app.py

import gradio as gr
from tqdm import tqdm
import requests, os, ctypes, json, argparse, os, array, sys

def ensure_file(filename, src):
    if not os.path.exists(filename):
        response = requests.get(src, stream=True)
        total_size = int(response.headers.get('content-length', 0))

        with open(filename, 'wb') as file:
            with tqdm(total=total_size, unit='B', unit_scale=True, desc=filename, ncols=80) as progress_bar:
                for data in response.iter_content(chunk_size=1024):
                    if data:
                        file.write(data)
                        progress_bar.update(len(data))

        print(f'Download Completed.')

ensure_file("mmproj-model-f16.gguf", "https://huggingface.co/mys/ggml_llava-v1.5-7b/resolve/main/mmproj-model-f16.gguf")
ensure_file("ggml-model-q4_k.gguf", "https://huggingface.co/mys/ggml_llava-v1.5-7b/resolve/main/ggml-model-q4_k.gguf")

from llama_cpp import Llama, clip_model_load, llava_image_embed_make_with_filename, llava_image_embed_make_with_bytes, llava_image_embed_p, llava_image_embed_free, llava_validate_embed_size, llava_eval_image_embed

ctx_clip = clip_model_load("mmproj-model-f16.gguf".encode('utf-8'))
llm = Llama(model_path="ggml-model-q4_k.gguf", n_ctx=2048)

def generate(image, ins='Describe the image'):
    if len(ins) < 1:
        ins = 'Describe the image'
    image_embed = llava_image_embed_make_with_filename(ctx_clip=ctx_clip, n_threads=1, filename=image.encode('utf8'))

    n_past = ctypes.c_int(llm.n_tokens)
    n_past_p = ctypes.byref(n_past)
    llava_eval_image_embed(llm.ctx, image_embed, llm.n_batch, n_past_p)
    llm.n_tokens = n_past.value
    llava_image_embed_free(image_embed)

    llm.eval(llm.tokenize(ins.encode('utf8')))

    max_target_len = 256
    res = ''
    for i in range(max_target_len):
        t_id = llm.sample(temp=0.3)
        t = llm.detokenize([t_id]).decode('utf8')
        if t == '</s>':
            break 
        res += t
        llm.eval([t_id])

    return res 


iface = gr.Interface(generate, inputs=[gr.Image(type='filepath'), gr.Textbox()], outpus='text')
iface.launch()