Model Card for Model ID
Русскоязычная версия Idefics, обученная на русифицированном сабсете LLaVA.
SFT был без текстовых данных, так что вполне возможно просадка по качеству на text-only данных.
Обучение было в int4 с QLoRA на consumer-grade железе.
Model Details
Model Description
- Model type: ruIdefics2
- Language(s) (NLP): Russian
- License: Apache-2.0
- Finetuned from model: Idefics2
How to Get Started
Запуск в fp16
import requests
import torch
from PIL import Image
from io import BytesIO
from transformers import AutoProcessor, AutoModelForVision2Seq
from transformers.image_utils import load_image
DEVICE = "cuda:0"
image1 = load_image("https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg")
image2 = load_image("https://cdn.britannica.com/59/94459-050-DBA42467/Skyline-Chicago.jpg")
image3 = load_image("https://cdn.britannica.com/68/170868-050-8DDE8263/Golden-Gate-Bridge-San-Francisco.jpg")
processor = AutoProcessor.from_pretrained("GeorgeBredis/ruIdefics2-ruLLaVA-merged")
model = AutoModelForVision2Seq.from_pretrained(
"GeorgeBredis/ruIdefics2-ruLLaVA-merged",
).to(DEVICE)
messages = [
{
"role": "user",
"content": [
{"type": "image"},
{"type": "text", "text": "Что изображено на данной картинке?"},
]
}
]
prompt = processor.apply_chat_template(messages, add_generation_prompt=True)
inputs = processor(text=prompt, images=[image1], return_tensors="pt")
inputs = {k: v.to(DEVICE) for k, v in inputs.items()}
generated_ids = model.generate(**inputs, max_new_tokens=500)
generated_texts = processor.batch_decode(generated_ids, skip_special_tokens=True)
print(generated_texts)
Вполне возможно что это не влезет в вашу GPU (если будете загружать на gpu), так что ниже вариант с bnb для запуска в colab'e.
Запуск в int4/int8 c bnb.
Требует установки peft
import requests
import torch
from PIL import Image
from io import BytesIO
from peft import LoraConfig
from transformers import AutoProcessor, BitsAndBytesConfig, Idefics2ForConditionalGeneration
from transformers.image_utils import load_image
DEVICE = "cuda:0"
image1 = load_image("https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg")
image2 = load_image("https://cdn.britannica.com/59/94459-050-DBA42467/Skyline-Chicago.jpg")
image3 = load_image("https://cdn.britannica.com/68/170868-050-8DDE8263/Golden-Gate-Bridge-San-Francisco.jpg")
processor = AutoProcessor.from_pretrained(
"GeorgeBredis/ruIdefics2-ruLLaVA-merged",
do_image_splitting=False
)
quantization_config = BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_quant_type="nf4",
bnb_4bit_use_double_quant=True,
bnb_4bit_compute_dtype=torch.float16
)
model = Idefics2ForConditionalGeneration.from_pretrained(
"GeorgeBredis/ruIdefics2-ruLLaVA-merged",
torch_dtype=torch.float16,
quantization_config=quantization_config,
)
# не нужно переносить на карту, так как в int4/8 заводятся сразу на них
messages = [
{
"role": "user",
"content": [
{"type": "image"},
{"type": "text", "text": "Что изображено на данной картинке?"},
]
}
]
prompt = processor.apply_chat_template(messages, add_generation_prompt=True)
inputs = processor(text=prompt, images=[image1], return_tensors="pt")
inputs = {k: v.to(DEVICE) for k, v in inputs.items()}
generated_ids = model.generate(**inputs, max_new_tokens=500)
generated_texts = processor.batch_decode(generated_ids, skip_special_tokens=True)
print(generated_texts)
- Downloads last month
- 21
Inference API (serverless) does not yet support transformers models for this pipeline type.