README.md

metadata

license: apache-2.0

Generation

The following is the sample code for inference.

from llava.model.builder import load_pretrained_model
from llava.mm_utils import process_images, tokenizer_image_token
from llava.constants import DEFAULT_IMAGE_TOKEN

from PIL import Image
import torch
import time
import warnings
import json 

# export PYTHONPATH="/thestack/LLM4CodeBeta/LLaVA-NeXT-FLAME:$PYTHONPATH"

warnings.filterwarnings("ignore")

pretrained = "/root/nfs3/flame_ft/res/checkpoints/flame-google_siglip-so400m-patch14-384-deepseek-ai_deepseek-coder-6.7b-instruct-mlp2x_gelu-selectlayer-2-onevision-1-pretrain_mmcoder-3NODE-Date1212-STAGE2v9-2-data_1220_no_code_v1-inst_data-STAGE2v9-eos-16k-1220-FINETUNE-2-data_1220/no_code_v1-inst_data-v5_v6-eos-16k-1223"

model_name = "flame"
device = "cuda"
device_map = "auto"
llava_model_args = {
    "multimodal": True,
    "attn_implementation": None,
}
tokenizer, model, image_processor, max_length = load_pretrained_model(pretrained, None, model_name, device_map=device_map,**llava_model_args)
model.config.tokenizer_padding_side = 'left'  # Use left padding for batch processing
# model.config.image_aspect_ratio = "resize"
model.eval()

url = "/root/nfs2/flame_ft/datasets/data_1220/TESTING_DATA/TEST80/imgs/000000034/000000034.png"
image = Image.open(url) 
image_tensor = process_images([image], image_processor, model.config)
image_tensor = [_image.to(dtype=torch.float16, device=device) for _image in image_tensor]

prompt = "Below is an image of the page to create. Generate React code and styles to replicate the design, including layout, typography, and styling. Format your response as follows:'// CSS\n[CSS/SCSS code]\n\n// [React Implementation (JS/TS/JSX/TSX)]\n[Component code]'.\n\n ### Input Image:\n{image}\n\n### Response:\n"

input_ids = tokenizer_image_token(prompt, tokenizer, return_tensors='pt')
input_ids = input_ids.unsqueeze(0)
input_ids=input_ids.to(device)
image_sizes = [image.size]
modalities = ["image"] 

cont = model.generate(
    input_ids,
    images=image_tensor,
    image_sizes=image_sizes,
    modalities=modalities,                   # Added this line with the modalities
    do_sample=True,
    num_beams=5, 
    temperature=0.1,
    max_new_tokens=4096,
    top_p=0.95,
    repetition_penalty=1.05
)

text_outputs = tokenizer.batch_decode(cont, skip_special_tokens=True)