BullseyeMxP
commited on
Commit
•
1a23e22
1
Parent(s):
a351b6b
Update app.py
Browse files
app.py
CHANGED
@@ -114,20 +114,25 @@ assert isinstance(tokenizer, PreTrainedTokenizer) or isinstance(tokenizer, PreTr
|
|
114 |
print("Loading LLM")
|
115 |
print("Loading VLM's custom text model")
|
116 |
|
117 |
-
# Configure 4-bit quantization
|
118 |
bnb_config = BitsAndBytesConfig(
|
119 |
load_in_4bit=True,
|
120 |
bnb_4bit_quant_type="nf4",
|
121 |
bnb_4bit_compute_dtype=torch.float16,
|
122 |
bnb_4bit_use_double_quant=True,
|
|
|
123 |
)
|
124 |
|
125 |
text_model = AutoModelForCausalLM.from_pretrained(
|
126 |
CHECKPOINT_PATH / "text_model",
|
127 |
device_map="auto",
|
128 |
quantization_config=bnb_config,
|
129 |
-
torch_dtype=torch.float16
|
|
|
130 |
)
|
|
|
|
|
|
|
131 |
text_model.gradient_checkpointing_enable()
|
132 |
text_model.eval()
|
133 |
text_model = torch.compile(text_model)
|
@@ -140,15 +145,27 @@ image_adapter.eval()
|
|
140 |
image_adapter.to("cuda")
|
141 |
image_adapter = torch.compile(image_adapter)
|
142 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
143 |
@spaces.GPU()
|
144 |
@torch.no_grad()
|
145 |
def stream_chat(input_image: Image.Image, caption_type: str, caption_length: str | int, extra_options: list[str], name_input: str, custom_prompt: str) -> tuple[str, str]:
|
|
|
146 |
torch.cuda.empty_cache()
|
147 |
gc.collect()
|
148 |
|
149 |
-
#
|
150 |
length = None if caption_length == "any" else caption_length
|
151 |
-
|
152 |
if isinstance(length, str):
|
153 |
try:
|
154 |
length = int(length)
|
@@ -176,57 +193,42 @@ def stream_chat(input_image: Image.Image, caption_type: str, caption_length: str
|
|
176 |
|
177 |
if custom_prompt.strip() != "":
|
178 |
prompt_str = custom_prompt.strip()
|
179 |
-
|
180 |
-
# For debugging
|
181 |
-
print(f"Prompt: {prompt_str}")
|
182 |
|
183 |
-
#
|
184 |
image = input_image.resize((384, 384), Image.LANCZOS)
|
185 |
-
image = image.convert('RGB')
|
186 |
pixel_values = TVF.pil_to_tensor(image).unsqueeze(0) / 255.0
|
187 |
-
pixel_values = TVF.normalize(pixel_values, [0.5, 0.5, 0.5], [0.5, 0.5, 0.5])
|
188 |
pixel_values = pixel_values.to('cuda', dtype=torch.float16)
|
189 |
|
190 |
-
#
|
191 |
-
with torch.amp.
|
192 |
vision_outputs = clip_model(pixel_values=pixel_values, output_hidden_states=True)
|
193 |
embedded_images = image_adapter(vision_outputs.hidden_states)
|
194 |
embedded_images = embedded_images.to('cuda', dtype=torch.float16)
|
195 |
|
196 |
-
# Build the conversation
|
197 |
convo = [
|
198 |
-
{
|
199 |
-
|
200 |
-
"content": "You are a helpful image captioner.",
|
201 |
-
},
|
202 |
-
{
|
203 |
-
"role": "user",
|
204 |
-
"content": prompt_str,
|
205 |
-
},
|
206 |
]
|
207 |
|
208 |
-
# Format
|
209 |
-
convo_string = tokenizer.apply_chat_template(convo, tokenize
|
210 |
-
assert isinstance(convo_string, str)
|
211 |
-
|
212 |
-
# Tokenize the conversation
|
213 |
convo_tokens = tokenizer.encode(convo_string, return_tensors="pt", add_special_tokens=False, truncation=False)
|
214 |
prompt_tokens = tokenizer.encode(prompt_str, return_tensors="pt", add_special_tokens=False, truncation=False)
|
215 |
-
|
216 |
-
convo_tokens = convo_tokens.squeeze(0)
|
217 |
prompt_tokens = prompt_tokens.squeeze(0)
|
218 |
|
219 |
-
# Calculate
|
220 |
eot_id_indices = (convo_tokens == tokenizer.convert_tokens_to_ids("<|eot_id|>")).nonzero(as_tuple=True)[0].tolist()
|
221 |
-
|
222 |
-
|
223 |
-
preamble_len = eot_id_indices[1] - prompt_tokens.shape[0] # Number of tokens before the prompt
|
224 |
|
225 |
-
#
|
226 |
-
convo_tokens = convo_tokens.unsqueeze(0).to('cuda')
|
227 |
convo_embeds = text_model.model.embed_tokens(convo_tokens)
|
228 |
|
229 |
-
# Construct the input
|
230 |
input_embeds = torch.cat([
|
231 |
convo_embeds[:, :preamble_len],
|
232 |
embedded_images,
|
@@ -240,27 +242,31 @@ def stream_chat(input_image: Image.Image, caption_type: str, caption_length: str
|
|
240 |
], dim=1)
|
241 |
attention_mask = torch.ones_like(input_ids)
|
242 |
|
243 |
-
#
|
244 |
-
|
245 |
-
|
246 |
-
with torch.amp.autocast_mode.autocast('cuda', dtype=torch.float16):
|
247 |
generate_ids = text_model.generate(
|
248 |
input_ids,
|
249 |
inputs_embeds=input_embeds,
|
250 |
attention_mask=attention_mask,
|
251 |
max_new_tokens=300,
|
252 |
do_sample=True,
|
253 |
-
|
254 |
-
|
|
|
|
|
|
|
|
|
255 |
)
|
256 |
|
257 |
-
#
|
258 |
generate_ids = generate_ids[:, input_ids.shape[1]:]
|
259 |
if generate_ids[0][-1] == tokenizer.eos_token_id or generate_ids[0][-1] == tokenizer.convert_tokens_to_ids("<|eot_id|>"):
|
260 |
generate_ids = generate_ids[:, :-1]
|
261 |
|
262 |
-
caption = tokenizer.batch_decode(generate_ids, skip_special_tokens=
|
263 |
|
|
|
|
|
264 |
torch.cuda.empty_cache()
|
265 |
gc.collect()
|
266 |
|
@@ -275,7 +281,7 @@ def process_directory(directory_path, caption_type, caption_length, extra_option
|
|
275 |
img_path = os.path.join(directory_path, filename)
|
276 |
img = Image.open(img_path)
|
277 |
|
278 |
-
|
279 |
|
280 |
# Save caption to a .txt file
|
281 |
txt_filename = os.path.splitext(filename)[0] + '.txt'
|
@@ -284,9 +290,29 @@ def process_directory(directory_path, caption_type, caption_length, extra_option
|
|
284 |
f.write(caption)
|
285 |
|
286 |
processed_images.append(img_path)
|
287 |
-
captions.append(
|
288 |
|
289 |
-
return processed_images, captions
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
290 |
|
291 |
# Custom CSS for a futuristic, neon-inspired theme
|
292 |
custom_css = """
|
@@ -439,27 +465,7 @@ with gr.Blocks(css=custom_css) as demo:
|
|
439 |
|
440 |
with gr.Row():
|
441 |
output_gallery = gr.Gallery(label="Processed Images", elem_classes="output-box")
|
442 |
-
output_text = gr.
|
443 |
-
|
444 |
-
def process_and_display(images, caption_type, caption_length, extra_options, name_input, custom_prompt):
|
445 |
-
processed_images = []
|
446 |
-
captions = []
|
447 |
-
|
448 |
-
for img_file in images:
|
449 |
-
img = Image.open(img_file.name)
|
450 |
-
prompt, caption = stream_chat(img, caption_type, caption_length, extra_options, name_input, custom_prompt)
|
451 |
-
processed_images.append(img_file.name)
|
452 |
-
captions.append({"filename": img_file.name, "caption": caption})
|
453 |
-
|
454 |
-
return processed_images, captions
|
455 |
-
|
456 |
-
def process_input(input_images, directory_path, caption_type, caption_length, extra_options, name_input, custom_prompt):
|
457 |
-
if directory_path:
|
458 |
-
return process_directory(directory_path, caption_type, caption_length, extra_options, name_input, custom_prompt)
|
459 |
-
elif input_images:
|
460 |
-
return process_and_display(input_images, caption_type, caption_length, extra_options, name_input, custom_prompt)
|
461 |
-
else:
|
462 |
-
return [], []
|
463 |
|
464 |
run_button.click(
|
465 |
fn=process_input,
|
|
|
114 |
print("Loading LLM")
|
115 |
print("Loading VLM's custom text model")
|
116 |
|
117 |
+
# Configure 4-bit quantization with more aggressive settings
|
118 |
bnb_config = BitsAndBytesConfig(
|
119 |
load_in_4bit=True,
|
120 |
bnb_4bit_quant_type="nf4",
|
121 |
bnb_4bit_compute_dtype=torch.float16,
|
122 |
bnb_4bit_use_double_quant=True,
|
123 |
+
llm_int8_enable_fp32_cpu_offload=True
|
124 |
)
|
125 |
|
126 |
text_model = AutoModelForCausalLM.from_pretrained(
|
127 |
CHECKPOINT_PATH / "text_model",
|
128 |
device_map="auto",
|
129 |
quantization_config=bnb_config,
|
130 |
+
torch_dtype=torch.float16,
|
131 |
+
low_cpu_mem_usage=True
|
132 |
)
|
133 |
+
|
134 |
+
# Enable memory efficient attention
|
135 |
+
text_model.config.use_memory_efficient_attention = True
|
136 |
text_model.gradient_checkpointing_enable()
|
137 |
text_model.eval()
|
138 |
text_model = torch.compile(text_model)
|
|
|
145 |
image_adapter.to("cuda")
|
146 |
image_adapter = torch.compile(image_adapter)
|
147 |
|
148 |
+
# Optimize CLIP model
|
149 |
+
clip_model = clip_model.half() # Convert to FP16
|
150 |
+
clip_model.eval()
|
151 |
+
clip_model.requires_grad_(False)
|
152 |
+
clip_model = torch.compile(clip_model)
|
153 |
+
|
154 |
+
# Optimize image adapter
|
155 |
+
image_adapter = image_adapter.half() # Convert to FP16
|
156 |
+
image_adapter.eval()
|
157 |
+
image_adapter.requires_grad_(False)
|
158 |
+
image_adapter = torch.compile(image_adapter)
|
159 |
+
|
160 |
@spaces.GPU()
|
161 |
@torch.no_grad()
|
162 |
def stream_chat(input_image: Image.Image, caption_type: str, caption_length: str | int, extra_options: list[str], name_input: str, custom_prompt: str) -> tuple[str, str]:
|
163 |
+
# Clear memory at the start
|
164 |
torch.cuda.empty_cache()
|
165 |
gc.collect()
|
166 |
|
167 |
+
# Build prompt string
|
168 |
length = None if caption_length == "any" else caption_length
|
|
|
169 |
if isinstance(length, str):
|
170 |
try:
|
171 |
length = int(length)
|
|
|
193 |
|
194 |
if custom_prompt.strip() != "":
|
195 |
prompt_str = custom_prompt.strip()
|
|
|
|
|
|
|
196 |
|
197 |
+
# Resize image to exact dimensions needed
|
198 |
image = input_image.resize((384, 384), Image.LANCZOS)
|
199 |
+
image = image.convert('RGB')
|
200 |
pixel_values = TVF.pil_to_tensor(image).unsqueeze(0) / 255.0
|
201 |
+
pixel_values = TVF.normalize(pixel_values, [0.5, 0.5, 0.5], [0.5, 0.5, 0.5])
|
202 |
pixel_values = pixel_values.to('cuda', dtype=torch.float16)
|
203 |
|
204 |
+
# Process image with optimized memory usage
|
205 |
+
with torch.amp.autocast('cuda', dtype=torch.float16):
|
206 |
vision_outputs = clip_model(pixel_values=pixel_values, output_hidden_states=True)
|
207 |
embedded_images = image_adapter(vision_outputs.hidden_states)
|
208 |
embedded_images = embedded_images.to('cuda', dtype=torch.float16)
|
209 |
|
210 |
+
# Build the conversation with minimal overhead
|
211 |
convo = [
|
212 |
+
{"role": "system", "content": "You are a helpful image captioner."},
|
213 |
+
{"role": "user", "content": prompt_str},
|
|
|
|
|
|
|
|
|
|
|
|
|
214 |
]
|
215 |
|
216 |
+
# Format and tokenize efficiently
|
217 |
+
convo_string = tokenizer.apply_chat_template(convo, tokenize=False, add_generation_prompt=True)
|
|
|
|
|
|
|
218 |
convo_tokens = tokenizer.encode(convo_string, return_tensors="pt", add_special_tokens=False, truncation=False)
|
219 |
prompt_tokens = tokenizer.encode(prompt_str, return_tensors="pt", add_special_tokens=False, truncation=False)
|
220 |
+
|
221 |
+
convo_tokens = convo_tokens.squeeze(0)
|
222 |
prompt_tokens = prompt_tokens.squeeze(0)
|
223 |
|
224 |
+
# Calculate injection point
|
225 |
eot_id_indices = (convo_tokens == tokenizer.convert_tokens_to_ids("<|eot_id|>")).nonzero(as_tuple=True)[0].tolist()
|
226 |
+
preamble_len = eot_id_indices[1] - prompt_tokens.shape[0]
|
|
|
|
|
227 |
|
228 |
+
# Prepare input tensors efficiently
|
229 |
+
convo_tokens = convo_tokens.unsqueeze(0).to('cuda')
|
230 |
convo_embeds = text_model.model.embed_tokens(convo_tokens)
|
231 |
|
|
|
232 |
input_embeds = torch.cat([
|
233 |
convo_embeds[:, :preamble_len],
|
234 |
embedded_images,
|
|
|
242 |
], dim=1)
|
243 |
attention_mask = torch.ones_like(input_ids)
|
244 |
|
245 |
+
# Generate with optimized settings
|
246 |
+
with torch.amp.autocast('cuda', dtype=torch.float16):
|
|
|
|
|
247 |
generate_ids = text_model.generate(
|
248 |
input_ids,
|
249 |
inputs_embeds=input_embeds,
|
250 |
attention_mask=attention_mask,
|
251 |
max_new_tokens=300,
|
252 |
do_sample=True,
|
253 |
+
use_cache=True,
|
254 |
+
pad_token_id=tokenizer.pad_token_id,
|
255 |
+
num_beams=1, # Disable beam search for faster generation
|
256 |
+
temperature=0.7, # Lower temperature for more focused generation
|
257 |
+
top_p=0.9, # Nucleus sampling for efficiency
|
258 |
+
repetition_penalty=1.2, # Prevent repetition
|
259 |
)
|
260 |
|
261 |
+
# Process output efficiently
|
262 |
generate_ids = generate_ids[:, input_ids.shape[1]:]
|
263 |
if generate_ids[0][-1] == tokenizer.eos_token_id or generate_ids[0][-1] == tokenizer.convert_tokens_to_ids("<|eot_id|>"):
|
264 |
generate_ids = generate_ids[:, :-1]
|
265 |
|
266 |
+
caption = tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=True)[0]
|
267 |
|
268 |
+
# Clear memory
|
269 |
+
del vision_outputs, embedded_images, input_embeds, generate_ids
|
270 |
torch.cuda.empty_cache()
|
271 |
gc.collect()
|
272 |
|
|
|
281 |
img_path = os.path.join(directory_path, filename)
|
282 |
img = Image.open(img_path)
|
283 |
|
284 |
+
_, caption = stream_chat(img, caption_type, caption_length, extra_options, name_input, custom_prompt)
|
285 |
|
286 |
# Save caption to a .txt file
|
287 |
txt_filename = os.path.splitext(filename)[0] + '.txt'
|
|
|
290 |
f.write(caption)
|
291 |
|
292 |
processed_images.append(img_path)
|
293 |
+
captions.append(caption)
|
294 |
|
295 |
+
return processed_images, "\n\n".join(captions) # Join captions with double newline for readability
|
296 |
+
|
297 |
+
def process_and_display(images, caption_type, caption_length, extra_options, name_input, custom_prompt):
|
298 |
+
processed_images = []
|
299 |
+
captions = []
|
300 |
+
|
301 |
+
for img_file in images:
|
302 |
+
img = Image.open(img_file.name)
|
303 |
+
_, caption = stream_chat(img, caption_type, caption_length, extra_options, name_input, custom_prompt)
|
304 |
+
processed_images.append(img_file.name)
|
305 |
+
captions.append(caption)
|
306 |
+
|
307 |
+
return processed_images, "\n\n".join(captions) # Join captions with double newline for readability
|
308 |
+
|
309 |
+
def process_input(input_images, directory_path, caption_type, caption_length, extra_options, name_input, custom_prompt):
|
310 |
+
if directory_path:
|
311 |
+
return process_directory(directory_path, caption_type, caption_length, extra_options, name_input, custom_prompt)
|
312 |
+
elif input_images:
|
313 |
+
return process_and_display(input_images, caption_type, caption_length, extra_options, name_input, custom_prompt)
|
314 |
+
else:
|
315 |
+
return [], ""
|
316 |
|
317 |
# Custom CSS for a futuristic, neon-inspired theme
|
318 |
custom_css = """
|
|
|
465 |
|
466 |
with gr.Row():
|
467 |
output_gallery = gr.Gallery(label="Processed Images", elem_classes="output-box")
|
468 |
+
output_text = gr.Textbox(label="Generated Captions", elem_classes="output-box", lines=10)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
469 |
|
470 |
run_button.click(
|
471 |
fn=process_input,
|