nvlm.py

import torch
from transformers import AutoTokenizer, AutoModel
import math
from PIL import Image
import torchvision.transforms as T
from torchvision.transforms.functional import InterpolationMode

from .base import BaseModel
from ..smp import *
from ..dataset import DATASET_TYPE


IMAGENET_MEAN = (0.485, 0.456, 0.406)
IMAGENET_STD = (0.229, 0.224, 0.225)


def split_model():
    device_map = {}

    num_gpus = torch.cuda.device_count()
    rank, world_size = get_rank_and_world_size()
    num_gpus = num_gpus // world_size
    num_layers = 80
    # Since the first GPU will be used for ViT, treat it as half a GPU.
    num_layers_per_gpu = math.ceil(num_layers / (num_gpus - 0.5))
    num_layers_per_gpu = [num_layers_per_gpu] * num_gpus
    num_layers_per_gpu[0] = math.ceil(num_layers_per_gpu[0] * 0.5)
    layer_cnt = 0

    for i, num_layer in enumerate(num_layers_per_gpu):
        for j in range(num_layer):
            device_map[f'language_model.model.layers.{layer_cnt}'] = rank + i * world_size
            layer_cnt += 1

    device_map['vision_model'] = rank
    device_map['mlp1'] = rank
    device_map['language_model.model.embed_tokens'] = rank
    device_map['language_model.model.norm'] = rank
    device_map['language_model.model.rotary_emb'] = rank
    device_map['language_model.lm_head'] = rank
    device_map[f'language_model.model.layers.{num_layers - 1}'] = rank
    return device_map


def build_transform(input_size):
    MEAN, STD = IMAGENET_MEAN, IMAGENET_STD
    transform = T.Compose([
        T.Lambda(lambda img: img.convert('RGB') if img.mode != 'RGB' else img),
        T.Resize((input_size, input_size), interpolation=InterpolationMode.BICUBIC),
        T.ToTensor(),
        T.Normalize(mean=MEAN, std=STD)
    ])
    return transform


def find_closest_aspect_ratio(aspect_ratio, target_ratios, width, height, image_size):
    best_ratio_diff = float('inf')
    best_ratio = (1, 1)
    area = width * height
    for ratio in target_ratios:
        target_aspect_ratio = ratio[0] / ratio[1]
        ratio_diff = abs(aspect_ratio - target_aspect_ratio)
        if ratio_diff < best_ratio_diff:
            best_ratio_diff = ratio_diff
            best_ratio = ratio
        elif ratio_diff == best_ratio_diff:
            if area > 0.5 * image_size * image_size * ratio[0] * ratio[1]:
                best_ratio = ratio
    return best_ratio


def dynamic_preprocess(image, min_num=1, max_num=12, image_size=448, use_thumbnail=False):
    orig_width, orig_height = image.size
    aspect_ratio = orig_width / orig_height

    # calculate the existing image aspect ratio
    target_ratios = set(
        (i, j) for n in range(min_num, max_num + 1) for i in range(1, n + 1) for j in range(1, n + 1) if
        i * j <= max_num and i * j >= min_num)
    target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1])

    # find the closest aspect ratio to the target
    target_aspect_ratio = find_closest_aspect_ratio(
        aspect_ratio, target_ratios, orig_width, orig_height, image_size)

    # calculate the target width and height
    target_width = image_size * target_aspect_ratio[0]
    target_height = image_size * target_aspect_ratio[1]
    blocks = target_aspect_ratio[0] * target_aspect_ratio[1]

    # resize the image
    resized_img = image.resize((target_width, target_height))
    processed_images = []
    for i in range(blocks):
        box = (
            (i % (target_width // image_size)) * image_size,
            (i // (target_width // image_size)) * image_size,
            ((i % (target_width // image_size)) + 1) * image_size,
            ((i // (target_width // image_size)) + 1) * image_size
        )
        # split the image
        split_img = resized_img.crop(box)
        processed_images.append(split_img)
    assert len(processed_images) == blocks
    if use_thumbnail and len(processed_images) != 1:
        thumbnail_img = image.resize((image_size, image_size))
        processed_images.append(thumbnail_img)
    return processed_images


def load_image(image_file, input_size=448, max_num=12):
    image = Image.open(image_file).convert('RGB')
    transform = build_transform(input_size=input_size)
    images = dynamic_preprocess(image, image_size=input_size, use_thumbnail=True, max_num=max_num)
    pixel_values = [transform(image) for image in images]
    pixel_values = torch.stack(pixel_values)
    return pixel_values


class NVLM(BaseModel):

    INSTALL_REQ = False
    INTERLEAVE = False

    def __init__(self, model_path='nvidia/NVLM-D-72B', **kwargs):
        assert model_path is not None
        self.model_path = model_path
        self.tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True, use_fast=False)
        kwargs_default = dict(max_new_tokens=1024, do_sample=False)
        kwargs_default.update(kwargs)
        self.kwargs = kwargs_default

        self.model = AutoModel.from_pretrained(
            model_path,
            torch_dtype=torch.bfloat16,
            low_cpu_mem_usage=True,
            use_flash_attn=False,
            trust_remote_code=True,
            device_map=split_model()).eval()

        logging.info(f'Following kwargs received: {self.kwargs}, will use as generation config. ')
        torch.cuda.empty_cache()

    def generate_inner(self, message, dataset=None):
        prompt, image_path = self.message_to_promptimg(message, dataset=dataset)
        pixel_values = load_image(image_path, max_num=6).to(torch.bfloat16).cuda()
        response = self.model.chat(self.tokenizer, pixel_values, prompt, self.kwargs)
        return response.strip()