app.py

import os
import random
import uuid
import json
import time
import asyncio
from threading import Thread
from typing import Iterable

import gradio as gr
import spaces
import torch
import numpy as np
from PIL import Image
# cv2 is no longer needed as video processing is removed

from transformers import (
    Qwen2VLForConditionalGeneration,
    Qwen2_5_VLForConditionalGeneration,
    AutoModelForImageTextToText,
    AutoProcessor,
    TextIteratorStreamer,
)
from transformers.image_utils import load_image
from gradio.themes import Soft
from gradio.themes.utils import colors, fonts, sizes

# --- Theme and CSS Definition ---

# Define the new SpringGreen color palette
colors.spring_green = colors.Color(
    name="spring_green",
    c50="#E5FFF2",
    c100="#CCFFEC",
    c200="#99FFD9",
    c300="#66FFC6",
    c400="#33FFB3",
    c500="#00FF7F",  # SpringGreen base color
    c600="#00E672",
    c700="#00CC66",
    c800="#00B359",
    c900="#00994D",
    c950="#008040",
)


class SpringGreenTheme(Soft):
    def __init__(
        self,
        *,
        primary_hue: colors.Color | str = colors.gray,
        secondary_hue: colors.Color | str = colors.spring_green, # Use the new color
        neutral_hue: colors.Color | str = colors.slate,
        text_size: sizes.Size | str = sizes.text_lg,
        font: fonts.Font | str | Iterable[fonts.Font | str] = (
            fonts.GoogleFont("Outfit"), "Arial", "sans-serif",
        ),
        font_mono: fonts.Font | str | Iterable[fonts.Font | str] = (
            fonts.GoogleFont("IBM Plex Mono"), "ui-monospace", "monospace",
        ),
    ):
        super().__init__(
            primary_hue=primary_hue,
            secondary_hue=secondary_hue,
            neutral_hue=neutral_hue,
            text_size=text_size,
            font=font,
            font_mono=font_mono,
        )
        super().set(
            background_fill_primary="*primary_50",
            background_fill_primary_dark="*primary_900",
            body_background_fill="linear-gradient(135deg, *primary_200, *primary_100)",
            body_background_fill_dark="linear-gradient(135deg, *primary_900, *primary_800)",
            button_primary_text_color="black",
            button_primary_text_color_hover="white",
            button_primary_background_fill="linear-gradient(90deg, *secondary_400, *secondary_500)",
            button_primary_background_fill_hover="linear-gradient(90deg, *secondary_500, *secondary_600)",
            button_primary_background_fill_dark="linear-gradient(90deg, *secondary_600, *secondary_700)",
            button_primary_background_fill_hover_dark="linear-gradient(90deg, *secondary_500, *secondary_600)",
            slider_color="*secondary_400",
            slider_color_dark="*secondary_600",
            block_title_text_weight="600",
            block_border_width="3px",
            block_shadow="*shadow_drop_lg",
            button_primary_shadow="*shadow_drop_lg",
            button_large_padding="11px",
            color_accent_soft="*primary_100",
            block_label_background_fill="*primary_200",
        )

# Instantiate the new theme
spring_green_theme = SpringGreenTheme()


css = """
#main-title h1 {
    font-size: 2.3em !important;
}
#output-title h2 {
    font-size: 2.1em !important;
}
"""

# Constants for text generation
MAX_MAX_NEW_TOKENS = 2048
DEFAULT_MAX_NEW_TOKENS = 1024
# Increased max_length to accommodate more complex inputs
MAX_INPUT_TOKEN_LENGTH = int(os.getenv("MAX_INPUT_TOKEN_LENGTH", "8192"))

device = torch.device("cuda:0" 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)

# --- Model Loading ---
# Load Nanonets-OCR2-3B
MODEL_ID_V = "nanonets/Nanonets-OCR2-3B"
processor_v = AutoProcessor.from_pretrained(MODEL_ID_V, trust_remote_code=True)
model_v = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    MODEL_ID_V,
    trust_remote_code=True,
    torch_dtype=torch.float16
).to(device).eval()

# Load Qwen2-VL-OCR-2B-Instruct
MODEL_ID_X = "prithivMLmods/Qwen2-VL-OCR-2B-Instruct"
processor_x = AutoProcessor.from_pretrained(MODEL_ID_X, trust_remote_code=True)
model_x = Qwen2VLForConditionalGeneration.from_pretrained(
    MODEL_ID_X,
    trust_remote_code=True,
    torch_dtype=torch.float16
).to(device).eval()

# Load Aya-Vision-8b
MODEL_ID_A = "CohereForAI/aya-vision-8b"
processor_a = AutoProcessor.from_pretrained(MODEL_ID_A, trust_remote_code=True)
model_a = AutoModelForImageTextToText.from_pretrained(
    MODEL_ID_A,
    trust_remote_code=True,
    torch_dtype=torch.float16
).to(device).eval()

# Load olmOCR-7B-0725
MODEL_ID_W = "allenai/olmOCR-7B-0725"
processor_w = AutoProcessor.from_pretrained(MODEL_ID_W, trust_remote_code=True)
model_w = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    MODEL_ID_W,
    trust_remote_code=True,
    torch_dtype=torch.float16
).to(device).eval()

# Load RolmOCR
MODEL_ID_M = "reducto/RolmOCR"
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
).to(device).eval()


@spaces.GPU
def generate_image(model_name: str, text: str, image: Image.Image,
                   max_new_tokens: int = 1024,
                   temperature: float = 0.6,
                   top_p: float = 0.9,
                   top_k: int = 50,
                   repetition_penalty: float = 1.2):
    """
    Generates responses using the selected model for image input.
    Yields raw text and Markdown-formatted text.
    """
    if model_name == "RolmOCR-7B":
        processor = processor_m
        model = model_m
    elif model_name == "Qwen2-VL-OCR-2B":
        processor = processor_x
        model = model_x
    elif model_name == "Nanonets-OCR2-3B":
        processor = processor_v
        model = model_v
    elif model_name == "Aya-Vision-8B":
        processor = processor_a
        model = model_a
    elif model_name == "olmOCR-7B-0725":
        processor = processor_w
        model = model_w
    else:
        yield "Invalid model selected.", "Invalid model selected."
        return

    if image is None:
        yield "Please upload an image.", "Please upload an image."
        return

    messages = [{
        "role": "user",
        "content": [
            {"type": "image"},
            {"type": "text", "text": text},
        ]
    }]
    prompt_full = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

    # FIX: Set truncation to False and rely on the model's context length.
    # The increased MAX_INPUT_TOKEN_LENGTH at the top also helps.
    inputs = processor(
        text=[prompt_full],
        images=[image],
        return_tensors="pt",
        padding=True
    ).to(device)

    streamer = TextIteratorStreamer(processor, skip_prompt=True, skip_special_tokens=True)
    generation_kwargs = {**inputs, "streamer": streamer, "max_new_tokens": max_new_tokens}
    thread = Thread(target=model.generate, kwargs=generation_kwargs)
    thread.start()
    buffer = ""
    for new_text in streamer:
        buffer += new_text
        buffer = buffer.replace("<|im_end|>", "")
        time.sleep(0.01)
        yield buffer, buffer


# Define examples for image inference
image_examples = [
    ["Extract the full page.", "images/ocr.png"],
    ["Extract the content.", "images/4.png"],
    ["Convert this page to doc [table] precisely for markdown.", "images/0.png"]
]


# Create the Gradio Interface
with gr.Blocks(css=css, theme=spring_green_theme) as demo:
    gr.Markdown("# **Multimodal OCR**", elem_id="main-title")
    with gr.Row():
        with gr.Column(scale=2):
            image_query = gr.Textbox(label="Query Input", placeholder="Enter your query here...")
            image_upload = gr.Image(type="pil", label="Upload Image", height=290)
            image_submit = gr.Button("Submit", variant="primary")
            gr.Examples(
                examples=image_examples,
                inputs=[image_query, image_upload]
            )
            with gr.Accordion("Advanced options", open=False):
                max_new_tokens = gr.Slider(label="Max new tokens", minimum=1, maximum=MAX_MAX_NEW_TOKENS, step=1,
                                           value=DEFAULT_MAX_NEW_TOKENS)
                temperature = gr.Slider(label="Temperature", minimum=0.1, maximum=4.0, step=0.1, value=0.6)
                top_p = gr.Slider(label="Top-p (nucleus sampling)", minimum=0.05, maximum=1.0, step=0.05, value=0.9)
                top_k = gr.Slider(label="Top-k", minimum=1, maximum=1000, step=1, value=50)
                repetition_penalty = gr.Slider(label="Repetition penalty", minimum=1.0, maximum=2.0, step=0.05,
                                               value=1.2)

        with gr.Column(scale=3):
            gr.Markdown("## Output", elem_id="output-title")
            output = gr.Textbox(label="Raw Output Stream", interactive=False, lines=11, show_copy_button=True)
            with gr.Accordion("(Result.md)", open=False):
                markdown_output = gr.Markdown(label="(Result.Md)")

            model_choice = gr.Radio(
                choices=["Nanonets-OCR2-3B", "olmOCR-7B-0725", "RolmOCR-7B",
                         "Aya-Vision-8B", "Qwen2-VL-OCR-2B"],
                label="Select Model",
                value="Nanonets-OCR2-3B"
            )

    image_submit.click(
        fn=generate_image,
        inputs=[model_choice, image_query, image_upload, max_new_tokens, temperature, top_p, top_k,
                repetition_penalty],
        outputs=[output, markdown_output]
    )

if __name__ == "__main__":
    demo.queue(max_size=50).launch(mcp_server=True, ssr_mode=False, show_error=True)