|
|
|
from PIL import Image |
|
from transformers import AutoTokenizer, AutoModel, AutoImageProcessor, AutoModelForCausalLM |
|
from transformers.generation.configuration_utils import GenerationConfig |
|
import torch |
|
|
|
from emu3.mllm.processing_emu3 import Emu3Processor |
|
|
|
|
|
|
|
EMU_HUB = "BAAI/Emu3-Chat" |
|
VQ_HUB = "BAAI/Emu3-VisionTokenizer" |
|
|
|
|
|
model = AutoModelForCausalLM.from_pretrained( |
|
EMU_HUB, |
|
device_map="cuda:0", |
|
torch_dtype=torch.bfloat16, |
|
attn_implementation="flash_attention_2", |
|
trust_remote_code=True, |
|
) |
|
|
|
tokenizer = AutoTokenizer.from_pretrained(EMU_HUB, trust_remote_code=True) |
|
image_processor = AutoImageProcessor.from_pretrained(VQ_HUB, trust_remote_code=True) |
|
image_tokenizer = AutoModel.from_pretrained(VQ_HUB, device_map="cuda:0", trust_remote_code=True).eval() |
|
processor = Emu3Processor(image_processor, image_tokenizer, tokenizer) |
|
|
|
|
|
text = "Please describe the image" |
|
image = Image.open("assets/demo.png") |
|
|
|
inputs = processor( |
|
text=text, |
|
image=image, |
|
mode='U', |
|
padding_side="left", |
|
padding="longest", |
|
return_tensors="pt", |
|
) |
|
|
|
|
|
GENERATION_CONFIG = GenerationConfig(pad_token_id=tokenizer.pad_token_id, bos_token_id=tokenizer.bos_token_id, eos_token_id=tokenizer.eos_token_id) |
|
|
|
|
|
outputs = model.generate( |
|
inputs.input_ids.to("cuda:0"), |
|
GENERATION_CONFIG, |
|
max_new_tokens=320, |
|
) |
|
|
|
outputs = outputs[:, inputs.input_ids.shape[-1]:] |
|
print(processor.batch_decode(outputs, skip_special_tokens=True)[0]) |
|
|