File size: 12,368 Bytes
eb25b77
 
 
 
 
 
 
 
 
 
 
 
 
 
8bfed10
eb25b77
 
8bfed10
 
 
 
 
 
eb25b77
8bfed10
 
eb25b77
8bfed10
 
 
 
 
 
 
 
 
 
 
 
 
 
eb25b77
 
8bfed10
 
 
 
 
 
 
 
 
 
 
 
eb25b77
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
8bfed10
eb25b77
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
8bfed10
 
 
eb25b77
 
8bfed10
 
eb25b77
 
 
 
 
 
8bfed10
eb25b77
8bfed10
 
 
 
eb25b77
 
 
 
 
 
 
 
 
 
 
 
8bfed10
eb25b77
 
 
 
 
 
 
 
 
 
 
 
 
 
 
8bfed10
 
 
 
 
 
 
 
 
 
 
eb25b77
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
8bfed10
 
 
 
 
 
 
 
 
 
 
 
 
eb25b77
 
8bfed10
eb25b77
 
 
 
 
8bfed10
 
 
 
 
eb25b77
 
 
 
 
 
 
 
 
 
 
 
 
 
8bfed10
 
 
 
eb25b77
 
 
 
8bfed10
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
eb25b77
 
8bfed10
 
 
eb25b77
 
 
8bfed10
 
 
eb25b77
 
 
8bfed10
eb25b77
8bfed10
 
eb25b77
 
 
8bfed10
eb25b77
8bfed10
eb25b77
 
 
 
8bfed10
eb25b77
8bfed10
eb25b77
 
8bfed10
 
 
 
eb25b77
 
 
 
 
 
 
 
 
 
8bfed10
eb25b77
 
 
 
 
8bfed10
eb25b77
 
 
 
 
8bfed10
 
eb25b77
 
 
8bfed10
 
eb25b77
8bfed10
 
eb25b77
 
 
8bfed10
eb25b77
 
 
 
 
 
8bfed10
 
 
 
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
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
#!/usr/bin/env python

import os
import re
import tempfile
from collections.abc import Iterator
from threading import Thread
import cv2
import gradio as gr
import spaces
import torch
from loguru import logger
from PIL import Image
from transformers import AutoProcessor, AutoModelForImageTextToText, TextIteratorStreamer
import gc

model_id = os.getenv("MODEL_ID", "google/medgemma-4b-it")

# Memory optimization settings
torch.backends.cudnn.benchmark = False
torch.backends.cudnn.deterministic = True

# Initialize processor and model with memory optimizations
processor = AutoProcessor.from_pretrained(model_id)

# Load model with aggressive memory optimizations
model = AutoModelForImageTextToText.from_pretrained(
    model_id, 
    device_map="auto", 
    torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
    token=os.environ.get("HF_TOKEN", "YOUR_HF_TOKEN"),
    # Memory optimization parameters
    low_cpu_mem_usage=True,
    trust_remote_code=True,
    # Use 8-bit quantization if available
    load_in_8bit=True if torch.cuda.is_available() else False,
    # Alternative: use 4-bit quantization for even more memory savings
    # load_in_4bit=True,
    # bnb_4bit_compute_dtype=torch.float16,
    # bnb_4bit_use_double_quant=True,
    # bnb_4bit_quant_type="nf4"
)

# Enable gradient checkpointing to save memory during inference
if hasattr(model, 'gradient_checkpointing_enable'):
    model.gradient_checkpointing_enable()

MAX_NUM_IMAGES = int(os.getenv("MAX_NUM_IMAGES", "3"))  # Reduced from 5 to 3

def cleanup_memory():
    """Aggressive memory cleanup"""
    gc.collect()
    if torch.cuda.is_available():
        torch.cuda.empty_cache()
        torch.cuda.synchronize()

def count_files_in_new_message(paths: list[str]) -> tuple[int, int]:
    image_count = 0
    video_count = 0
    for path in paths:
        if path.endswith(".mp4"):
            video_count += 1
        else:
            image_count += 1
    return image_count, video_count

def count_files_in_history(history: list[dict]) -> tuple[int, int]:
    image_count = 0
    video_count = 0
    for item in history:
        if item["role"] != "user" or isinstance(item["content"], str):
            continue
        if item["content"][0].endswith(".mp4"):
            video_count += 1
        else:
            image_count += 1
    return image_count, video_count

def validate_media_constraints(message: dict, history: list[dict]) -> bool:
    new_image_count, new_video_count = count_files_in_new_message(message["files"])
    history_image_count, history_video_count = count_files_in_history(history)
    image_count = history_image_count + new_image_count
    video_count = history_video_count + new_video_count
    
    if video_count > 1:
        gr.Warning("Only one video is supported.")
        return False
    if video_count == 1:
        if image_count > 0:
            gr.Warning("Mixing images and videos is not allowed.")
            return False
        if "<image>" in message["text"]:
            gr.Warning("Using <image> tags with video files is not supported.")
            return False
    if video_count == 0 and image_count > MAX_NUM_IMAGES:
        gr.Warning(f"You can upload up to {MAX_NUM_IMAGES} images.")
        return False
    if "<image>" in message["text"] and message["text"].count("<image>") != new_image_count:
        gr.Warning("The number of <image> tags in the text does not match the number of images.")
        return False
    return True

def downsample_video(video_path: str) -> list[tuple[Image.Image, float]]:
    vidcap = cv2.VideoCapture(video_path)
    fps = vidcap.get(cv2.CAP_PROP_FPS)
    total_frames = int(vidcap.get(cv2.CAP_PROP_FRAME_COUNT))

    # Reduce max frames for memory efficiency
    max_frames = min(MAX_NUM_IMAGES, 8)
    frame_interval = max(total_frames // max_frames, 1)
    frames: list[tuple[Image.Image, float]] = []

    for i in range(0, min(total_frames, max_frames * frame_interval), frame_interval):
        if len(frames) >= max_frames:
            break

        vidcap.set(cv2.CAP_PROP_POS_FRAMES, i)
        success, image = vidcap.read()
        if success:
            image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
            # Resize image to reduce memory usage
            pil_image = Image.fromarray(image)
            # Resize if too large
            max_size = 512
            if max(pil_image.size) > max_size:
                pil_image.thumbnail((max_size, max_size), Image.Resampling.LANCZOS)
            timestamp = round(i / fps, 2)
            frames.append((pil_image, timestamp))

    vidcap.release()
    return frames

def process_video(video_path: str) -> list[dict]:
    content = []
    frames = downsample_video(video_path)
    for frame in frames:
        pil_image, timestamp = frame
        with tempfile.NamedTemporaryFile(delete=False, suffix=".png") as temp_file:
            pil_image.save(temp_file.name, optimize=True)
            content.append({"type": "text", "text": f"Frame {timestamp}:"})
            content.append({"type": "image", "url": temp_file.name})
    logger.debug(f"{content=}")
    return content

def process_interleaved_images(message: dict) -> list[dict]:
    logger.debug(f"{message['files']=}")
    parts = re.split(r"(<image>)", message["text"])
    logger.debug(f"{parts=}")

    content = []
    image_index = 0
    for part in parts:
        logger.debug(f"{part=}")
        if part == "<image>":
            # Resize images before processing
            img = Image.open(message["files"][image_index])
            max_size = 512
            if max(img.size) > max_size:
                img.thumbnail((max_size, max_size), Image.Resampling.LANCZOS)
                # Save resized image
                with tempfile.NamedTemporaryFile(delete=False, suffix=".jpg") as temp_file:
                    img.save(temp_file.name, optimize=True, quality=85)
                    content.append({"type": "image", "url": temp_file.name})
            else:
                content.append({"type": "image", "url": message["files"][image_index]})
            logger.debug(f"file: {message['files'][image_index]}")
            image_index += 1
        elif part.strip():
            content.append({"type": "text", "text": part.strip()})
        elif isinstance(part, str) and part != "<image>":
            content.append({"type": "text", "text": part})
    logger.debug(f"{content=}")
    return content

def process_new_user_message(message: dict) -> list[dict]:
    if not message["files"]:
        return [{"type": "text", "text": message["text"]}]

    if message["files"][0].endswith(".mp4"):
        return [{"type": "text", "text": message["text"]}, *process_video(message["files"][0])]

    if "<image>" in message["text"]:
        return process_interleaved_images(message)

    # Process regular images with resizing
    processed_images = []
    for path in message["files"]:
        img = Image.open(path)
        max_size = 512
        if max(img.size) > max_size:
            img.thumbnail((max_size, max_size), Image.Resampling.LANCZOS)
            with tempfile.NamedTemporaryFile(delete=False, suffix=".jpg") as temp_file:
                img.save(temp_file.name, optimize=True, quality=85)
                processed_images.append({"type": "image", "url": temp_file.name})
        else:
            processed_images.append({"type": "image", "url": path})

    return [
        {"type": "text", "text": message["text"]},
        *processed_images,
    ]

def process_history(history: list[dict]) -> list[dict]:
    messages = []
    current_user_content: list[dict] = []
    
    # Limit history to prevent memory overflow
    recent_history = history[-10:] if len(history) > 10 else history
    
    for item in recent_history:
        if item["role"] == "assistant":
            if current_user_content:
                messages.append({"role": "user", "content": current_user_content})
                current_user_content = []
            messages.append({"role": "assistant", "content": [{"type": "text", "text": item["content"]}]})
        else:
            content = item["content"]
            if isinstance(content, str):
                current_user_content.append({"type": "text", "text": content})
            else:
                current_user_content.append({"type": "image", "url": content[0]})
    return messages

@spaces.GPU(duration=120)
def run(message: dict, history: list[dict], system_prompt: str = "", max_new_tokens: int = 1024) -> Iterator[str]:
    # Cleanup memory before processing
    cleanup_memory()
    
    if not validate_media_constraints(message, history):
        yield ""
        return

    try:
        messages = []
        if system_prompt:
            messages.append({"role": "system", "content": [{"type": "text", "text": system_prompt}]})
        messages.extend(process_history(history))
        messages.append({"role": "user", "content": process_new_user_message(message)})

        # Apply chat template with memory optimization
        inputs = processor.apply_chat_template(
            messages,
            add_generation_prompt=True,
            tokenize=True,
            return_dict=True,
            return_tensors="pt",
        ).to(device=model.device)

        # Reduce max_new_tokens to save memory
        max_new_tokens = min(max_new_tokens, 512)

        streamer = TextIteratorStreamer(processor, timeout=30.0, skip_prompt=True, skip_special_tokens=True)
        generate_kwargs = dict(
            inputs,
            max_new_tokens=max_new_tokens,
            streamer=streamer,
            temperature=0.8,  # Slightly reduced for more focused responses
            top_p=0.9,        # Reduced for efficiency
            top_k=50,         # Reduced for efficiency
            min_p=0.05,       # Added for better token filtering
            do_sample=True,
            pad_token_id=processor.tokenizer.eos_token_id,
            use_cache=True,
        )
        
        t = Thread(target=model.generate, kwargs=generate_kwargs)
        t.start()

        output = ""
        for delta in streamer:
            output += delta
            yield output
            
    except Exception as e:
        logger.error(f"Error during generation: {e}")
        yield f"Error: {str(e)}"
    finally:
        # Cleanup after generation
        cleanup_memory()

# Streamlined CSS
custom_css = """
.gr-chatbot {
    background-color: #ffffff;
    border-radius: 8px;
    border: 1px solid #e5e7eb;
}

.gr-textbox textarea {
    min-height: 100px;
    border-radius: 8px;
    padding: 12px;
}

.gr-button {
    background-color: #4f46e5 !important;
    color: white !important;
    border-radius: 6px !important;
    padding: 8px 16px !important;
}

.gr-interface {
    max-width: 800px;
    margin: 0 auto;
    padding: 16px;
}
"""

DESCRIPTION = """\
## Medical Vision-Language Assistant (Memory Optimized)

This AI assistant analyzes medical images and videos with memory efficiency optimizations.

**Features:**
- Medical image analysis (max 512px resolution for efficiency)
- Video frame processing (limited frames)
- Reduced memory footprint
- Optimized for resource-constrained environments
"""

demo = gr.ChatInterface(
    fn=run,
    type="messages",
    chatbot=gr.Chatbot(
        type="messages", 
        scale=1, 
        allow_tags=["image"],
        bubble_full_width=False,
        height=400,  # Fixed height to save memory
    ),
    textbox=gr.MultimodalTextbox(
        file_types=["image", ".mp4"], 
        file_count="multiple", 
        autofocus=True,
        placeholder="Upload images/video and ask questions...",
    ),
    multimodal=True,
    additional_inputs=[
        gr.Textbox(
            label="System Prompt", 
            value="You are a medical AI assistant. Provide concise, accurate analysis.",
            lines=2,
        ),
        gr.Slider(
            label="Response Length", 
            minimum=50, 
            maximum=512, 
            step=10, 
            value=256,
            info="Shorter responses use less memory"
        ),
    ],
    stop_btn=None,
    title="Medical Vision Assistant",
    description=DESCRIPTION,
    cache_examples=False,
    css=custom_css,
)

if __name__ == "__main__":
    demo.launch(
        server_name="0.0.0.0",
        server_port=7860,
        share=True)