cli.py

import torch
from constants import IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN
from conversation import conv_templates, SeparatorStyle
from builder import load_pretrained_model
from utils import disable_torch_init
from mm_utils import process_images, tokenizer_image_token, get_model_name_from_path
from PIL import Image
import requests
from io import BytesIO
from transformers import TextStreamer
import spaces
from functools import partial
import traceback
import sys
# def load_image(image_file):
#     if image_file.startswith('http://') or image_file.startswith('https://'):
#         response = requests.get(image_file)
#         image = Image.open(BytesIO(response.content)).convert('RGB')
#     else:
#         image = Image.open(image_file).convert('RGB')
#     return image

def load_image(image_file):
    print("the image file : ", image_file)
    
    image = Image.open(image_file).convert('RGB')

    if image is None:
        print("image is None")
        sys.exit("Aborting program: Image is None.")
    
    return image

@spaces.GPU()
def run_inference(
    model_path,
    image_file,
    prompt_text,
    model_base=None,
    device="cuda",
    conv_mode=None,
    temperature=0.2,
    max_new_tokens=512,
    load_8bit=False,
    load_4bit=False,
    debug=False
):
    # Model initialization
    disable_torch_init()

    model_name = get_model_name_from_path(model_path)
    tokenizer, model, image_processor, context_len = load_pretrained_model(
        model_path, model_base, model_name, load_8bit, load_4bit
    )

    # Determine conversation mode
    if "llama-2" in model_name.lower():
        conv_mode_inferred = "llava_llama_2"
    elif "mistral" in model_name.lower():
        conv_mode_inferred = "mistral_instruct"
    elif "v1.6-34b" in model_name.lower():
        conv_mode_inferred = "chatml_direct"
    elif "v1" in model_name.lower():
        conv_mode_inferred = "llava_v1"
    elif "mpt" in model_name.lower():
        conv_mode_inferred = "mpt"
    elif "gemma" in model_name.lower():
        conv_mode_inferred = "ferret_gemma_instruct"
    elif "llama" in model_name.lower():
        conv_mode_inferred = "ferret_llama_3"
    else:
        conv_mode_inferred = "llava_v0"

    # Use user-specified conversation mode if provided
    conv_mode = conv_mode or conv_mode_inferred

    if conv_mode != conv_mode_inferred:
        print(f'[WARNING] the auto inferred conversation mode is {conv_mode_inferred}, while `conv_mode` is {conv_mode}, using {conv_mode}')

    conv = conv_templates[conv_mode].copy()

    if "mpt" in model_name.lower():
        roles = ('user', 'assistant')
    else:
        roles = conv.roles

    # Load and process image
    print("loading image", image_file)
    image = load_image(image_file)
    if image is None:
        print("image is None")
    image_size = image.size
    image_h = 336  # Height of the image
    image_w = 336
    #ERROR
    # image_tensor = process_images([image], image_processor, model.config)
    # if type(image_tensor) is list:
    #     image_tensor = [image.to(model.device, dtype=torch.float16) for image in image_tensor]
    # else:
    #     image_tensor = image_tensor.to(model.device, dtype=torch.float16)
    if model.config.image_aspect_ratio == "square_nocrop":
            image_tensor = image_processor.preprocess(image, return_tensors='pt', do_resize=True, 
                                                  do_center_crop=False, size=[image_h, image_w])['pixel_values'][0]
    elif model.config.image_aspect_ratio == "anyres":
        image_process_func = partial(image_processor.preprocess, return_tensors='pt', do_resize=True, do_center_crop=False, size=[image_h, image_w])
        image_tensor = process_images([image], image_processor, model.config, image_process_func=image_process_func)[0]
    else:
        image_tensor = process_images([image], image_processor, model.config)[0]

    if model.dtype == torch.float16:
        image_tensor = image_tensor.half()  # Convert image tensor to float16
        data_type = torch.float16
    else:
        image_tensor = image_tensor.float()  # Keep it in float32
        data_type = torch.float32

    # Now, add the batch dimension and move to GPU
    images = image_tensor.unsqueeze(0).to(data_type).cuda()


    # Process the first message with the image
    if model.config.mm_use_im_start_end:
        prompt_text = DEFAULT_IM_START_TOKEN + DEFAULT_IMAGE_TOKEN + DEFAULT_IM_END_TOKEN + '\n' + prompt_text
    else:
        prompt_text = DEFAULT_IMAGE_TOKEN + '\n' + prompt_text

    # Prepare conversation
    conv.append_message(conv.roles[0], prompt_text)
    conv.append_message(conv.roles[1], None)
    prompt = conv.get_prompt()

    input_ids = tokenizer_image_token(prompt, tokenizer, IMAGE_TOKEN_INDEX, return_tensors='pt').unsqueeze(0).to(model.device)
    stop_str = conv.sep if conv.sep_style != SeparatorStyle.TWO else conv.sep2
    keywords = [stop_str]
    streamer = TextStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)
    print("image size: ", image_size)
    # Generate the model's response
    
    with torch.inference_mode():
        output_ids = model.generate(
            input_ids,
            images=images,
            image_sizes=[image_size],
            do_sample=True if temperature > 0 else False,
            temperature=temperature,
            max_new_tokens=max_new_tokens,
            streamer=streamer,
            num_beams=1,
            use_cache=True
        )

    # Decode and return the output
    outputs = tokenizer.decode(output_ids[0]).strip()
    conv.messages[-1][-1] = outputs

    if debug:
        print("\n", {"prompt": prompt, "outputs": outputs}, "\n")

    return outputs


# Example usage:
# response = run_inference("path_to_model", "path_to_image", "your_prompt")
# print(response)