensions from aspect ratio
-    width, height = compute_image_dimensions(aspect_ratio)
-    
-    # Generate the image
-    final_image = create_image(prompt_mash, steps, seed, cfg_scale, width, height, lora_scale)
-    
-    return final_image, seed
-
-def fetch_hf_adapter_files(link):
-    split_link = link.split("/")
-    if len(split_link) != 2:
-        raise Exception("Invalid Hugging Face repository link format.")
-
-    print(f"Repository attempted: {split_link}")
-    
-    # Load model card
-    model_card = ModelCard.load(link)
-    base_model = model_card.data.get("base_model")
-    print(f"Base model: {base_model}")
-
-    # Validate model type (for Qwen-Image)
-    acceptable_models = {"Qwen/Qwen-Image"}
-    
-    models_to_check = base_model if isinstance(base_model, list) else [base_model]
-    
-    if not any(model in acceptable_models for model in models_to_check):
-        raise Exception("Not a Qwen-Image LoRA!")
-        
-    # Extract image and trigger word
-    image_path = model_card.data.get("widget", [{}])[0].get("output", {}).get("url", None)
-    trigger_word = model_card.data.get("instance_prompt", "")
-    image_url = f"https://huggingface.co/{link}/resolve/main/{image_path}" if image_path else None
-
-    # Initialize Hugging Face file system
-    fs = HfFileSystem()
+        raise RuntimeError("No PDF rendering backend available. Install PyMuPDF (fitz) or pdf2image + poppler.")
+
+
+def build_parser_prompt(filename: str, num_pages: int) -> str:
+    """
+    Construct a prompt instructing the model to parse a document and return:
+    - mmd (markdown-like MMD)
+    - qwenHtml (HTML)
+    - mmdHtml (rendered HTML)
+    Return format should be clearly mark-delimited so we can extract parts.
+    """
+    prompt = f"""You are a document parsing assistant.
+Input: a multi-page PDF document named "{filename}" with {num_pages} pages (images of pages are provided).
+Task: Extract the document content in three outputs:
+---BEGIN_MMD---
+Provide the document structure and content in MMD (Markdown-like) format. Keep code blocks, equations and tables preserved.
+---END_MMD---
+
+---BEGIN_MMD_HTML---
+Provide an HTML rendering of the MMD (a full HTML fragment).
+---END_MMD_HTML---
+
+---BEGIN_QWEN_HTML---
+Provide the Qwen-specific HTML (if applicable). If none, output a short HTML wrapper.
+---END_QWEN_HTML---
+
+Only output the three sections exactly between the markers above. Do not output other commentary."""
+    return prompt
+
+
+def run_model_on_pages(images: list, filename: str, max_new_tokens: int = 2048) -> str:
+    """
+    Run the vision-language model on the provided list of PIL images and return the generated text.
+    This is a blocking function — call it inside asyncio.to_thread when used from async context.
+    """
+    if processor_m is None or model_m is None:
+        raise RuntimeError("Processor or model not loaded. Check logs — model loading failed earlier.")
+
+    # Build a single prompt (the processor will accept images+text)
+    prompt = build_parser_prompt(filename, len(images))
+
+    # Prepare inputs for processor.
+    # Many VL processors accept a single text prompt + list of images.
+    # We pass the first N images (or all) and let the model generate.
     try:
-        list_of_files = fs.ls(link, detail=False)
-        
-        # Find safetensors file
-        safetensors_name = None
-        for file in list_of_files:
-            filename = file.split("/")[-1]
-            if filename.endswith(".safetensors"):
-                safetensors_name = filename
-                break
-
-        if not safetensors_name:
-            raise Exception("No valid *.safetensors file found in the repository.")
-
+        inputs = processor_m(
+            text=[prompt],
+            images=images,
+            return_tensors="pt",
+            padding=True,
+            truncation=False,
+            max_length=4096  # conservative; may vary by model
+        )
     except Exception as e:
-        print(e)
-        raise Exception("You didn't include a valid Hugging Face repository with a *.safetensors LoRA")
-    
-    return split_link[1], link, safetensors_name, trigger_word, image_url
-
-def validate_custom_adapter(link):
-    print(f"Checking a custom model on: {link}")
-    
-    if link.endswith('.safetensors'):
-        if 'huggingface.co' in link:
-            parts = link.split('/')
-            try:
-                hf_index = parts.index('huggingface.co')
-                username = parts[hf_index + 1]
-                repo_name = parts[hf_index + 2]
-                repo = f"{username}/{repo_name}"
-                
-                safetensors_name = parts[-1]
-                
-                try:
-                    model_card = ModelCard.load(repo)
-                    trigger_word = model_card.data.get("instance_prompt", "")
-                    image_path = model_card.data.get("widget", [{}])[0].get("output", {}).get("url", None)
-                    image_url = f"https://huggingface.co/{repo}/resolve/main/{image_path}" if image_path else None
-                except:
-                    trigger_word = ""
-                    image_url = None
-                
-                return repo_name, repo, safetensors_name, trigger_word, image_url
-            except:
-                raise Exception("Invalid safetensors URL format")
-    
-    if link.startswith("https://"):
-        if link.startswith("https://huggingface.co") or link.startswith("https://www.huggingface.co"):
-            link_split = link.split("huggingface.co/")
-            return fetch_hf_adapter_files(link_split[1])
-    else: 
-        return fetch_hf_adapter_files(link)
-
-def incorporate_custom_adapter(custom_lora):
-    global loras
-    if custom_lora:
+        # Fallback: try tokenizing text only and provide images separately if needed
+        logger.warning(f"Processor call with images failed: {e}. Trying text-only processing.")
+        inputs = processor_m(text=[prompt], return_tensors="pt", padding=True).to(device)
+
+    # Move tensors to device if present
+    for k, v in list(inputs.items()):
+        try:
+            inputs[k] = v.to(device)
+        except Exception:
+            pass
+
+    gen_kwargs = {
+        **inputs,
+        "max_new_tokens": max_new_tokens,
+        "temperature": 0.2,
+        # you can add other generation params if needed
+    }
+
+    # Some models expect generate() to be called with different keys — keep try/except.
+    with torch.no_grad():
         try:
-            title, repo, path, trigger_word, image = validate_custom_adapter(custom_lora)
-            print(f"Loaded custom LoRA: {repo}")
-            card = f'''
-            <div class="custom_lora_card">
-              <span>Loaded custom LoRA:</span>
-              <div class="card_internal">
-                <img src="{image}" />
-                <div>
-                    <h3>{title}</h3>
-                    <small>{"Using: <code><b>"+trigger_word+"</code></b> as the trigger word" if trigger_word else "No trigger word found. If there's a trigger word, include it in your prompt"}<br></small>
-                </div>
-              </div>
-            </div>
-            '''
-            existing_item_index = next((index for (index, item) in enumerate(loras) if item['repo'] == repo), None)
-            if existing_item_index is None:
-                new_item = {
-                    "image": image,
-                    "title": title,
-                    "repo": repo,
-                    "weights": path,
-                    "trigger_word": trigger_word
-                }
-                print(new_item)
-                loras.append(new_item)
-                existing_item_index = len(loras) - 1  # Get the actual index after adding
-        
-            return gr.update(visible=True, value=card), gr.update(visible=True), gr.Gallery(selected_index=None), f"Custom: {path}", existing_item_index, trigger_word
+            outputs = model_m.generate(**gen_kwargs)
+            # decode output
+            generated_text = processor_m.decode(outputs[0], skip_special_tokens=True)
         except Exception as e:
-            gr.Warning(f"Invalid LoRA: either you entered an invalid link, or a non-Qwen-Image LoRA, this was the issue: {e}")
-            return gr.update(visible=True, value=f"Invalid LoRA: either you entered an invalid link, a non-Qwen-Image LoRA"), gr.update(visible=True), gr.update(), "", None, ""
+            logger.warning(f"Direct model.generate failed: {e}. Trying streaming decoder approach.")
+            # fallback: run a smaller streaming loop or different interface
+            # Here we attempt to use the model's text generation with .generate from input_ids
+            if "input_ids" in inputs:
+                try:
+                    outputs = model_m.generate(input_ids=inputs["input_ids"].to(device), max_new_tokens=max_new_tokens)
+                    generated_text = processor_m.decode(outputs[0], skip_special_tokens=True)
+                except Exception as e2:
+                    logger.error(f"Fallback generation also failed: {e2}")
+                    raise
+            else:
+                raise
+
+    # Basic cleanup
+    generated_text = generated_text.strip()
+    return generated_text
+
+
+async def call_pdf_parse_async(file_name: str, pdf_url: str, app_key: str, user_id: str) -> Dict[str, Any]:
+    """
+    Replacement for the remote API. Downloads the PDF (signed URL), runs the local VL model to parse,
+    and returns a dict shaped like the remote API response: {'data': {...}}.
+    """
+    start_time = time.time()
+    # 1) download pdf bytes from the signed URL
+    try:
+        logger.info(f"Downloading PDF from url (signed) for local parsing: {pdf_url}")
+        async with httpx.AsyncClient() as client:
+            resp = await client.get(pdf_url, timeout=120)
+            if resp.status_code != 200:
+                raise Exception(f"Failed to download pdf from {pdf_url}, status {resp.status_code}")
+            pdf_bytes = resp.content
+    except Exception as e:
+        logger.error(f"Failed to download pdf bytes: {e}")
+        raise
+
+    # 2) convert pdf bytes to images (pages)
+    try:
+        page_images = await asyncio.to_thread(pdf_bytes_to_images, pdf_bytes, 20)
+    except Exception as e:
+        logger.error(f"PDF to images conversion failed: {e}")
+        raise
+
+    # 3) run model on pages (blocking, put into thread)
+    parse_start = time.time()
+    try:
+        generated_text = await asyncio.to_thread(run_model_on_pages, page_images, file_name, 4096)
+    except Exception as e:
+        logger.error(f"Model generation failed: {e}")
+        raise
+    parse_end = time.time()
+
+    # 4) extract sections from generated_text using markers (best-effort)
+    def extract_section(full_text: str, start_marker: str, end_marker: str) -> str:
+        s = full_text.find(start_marker)
+        e = full_text.find(end_marker, s + len(start_marker)) if s != -1 else -1
+        if s != -1 and e != -1:
+            return full_text[s + len(start_marker):e].strip()
+        return ""
+
+    mmd = extract_section(generated_text, "---BEGIN_MMD---", "---END_MMD---")
+    mmd_html = extract_section(generated_text, "---BEGIN_MMD_HTML---", "---END_MMD_HTML---")
+    qwen_html = extract_section(generated_text, "---BEGIN_QWEN_HTML---", "---END_QWEN_HTML---")
+
+    # If extraction failed, fallback to using the whole generated_text in all fields (graceful)
+    if not any([mmd, mmd_html, qwen_html]):
+        mmd = generated_text
+        mmd_html = f"<pre>{generated_text}</pre>"
+        qwen_html = f"<pre>{generated_text}</pre>"
+
+    end_time = time.time()
+    waiting_cost = 0  # local parse no queue
+    parsing_cost = parse_end - parse_start
+    total_cost = end_time - start_time
+
+    result = {
+        "data": {
+            "mmd": mmd,
+            "qwenHtml": qwen_html,
+            "downloadUrl": pdf_url,
+            "mmdHtml": mmd_html,
+            "waitingCostTime": waiting_cost,
+            "parsingCostTime": parsing_cost,
+            "totalCostTime": total_cost
+        }
+    }
+
+    logger.info(f"Local parsing finished: parsing_time={parsing_cost:.2f}s total_time={total_cost:.2f}s")
+    return result
+
+# ----------------- end of model integration -----------------
+
+# The rest of your code remains functionally identical. I kept it verbatim below.
+
+async def pdf_parse(file_path, request: gr.Request):
+    headers = request.headers
+    print(f'headers: {headers}')
+    user_id = headers.get("X-Modelscope-Router-Id")
+    cookies = request.cookies
+    print(f'cookies: {cookies}')
+    cna = cookies.get('cna')
+    print(f'user_id: {user_id}, cna: {cna}')
+    ip = request.client.host
+    print(f'ip: {ip}')
+    if (user_id is None or user_id == '') and (cna is None or cna == ''):
+        user_id = "visitor"
+    if file_path is None:
+        logger.warning("file_path is None")
+        return (
+            "<p>Please upload a PDF file</p>",
+            "",
+            "<p>No input file</p>",
+            None,
+            None,
+            "Error: No file provided"
+        )
+    logger.info(f'file_path: {file_path}')
+    today = datetime.now().strftime("%Y-%m-%d")
+    file_name = Path(file_path).name
+    safe_file_name = sanitize_filename(file_name, 12)
+    print(f'safe_file_name: {safe_file_name}')
+    oss_path = f"model_scope/pdf_parse/input/{today}/{safe_file_name}"
+    url = upload_file_to_oss(file_path, oss_path)
+    logger.info(f'url: {url}')
+
+    # IMPORTANT: we now call the local model-based parser (no remote API)
+    result = await call_pdf_parse_async(safe_file_name, url, APP_KEY, user_id or cna)
+
+    if result is None:
+        logger.info(f'result is None')
+        return (
+            "<p>The parsing service is not responding. Please try again later.</p>",
+            "",
+            "<p>Service temporarily unavailable</p>",
+            None,
+            None,
+            "Error: The service did not return a response"
+        )
+    data = result.get('data', {})
+    mmd = data.get('mmd')
+    qwen_html = data.get('qwenHtml')
+    download_url = data.get('downloadUrl')
+    logger.info(f'download_url: {download_url}')
+    mmd_html = data.get('mmdHtml')
+    waiting_cost_time = data.get('waitingCostTime')
+    parsing_cost_time = data.get('parsingCostTime')
+    total_cost_time = data.get('totalCostTime')
+    # qwen_html = data.get('qwen_html')
+    # download_url = data.get('download_url')
+    # mmd_html = data.get('mmd_html')
+    cost_time = f'Queue waiting time: {waiting_cost_time}, Parsing time: {parsing_cost_time}, Total time: {total_cost_time}'
+
+    return mmd_html, mmd, qwen_html, download_url, url, cost_time
+
+
+@click.command(context_settings=dict(ignore_unknown_options=True, allow_extra_args=True))
+@click.pass_context
+@click.option(
+    '--latex-delimiters-type',
+    'latex_delimiters_type',
+    type=click.Choice(['a', 'b', 'all']),
+    help="Set the type of LaTeX delimiters to use in Markdown rendering:"
+         "'a' for type '$', 'b' for type '()[]', 'all' for both types.",
+    default='all',
+)
+def main(ctx, latex_delimiters_type, **kwargs):
+    kwargs.update(arg_parse(ctx))
+    if latex_delimiters_type == 'a':
+        latex_delimiters = latex_delimiters_type_a
+    elif latex_delimiters_type == 'b':
+        latex_delimiters = latex_delimiters_type_b
+    elif latex_delimiters_type == 'all':
+        latex_delimiters = latex_delimiters_type_all
     else:
-        return gr.update(visible=False), gr.update(visible=False), gr.update(), "", None, ""
-
-def discard_custom_adapter():
-    return gr.update(visible=False), gr.update(visible=False), gr.update(), "", None, ""
-
-process_adapter_generation.zerogpu = True
-
-css = '''
-#gen_btn{height: 100%}
-#gen_column{align-self: stretch}
-#title{text-align: center}
-#title h1{font-size: 3em; display:inline-flex; align-items:center}
-#title img{width: 100px; margin-right: 0.5em}
-#gallery .grid-wrap{height: 10vh}
-#lora_list{background: var(--block-background-fill);padding: 0 1em .3em; font-size: 90%}
-.card_internal{display: flex;height: 100px;margin-top: .5em}
-.card_internal img{margin-right: 1em}
-.styler{--form-gap-width: 0px !important}
-#speed_status{padding: .5em; border-radius: 5px; margin: 1em 0}
-'''
-
-with gr.Blocks(theme="bethecloud/storj_theme", css=css, delete_cache=(240, 240)) as app:
-    title = gr.HTML("""<h1>Qwen Image LoRA DLC⛵</h1>""", elem_id="title")
-    selected_index = gr.State(None)
-    
-    with gr.Row():
-        with gr.Column(scale=3):
-            prompt = gr.Textbox(label="Prompt", lines=1, placeholder="✦︎ Choose the LoRA and type the prompt")
-        with gr.Column(scale=1, elem_id="gen_column"):
-            generate_button = gr.Button("Generate", variant="primary", elem_id="gen_btn")
-    
-    with gr.Row():
-        with gr.Column():
-            selected_info = gr.Markdown("")
-            gallery = gr.Gallery(
-                [(item["image"], item["title"]) for item in loras],
-                label="LoRA Gallery",
-                allow_preview=False,
-                columns=3,
-                elem_id="gallery",
-                show_share_button=False
-            )
-            with gr.Group():
-                custom_lora = gr.Textbox(label="Custom LoRA", placeholder="username/lora-model-name")
-                gr.Markdown("[Check Qwen-Image LoRAs](https://huggingface.co/models?other=base_model:adapter:Qwen/Qwen-Image)", elem_id="lora_list")
-            custom_lora_info = gr.HTML(visible=False)
-            custom_lora_button = gr.Button("Remove custom LoRA", visible=False)
-        
-        with gr.Column():
-            result = gr.Image(label="Generated Image", format="png")
-
-            with gr.Row():
-                aspect_ratio = gr.Dropdown(
-                    label="Aspect Ratio",
-                    choices=["1:1", "16:9", "9:16", "4:3", "3:4", "3:2", "2:3"],
-                    value="3:2"
-                    )
-            with gr.Row():
-                speed_mode = gr.Dropdown(
-                    label="Output Mode",
-                    choices=["Fast (8 steps)", "Base (50 steps)"],
-                    value="Base (50 steps)",
-                )
-            
-            speed_status = gr.Markdown("Base mode selected - 50 steps for best quality", elem_id="speed_status")
-
-    with gr.Row():
-        with gr.Accordion("Advanced Settings", open=False):
-            with gr.Column():
+        raise ValueError(f"Invalid latex delimiters type: {latex_delimiters_type}.")
+
+    suffixes = pdf_suffixes + image_suffixes
+    with gr.Blocks(head='''
+        <meta name="data-spm" content="label" />
+      <meta name="aplus-core" content="aplus.js" />
+      <meta name="aplus-ifr-pv" content="1"/>
+      <meta name="aplus-iframe-ignore-i" content="on" />
+      <script>
+        window.APLUS_CONFIG = {
+            pid: 'aidata',
+        };
+        (function (w, d, s, q) {
+            w[q] = w[q] || [];
+            var f = d.getElementsByTagName(s)[0],
+                j = d.createElement(s);
+            j.async = true;
+            j.id = 'beacon-aplus';
+            var userIdParam = '';
+            j.setAttribute(
+                'exparams',
+                'userid=' +
+                    userIdParam +
+                    '&aplus&sidx=aplusSidex&ckx=aplusCkx'
+            );
+            j.src = '//g.alicdn.com/alilog/mlog/aplus_v2.js';
+            j.crossorigin = 'anonymous';
+            f.parentNode.insertBefore(j, f);
+        })(window, document, 'script', 'aplus_queue');
+      </script>
+    ''') as demo:
+        gr.HTML(header)
+        with gr.Row():
+            with gr.Column(variant='panel', scale=5):
                 with gr.Row():
-                    cfg_scale = gr.Slider(
-                        label="Guidance Scale (True CFG)", 
-                        minimum=1.0, 
-                        maximum=5.0, 
-                        step=0.1, 
-                        value=4.0,
-                        info="Lower for speed mode, higher for quality"
-                    )
-                    steps = gr.Slider(
-                        label="Steps", 
-                        minimum=4, 
-                        maximum=50, 
-                        step=1, 
-                        value=50,
-                        info="Automatically set by speed mode"
-                    )
-                
+                    input_file = gr.File(label='Please upload a PDF or image (Max 20 pages for conversion)',
+                                         file_types=suffixes)
                 with gr.Row():
-                    randomize_seed = gr.Checkbox(True, label="Randomize seed")
-                    seed = gr.Slider(label="Seed", minimum=0, maximum=MAX_SEED, step=1, value=0, randomize=True)
-                    lora_scale = gr.Slider(label="LoRA Scale", minimum=0, maximum=2, step=0.01, value=1.0)
-
-    # Event handlers
-    gallery.select(
-        handle_lora_selection,
-        inputs=[aspect_ratio],
-        outputs=[prompt, selected_info, selected_index, aspect_ratio]
-    )
-    
-    speed_mode.change(
-        adjust_generation_mode,
-        inputs=[speed_mode],
-        outputs=[speed_status, steps, cfg_scale]
-    )
-    
-    custom_lora.input(
-        incorporate_custom_adapter,
-        inputs=[custom_lora],
-        outputs=[custom_lora_info, custom_lora_button, gallery, selected_info, selected_index, prompt]
-    )
-    
-    custom_lora_button.click(
-        discard_custom_adapter,
-        outputs=[custom_lora_info, custom_lora_button, gallery, selected_info, selected_index, custom_lora]
-    )
-    
-    gr.on(
-        triggers=[generate_button.click, prompt.submit],
-        fn=process_adapter_generation,
-        inputs=[prompt, cfg_scale, steps, selected_index, randomize_seed, seed, aspect_ratio, lora_scale, speed_mode],
-        outputs=[result, seed]
-    )
-
-app.queue()
-app.launch(share=False, ssr_mode=False, show_error=True)
+                    change_bu = gr.Button('Convert')
+                    clear_bu = gr.ClearButton(value='Clear')
+                pdf_show = PDF(label='PDF Preview', interactive=False, visible=True, height=800)
+
+                example_root = os.path.join(os.getcwd(), 'parsing/examples')
+                print(example_root)
+                logger.info(f'example_root: {example_root}')
+                if os.path.exists(example_root):
+                    with gr.Accordion('Examples:'):
+                        gr.Examples(
+                            examples=[os.path.join(example_root, _) for _ in os.listdir(example_root) if
+                                      _.endswith(tuple(suffixes))],
+                            inputs=input_file
+                        )
+
+            with gr.Column(variant='panel', scale=5):
+                output_file = gr.File(label='Conversion Result', interactive=False)
+                cost_time = gr.Text(label='Time Cost')
+                with gr.Tabs():
+                    with gr.Tab('MMD Rendering'):
+                        mmd_html = gr.HTML(label='MMD Rendering')
+                    # with gr.Tab('mmd html text'):
+                    #     mmd_html_text = gr.TextArea(lines=45, show_copy_button=True)
+                    with gr.Tab('MMD'):
+                        mmd = gr.TextArea(lines=45, show_copy_button=True)
+                    with gr.Tab('Qwen HTML'):
+                        raw_html = gr.TextArea(lines=45, show_copy_button=True)
+
+        clear_bu.add([input_file, pdf_show, mmd, raw_html, output_file, mmd_html, cost_time])
+        cna = gr.Textbox(visible=False)
+        input_file.change(fn=to_pdf, inputs=input_file, outputs=pdf_show)
+        change_bu.click(
+            fn=pdf_parse,
+            inputs=[input_file],
+            outputs=[mmd_html, mmd, raw_html, output_file, pdf_show, cost_time],
+            concurrency_limit=15
+        )
+
+    demo.launch()
+
+
+if __name__ == '__main__':
+    main()