Spaces:
Running
on
Zero
Running
on
Zero
File size: 10,011 Bytes
8e31ab1 2144e66 8e31ab1 8ec9ef4 8e31ab1 8ec9ef4 8e31ab1 8ec9ef4 066eb01 8ec9ef4 2144e66 8e31ab1 8ec9ef4 8e31ab1 8ec9ef4 066eb01 2144e66 8ec9ef4 8e31ab1 066eb01 2144e66 8ec9ef4 066eb01 8ec9ef4 2144e66 8ec9ef4 2144e66 066eb01 2144e66 066eb01 8ec9ef4 8e31ab1 8ec9ef4 8e31ab1 066eb01 8e31ab1 2144e66 8e31ab1 2144e66 8e31ab1 2144e66 8e31ab1 2144e66 8ec9ef4 2144e66 8e31ab1 2144e66 8e31ab1 2144e66 |
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 |
import gradio as gr
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, AutoProcessor, AutoModel
from PIL import Image
import logging
import spaces
# Setup logging
logging.basicConfig(level=logging.INFO)
class LLaVAPhiModel:
def __init__(self, model_id="sagar007/Lava_phi"):
self.device = "cuda" # Always use cuda with ZeroGPU
self.model_id = model_id
logging.info("Initializing LLaVA-Phi model...")
# Initialize tokenizer (can be done outside GPU context)
self.tokenizer = AutoTokenizer.from_pretrained(model_id)
if self.tokenizer.pad_token is None:
self.tokenizer.pad_token = self.tokenizer.eos_token
try:
# Initialize processor (can be done outside GPU context)
self.processor = AutoProcessor.from_pretrained("huggingface/clip-vit-base-patch32")
except Exception as e:
logging.warning(f"Failed to load CLIP processor: {str(e)}")
# Fallback to basic tokenizer if needed
self.processor = None
# Store conversation history
self.history = []
# Lazy loading of models - will be initialized in GPU context
self.model = None
self.clip = None
@spaces.GPU
def ensure_models_loaded(self):
"""Ensure models are loaded in GPU context"""
if self.model is None:
# Load main model
from transformers import BitsAndBytesConfig
quantization_config = BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_compute_dtype=torch.float16,
bnb_4bit_use_double_quant=True,
bnb_4bit_quant_type="nf4"
)
self.model = AutoModelForCausalLM.from_pretrained(
self.model_id,
quantization_config=quantization_config,
device_map="auto",
torch_dtype=torch.bfloat16,
trust_remote_code=True
)
self.model.config.pad_token_id = self.tokenizer.eos_token_id
if self.clip is None:
# Load CLIP model if not already loaded
if self.clip is None:
try:
self.clip = AutoModel.from_pretrained("huggingface/clip-vit-base-patch32").to(self.device)
except Exception as e:
logging.warning(f"Failed to load CLIP model: {str(e)}")
self.clip = None
@spaces.GPU
def process_image(self, image):
"""Process image through CLIP if available, otherwise return None"""
try:
# Ensure models are loaded
self.ensure_models_loaded()
# If CLIP isn't available, return None
if self.clip is None or self.processor is None:
logging.warning("CLIP model or processor not available - skipping image processing")
return None
# Convert image to correct format
if isinstance(image, str):
image = Image.open(image)
elif isinstance(image, numpy.ndarray):
image = Image.fromarray(image)
with torch.no_grad():
try:
image_inputs = self.processor(images=image, return_tensors="pt")
image_features = self.clip.get_image_features(
pixel_values=image_inputs.pixel_values.to(self.device)
)
return image_features
except Exception as e:
logging.error(f"Error during image processing: {str(e)}")
return None
except Exception as e:
logging.error(f"Error in process_image: {str(e)}")
return None
@spaces.GPU(duration=120) # Set longer duration for generation
def generate_response(self, message, image=None):
try:
# Ensure models are loaded
self.ensure_models_loaded()
if image is not None:
image_features = self.process_image(image)
has_image = image_features is not None
if not has_image:
message = "Note: Image processing is not available - continuing with text only.\n" + message
prompt = f"human: {'<image>' if has_image else ''}\n{message}\ngpt:"
context = ""
for turn in self.history[-3:]:
context += f"human: {turn[0]}\ngpt: {turn[1]}\n"
full_prompt = context + prompt
inputs = self.tokenizer(
full_prompt,
return_tensors="pt",
padding=True,
truncation=True,
max_length=512
)
inputs = {k: v.to(self.device) for k, v in inputs.items()}
if has_image:
inputs["image_features"] = image_features
with torch.no_grad():
outputs = self.model.generate(
**inputs,
max_new_tokens=256,
min_length=20,
temperature=0.7,
do_sample=True,
top_p=0.9,
top_k=40,
repetition_penalty=1.5,
no_repeat_ngram_size=3,
use_cache=True,
pad_token_id=self.tokenizer.pad_token_id,
eos_token_id=self.tokenizer.eos_token_id
)
else:
prompt = f"human: {message}\ngpt:"
context = ""
for turn in self.history[-3:]:
context += f"human: {turn[0]}\ngpt: {turn[1]}\n"
full_prompt = context + prompt
inputs = self.tokenizer(
full_prompt,
return_tensors="pt",
padding=True,
truncation=True,
max_length=512
)
inputs = {k: v.to(self.device) for k, v in inputs.items()}
with torch.no_grad():
outputs = self.model.generate(
**inputs,
max_new_tokens=150,
min_length=20,
temperature=0.6,
do_sample=True,
top_p=0.85,
top_k=30,
repetition_penalty=1.8,
no_repeat_ngram_size=4,
use_cache=True,
pad_token_id=self.tokenizer.pad_token_id,
eos_token_id=self.tokenizer.eos_token_id
)
response = self.tokenizer.decode(outputs[0], skip_special_tokens=True)
if "gpt:" in response:
response = response.split("gpt:")[-1].strip()
if "human:" in response:
response = response.split("human:")[0].strip()
if "<image>" in response:
response = response.replace("<image>", "").strip()
self.history.append((message, response))
return response
except Exception as e:
logging.error(f"Error generating response: {str(e)}")
logging.error(f"Full traceback:", exc_info=True)
return f"Error: {str(e)}"
def clear_history(self):
self.history = []
return None
def create_demo():
try:
model = LLaVAPhiModel()
with gr.Blocks(css="footer {visibility: hidden}") as demo:
gr.Markdown(
"""
# LLaVA-Phi Demo (ZeroGPU)
Chat with a vision-language model that can understand both text and images.
"""
)
chatbot = gr.Chatbot(height=400)
with gr.Row():
with gr.Column(scale=0.7):
msg = gr.Textbox(
show_label=False,
placeholder="Enter text and/or upload an image",
container=False
)
with gr.Column(scale=0.15, min_width=0):
clear = gr.Button("Clear")
with gr.Column(scale=0.15, min_width=0):
submit = gr.Button("Submit", variant="primary")
image = gr.Image(type="pil", label="Upload Image (Optional)")
def respond(message, chat_history, image):
if not message and image is None:
return chat_history
response = model.generate_response(message, image)
chat_history.append((message, response))
return "", chat_history
def clear_chat():
model.clear_history()
return None, None
submit.click(
respond,
[msg, chatbot, image],
[msg, chatbot],
)
clear.click(
clear_chat,
None,
[chatbot, image],
)
msg.submit(
respond,
[msg, chatbot, image],
[msg, chatbot],
)
return demo
except Exception as e:
logging.error(f"Error creating demo: {str(e)}")
raise
if __name__ == "__main__":
demo = create_demo()
demo.launch(
server_name="0.0.0.0",
server_port=7860,
share=True
) |