Image-Text-to-Text
Transformers
Safetensors
Polish
llama
text-generation
vision
multimodal
llava
siglip
bielik
llm
conversational
text-generation-inference
Instructions to use Wojtekb30/Bielik-1.5B-v3.0-VLM-Instruct with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use Wojtekb30/Bielik-1.5B-v3.0-VLM-Instruct with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("image-text-to-text", model="Wojtekb30/Bielik-1.5B-v3.0-VLM-Instruct") messages = [ { "role": "user", "content": [ {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"}, {"type": "text", "text": "What animal is on the candy?"} ] }, ] pipe(text=messages)# Load model directly from transformers import AutoTokenizer, AutoModelForCausalLM tokenizer = AutoTokenizer.from_pretrained("Wojtekb30/Bielik-1.5B-v3.0-VLM-Instruct") model = AutoModelForCausalLM.from_pretrained("Wojtekb30/Bielik-1.5B-v3.0-VLM-Instruct", device_map="auto") messages = [ { "role": "user", "content": [ {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"}, {"type": "text", "text": "What animal is on the candy?"} ] }, ] inputs = tokenizer.apply_chat_template( messages, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt", ).to(model.device) outputs = model.generate(**inputs, max_new_tokens=40) print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:])) - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- vLLM
How to use Wojtekb30/Bielik-1.5B-v3.0-VLM-Instruct with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "Wojtekb30/Bielik-1.5B-v3.0-VLM-Instruct" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "Wojtekb30/Bielik-1.5B-v3.0-VLM-Instruct", "messages": [ { "role": "user", "content": [ { "type": "text", "text": "Describe this image in one sentence." }, { "type": "image_url", "image_url": { "url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg" } } ] } ] }'Use Docker
docker model run hf.co/Wojtekb30/Bielik-1.5B-v3.0-VLM-Instruct
- SGLang
How to use Wojtekb30/Bielik-1.5B-v3.0-VLM-Instruct with SGLang:
Install from pip and serve model
# Install SGLang from pip: pip install sglang # Start the SGLang server: python3 -m sglang.launch_server \ --model-path "Wojtekb30/Bielik-1.5B-v3.0-VLM-Instruct" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "Wojtekb30/Bielik-1.5B-v3.0-VLM-Instruct", "messages": [ { "role": "user", "content": [ { "type": "text", "text": "Describe this image in one sentence." }, { "type": "image_url", "image_url": { "url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg" } } ] } ] }'Use Docker images
docker run --gpus all \ --shm-size 32g \ -p 30000:30000 \ -v ~/.cache/huggingface:/root/.cache/huggingface \ --env "HF_TOKEN=<secret>" \ --ipc=host \ lmsysorg/sglang:latest \ python3 -m sglang.launch_server \ --model-path "Wojtekb30/Bielik-1.5B-v3.0-VLM-Instruct" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "Wojtekb30/Bielik-1.5B-v3.0-VLM-Instruct", "messages": [ { "role": "user", "content": [ { "type": "text", "text": "Describe this image in one sentence." }, { "type": "image_url", "image_url": { "url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg" } } ] } ] }' - Docker Model Runner
How to use Wojtekb30/Bielik-1.5B-v3.0-VLM-Instruct with Docker Model Runner:
docker model run hf.co/Wojtekb30/Bielik-1.5B-v3.0-VLM-Instruct
File size: 8,850 Bytes
fe45cbb 8350b14 fe45cbb 8350b14 fe45cbb | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 | import torch
import os
from PIL import Image
from transformers import AutoModelForCausalLM, AutoTokenizer, AutoModel, AutoImageProcessor
import tkinter as tk
from tkinter import filedialog
from safetensors.torch import load_file
# ============================================================
# KONFIGURACJA 艢CIE呕EK MODELI ORAZ URZ膭DZENIA OBLICZENIOWEGO
# ============================================================
# 艢cie偶ka do modelu wizyjnego
VISION_MODEL_PATH = "google/siglip-so400m-patch14-384"
#VISION_MODEL_PATH = "siglip-so400m-patch14-384"
# 艢cie偶ka do scalonego modelu j臋zykowego (LLM + adaptery)
MERGED_MODEL_PATH = "./"
# 艢cie偶ka do pliku z wagami projektora multimodalnego
PROJECTOR_FILE = "./mm_projector.safetensors"
# Automatyczny wyb贸r urz膮dzenia: GPU (CUDA) je艣li dost臋pne, w przeciwnym razie CPU
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
def wybierz_plik():
"""
Otwiera systemowe okno dialogowe wyboru pliku.
Zwraca 艣cie偶k臋 do wybranego pliku lub pusty string,
je艣li u偶ytkownik anulowa艂 wyb贸r.
"""
root = tk.Tk()
root.withdraw()
sciezka = filedialog.askopenfilename()
root.destroy() # Zamkni臋cie instancji Tkinter w celu zwolnienia zasob贸w
return sciezka if sciezka else ""
# ============================================================
# DEFINICJA PROJEKTORA MULTIMODALNEGO
# ============================================================
class MultimodalProjector(torch.nn.Module):
"""
Projektor odpowiedzialny za mapowanie reprezentacji
wizualnych (vision tower) do przestrzeni osadze艅 (embedding贸w)
modelu j臋zykowego (LLM).
Sk艂ada si臋 z dw贸ch warstw liniowych z aktywacj膮 GELU.
"""
def __init__(self, vision_dim, llm_dim):
"""
:param vision_dim: Rozmiar wektora cech z modelu vision.
:param llm_dim: Rozmiar przestrzeni ukrytej modelu j臋zykowego.
"""
super().__init__()
self.net = torch.nn.Sequential(
torch.nn.Linear(vision_dim, llm_dim),
torch.nn.GELU(),
torch.nn.Linear(llm_dim, llm_dim)
)
def forward(self, x):
"""
Przekszta艂ca wej艣ciowe cechy wizualne do przestrzeni LLM.
"""
return self.net(x)
def load_models():
"""
艁aduje wszystkie wymagane komponenty systemu multimodalnego:
- model j臋zykowy (LLM),
- tokenizer,
- model vision (vision tower),
- procesor obrazu,
- projektor multimodalny.
Zwraca komplet za艂adowanych obiekt贸w.
"""
print(f"艁adowanie Scalonego Bielika z {MERGED_MODEL_PATH}...")
# 1. 艁adowanie modelu j臋zykowego (LLM)
model = AutoModelForCausalLM.from_pretrained(
MERGED_MODEL_PATH,
torch_dtype=torch.float16,
trust_remote_code=True,
device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained(MERGED_MODEL_PATH)
# 2. 艁adowanie modelu vision oraz procesora obrazu
print("艁adowanie Vision Tower...")
vision_tower = AutoModel.from_pretrained(
VISION_MODEL_PATH,
torch_dtype=torch.float16
).to(DEVICE)
# Procesor odpowiada za preprocessing obrazu (resize, normalizacja itd.)
vision_processor = AutoImageProcessor.from_pretrained(VISION_MODEL_PATH)
# 3. 艁adowanie wag projektora multimodalnego
print("艁adowanie Projektora...")
projector_weights = load_file(PROJECTOR_FILE)
# Dynamiczne pobranie wymiar贸w ukrytych z konfiguracji modeli
vision_dim = vision_tower.config.vision_config.hidden_size
llm_dim = model.config.hidden_size
print(f"Wykryte wymiary: Vision={vision_dim}, LLM={llm_dim}")
# Inicjalizacja projektora z odpowiednimi wymiarami
projector = MultimodalProjector(
vision_dim=vision_dim,
llm_dim=llm_dim
).to(DEVICE).to(torch.float16)
# Za艂adowanie wag do projektora
projector.load_state_dict(projector_weights)
return model, vision_tower, vision_processor, projector, tokenizer
def chat():
"""
G艂贸wna p臋tla interakcyjna aplikacji.
Umo偶liwia prac臋 w trybie:
- tekstowym,
- multimodalnym (tekst + obraz).
"""
# Za艂adowanie wszystkich komponent贸w systemu
model, vision_tower, vision_processor, projector, tokenizer = load_models()
print("\n" + "=" * 50)
print("BIELIK LLaVA - GOTOWY DO ROZMOWY")
print("Otworzy si臋 okno wyboru pliku. Anulowanie wyboru = tryb tekstowy.")
print("Wpisz 'exit' w konsoli aby wyj艣膰.")
print("=" * 50 + "\n")
while True:
# ============================================
# A. Wyb贸r obrazu (opcjonalnie)
# ============================================
print("\nWybierz plik obrazka w oknie...")
img_path = wybierz_plik()
pixel_values = None
has_image = False
if img_path:
if os.path.exists(img_path):
try:
print(f"Wybrano: {img_path}")
# Wczytanie i konwersja obrazu do RGB
image = Image.open(img_path).convert('RGB')
# Przetwarzanie obrazu do tensora zgodnego z vision tower
pixel_values = vision_processor(
images=image,
return_tensors="pt"
).pixel_values.to(DEVICE, dtype=torch.float16)
has_image = True
except Exception as e:
print(f"B艂膮d 艂adowania obrazka: {e}")
continue
else:
print("艢cie偶ka nieprawid艂owa.")
else:
print("Tryb tekstowy (bez obrazka).")
# ============================================
# B. Pobranie promptu tekstowego
# ============================================
prompt = input("Tw贸j Prompt: ").strip()
if prompt.lower() == 'exit':
break
if not prompt:
continue
# ============================================
# C. Generowanie odpowiedzi
# ============================================
with torch.no_grad():
# 1. Ekstrakcja cech wizualnych (je艣li obraz zosta艂 podany)
img_embeds = None
if has_image:
vision_feats = vision_tower.vision_model(pixel_values).last_hidden_state
img_embeds = projector(vision_feats)
# 2. Przygotowanie wej艣cia tekstowego w odpowiednim formacie czatu
if has_image:
text_input = (
f"<|im_start|>user\n<image>\n{prompt}<|im_end|>\n"
f"<|im_start|>assistant\n"
)
else:
text_input = (
f"<|im_start|>user\n{prompt}<|im_end|>\n"
f"<|im_start|>assistant\n"
)
# Tokenizacja tekstu
input_ids = tokenizer.encode(
text_input,
return_tensors="pt"
).to(DEVICE)
# Konwersja token贸w do embedding贸w
inputs_embeds = model.model.embed_tokens(input_ids)
# 3. 艁膮czenie embedding贸w obrazu i tekstu
final_embeds = inputs_embeds
if has_image:
input_ids_clean = tokenizer.encode(
text_input,
return_tensors="pt"
).to(DEVICE)
inputs_embeds_clean = model.model.embed_tokens(input_ids_clean)
# Konkatenacja embedding贸w obrazu oraz tekstu w osi sekwencji
final_embeds = torch.cat(
[img_embeds, inputs_embeds_clean],
dim=1
)
# 4. Generowanie odpowiedzi przez model j臋zykowy
print("Generowanie...", end="", flush=True)
# Utworzenie attention_mask o tej samej d艂ugo艣ci co final_embeds
batch_size = final_embeds.shape[0]
seq_len = final_embeds.shape[1]
attention_mask = torch.ones(
(batch_size, seq_len),
device=DEVICE
)
output_ids = model.generate(
inputs_embeds=final_embeds,
attention_mask=attention_mask,
max_new_tokens=256,
temperature=0.3,
do_sample=True,
pad_token_id=tokenizer.eos_token_id,
eos_token_id=tokenizer.eos_token_id
)
# Dekodowanie wygenerowanej sekwencji token贸w
generated_text = tokenizer.decode(
output_ids[0],
skip_special_tokens=True
)
print("\r", end="")
print(f"Bielik: {generated_text}")
print("-" * 30)
if __name__ == "__main__":
chat() |