File size: 8,658 Bytes
5c50991
36dacc6
970f416
dd0d47d
221e4b6
 
ca97f63
221e4b6
970f416
45c2159
34cd1e5
221e4b6
5c50991
 
 
 
 
 
 
 
07f3263
5c50991
 
 
 
 
 
 
07f3263
5c50991
 
 
 
 
 
 
07f3263
5c50991
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
221e4b6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5c50991
 
 
 
 
 
34cd1e5
 
5c50991
 
 
 
 
 
 
 
 
 
 
34cd1e5
5c50991
 
34cd1e5
 
 
 
 
5c50991
34cd1e5
 
 
 
 
 
 
 
 
45c2159
34cd1e5
 
 
 
45c2159
 
 
 
 
 
 
 
34cd1e5
 
 
 
 
 
 
 
b56b6ec
 
34cd1e5
 
45c2159
 
 
 
 
34cd1e5
 
 
 
 
 
 
238a95a
 
34cd1e5
 
238a95a
34cd1e5
5c50991
 
 
 
 
 
 
 
 
34cd1e5
5c50991
 
 
 
 
e1ad065
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
34cd1e5
 
e1ad065
 
c9c43a8
34cd1e5
 
 
 
 
5c50991
 
 
 
34cd1e5
 
292fb3c
221e4b6
970f416
65b3c3a
5c50991
 
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
import time
import logging
import gradio as gr
import cv2
import os
from pathlib import Path
from huggingface_hub import hf_hub_download
from llama_cpp import Llama
from llama_cpp.llama_chat_format import Llava15ChatHandler
import base64
import gc

# ----------------------------------------
# Model configurations: per-size prefixes and repos
MODELS = {
    "256M": {
        "model_repo": "mradermacher/SmolVLM2-256M-Video-Instruct-GGUF",
        "clip_repo":  "ggml-org/SmolVLM2-256M-Video-Instruct-GGUF",
        "model_prefix": "SmolVLM2-256M-Video-Instruct",
        "clip_prefix":  "mmproj-SmolVLM2-256M-Video-Instruct",
        "model_variants": ["Q2_K","Q8_0", "f16"],
        "clip_variants":  ["Q8_0", "f16"],
    },
    "500M": {
        "model_repo": "mradermacher/SmolVLM2-500M-Video-Instruct-GGUF",
        "clip_repo":  "ggml-org/SmolVLM2-500M-Video-Instruct-GGUF",
        "model_prefix": "SmolVLM2-500M-Video-Instruct",
        "clip_prefix":  "mmproj-SmolVLM2-500M-Video-Instruct",
        "model_variants": ["Q2_K","Q8_0", "f16"],
        "clip_variants":  ["Q8_0", "f16"],
    },
    "2.2B": {
        "model_repo": "mradermacher/SmolVLM2-2.2B-Instruct-GGUF",
        "clip_repo":  "ggml-org/SmolVLM2-2.2B-Instruct-GGUF",
        "model_prefix": "SmolVLM2-2.2B-Instruct",
        "clip_prefix":  "mmproj-SmolVLM2-2.2B-Instruct",
        "model_variants": ["Q2_K","Q4_K_M", "Q8_0", "f16"],
        "clip_variants":  ["Q8_0", "f16"],
    },
}

# ----------------------------------------
# Cache for loaded model instance
model_cache = {
    'size': None,
    'model_file': None,
    'clip_file': None,
    'llm': None
}

# Helper to download & symlink weights

def ensure_weights(size, model_file, clip_file):
    cfg = MODELS[size]
    if not os.path.exists(model_file):
        logging.info(f"Downloading model file {model_file} from {cfg['model_repo']}...")
        path = hf_hub_download(repo_id=cfg['model_repo'], filename=model_file)
        os.symlink(path, model_file)
    if not os.path.exists(clip_file):
        logging.info(f"Downloading CLIP file {clip_file} from {cfg['clip_repo']}...")
        path = hf_hub_download(repo_id=cfg['clip_repo'], filename=clip_file)
        os.symlink(path, clip_file)
    return model_file, clip_file

# Custom chat handler
class SmolVLM2ChatHandler(Llava15ChatHandler):
    CHAT_FORMAT = (
        "<|im_start|>"
        "{% for message in messages %}"
        "{{ message['role'] | capitalize }}"
        "{% if message['role']=='user' and message['content'][0]['type']=='image_url' %}:"
        "{% else %}: "
        "{% endif %}"
        "{% for content in message['content'] %}"
        "{% if content['type']=='text' %}{{ content['text'] }}"
        "{% elif content['type']=='image_url' %}"
        "{% if content['image_url'] is string %}"
        "{{ content['image_url'] }}\n"
        "{% elif content['image_url'] is mapping %}"
        "{{ content['image_url']['url'] }}\n"
        "{% endif %}"
        "{% endif %}"
        "{% endfor %}"
        "<end_of_utterance>\n"
        "{% endfor %}"
        "{% if add_generation_prompt %}Assistant:{% endif %}"
    )

# Load and cache LLM (only on dropdown change)

def update_llm(size, model_file, clip_file):
    if (model_cache['size'], model_cache['model_file'], model_cache['clip_file']) != (size, model_file, clip_file):
        mf, cf = ensure_weights(size, model_file, clip_file)
        handler = SmolVLM2ChatHandler(clip_model_path=cf, verbose=False)
        llm = Llama(model_path=mf, chat_handler=handler, n_ctx=1024,
                    verbose=False, n_threads=min(2, os.cpu_count()))
        model_cache.update({'size': size, 'model_file': mf, 'clip_file': cf, 'llm': llm})
    return None  # no UI output

# Build weight filename lists

def get_weight_files(size):
    cfg = MODELS[size]
    model_files = [f"{cfg['model_prefix']}.{v}.gguf" for v in cfg['model_variants']]
    clip_files  = [f"{cfg['clip_prefix']}-{v}.gguf"  for v in cfg['clip_variants']]
    return model_files, clip_files

# Caption using cached llm with real-time debug logs

def caption_frame(frame, size, model_file, clip_file, interval_ms, sys_prompt, usr_prompt):
    debug_msgs = []
    timestamp = time.strftime('%H:%M:%S')
    debug_msgs.append(f"[{timestamp}] Received frame shape: {frame.shape}")

    t_resize = time.time()
    img = cv2.resize(frame.copy(), (384, 384))
    elapsed = (time.time() - t_resize) * 1000
    timestamp = time.strftime('%H:%M:%S')
    debug_msgs.append(f"[{timestamp}] Resized to 384x384 in {elapsed:.1f} ms")

    timestamp = time.strftime('%H:%M:%S')
    debug_msgs.append(f"[{timestamp}] Sleeping for {interval_ms} ms")
    time.sleep(interval_ms / 1000)

    t_enc = time.time()
    success, jpeg = cv2.imencode('.jpg', img)
    elapsed = (time.time() - t_enc) * 1000
    timestamp = time.strftime('%H:%M:%S')
    debug_msgs.append(f"[{timestamp}] JPEG encode: success={success}, bytes={len(jpeg)} in {elapsed:.1f} ms")

    uri = 'data:image/jpeg;base64,' + base64.b64encode(jpeg.tobytes()).decode()
    messages = [
        {"role": "system", "content": sys_prompt},
        {"role": "user",   "content": [
            {"type": "image_url", "image_url": uri},
            {"type": "text",      "text": usr_prompt}
        ]}
    ]

    timestamp = time.strftime('%H:%M:%S')
    debug_msgs.append(f"[{timestamp}] Sending prompt of length {len(usr_prompt)} to LLM")
    # re-init handler for image
    model_cache['llm'].chat_handler = SmolVLM2ChatHandler(clip_model_path=clip_file, verbose=False)
    timestamp = time.strftime('%H:%M:%S')
    debug_msgs.append(f"[{timestamp}] Reinitialized chat handler")

    debug_msgs.append(f"[{timestamp}] CPU count = {os.cpu_count()}")

    t_start = time.time()
    resp = model_cache['llm'].create_chat_completion(
        messages=messages,
        max_tokens=128,
        temperature=0.1,
        stop=["<end_of_utterance>"]
    )
    elapsed = (time.time() - t_start) * 1000
    timestamp = time.strftime('%H:%M:%S')
    debug_msgs.append(f"[{timestamp}] LLM response in {elapsed:.1f} ms")

    content = resp.get('choices', [{}])[0].get('message', {}).get('content', '').strip()
    timestamp = time.strftime('%H:%M:%S')
    debug_msgs.append(f"[{timestamp}] Caption length: {len(content)} chars")

    gc.collect()
    timestamp = time.strftime('%H:%M:%S')
    debug_msgs.append(f"[{timestamp}] Garbage collected")

    return content, "\n".join(debug_msgs)

# Gradio UI

def main():
    logging.basicConfig(level=logging.INFO)
    default = '2.2B'
    mf, cf = get_weight_files(default)

    with gr.Blocks() as demo:
        gr.Markdown("## 🎥 Real-Time Camera Captioning with Debug Logs")
        with gr.Row():
            size_dd   = gr.Dropdown(list(MODELS.keys()), value=default, label='Model Size')
            model_dd  = gr.Dropdown(mf, value=mf[0], label='Decoder Weights')
            clip_dd   = gr.Dropdown(cf, value=cf[0], label='CLIP Weights')

                # When size changes: update dropdowns AND preload llm with the new first weights
        def on_size_change(sz):
            mlist, clist = get_weight_files(sz)
            # update dropdown choices and default values
            update_ui = (
                gr.update(choices=mlist, value=mlist[0]),
                gr.update(choices=clist, value=clist[0])
            )
            # preload with first weights
            update_llm(sz, mlist[0], clist[0])
            return update_ui
        size_dd.change(
            fn=on_size_change,
            inputs=[size_dd],
            outputs=[model_dd, clip_dd]
        )
        model_dd.change(lambda sz, mf, cf: update_llm(sz, mf, cf), inputs=[size_dd, model_dd, clip_dd], outputs=[])
        clip_dd.change(lambda sz, mf, cf: update_llm(sz, mf, cf), inputs=[size_dd, model_dd, clip_dd], outputs=[])
        update_llm(default, mf[0], cf[0])

        interval = gr.Slider(100, 20000, step=100, value=3000, label='Interval (ms)')
        sys_p    = gr.Textbox(lines=2, value="Focus on key dramatic action…", label='System Prompt')
        usr_p    = gr.Textbox(lines=1, value="What is happening in this image?", label='User Prompt')
        cam      = gr.Image(sources=['webcam'], streaming=True, label='Webcam Feed')
        cap      = gr.Textbox(interactive=False, label='Caption')
        log_box  = gr.Textbox(lines=8, interactive=False, label='Debug Log')

        cam.stream(
            fn=caption_frame,
            inputs=[cam, size_dd, model_dd, clip_dd, interval, sys_p, usr_p],
            outputs=[cap, log_box],
            time_limit=600
        )

    demo.launch()

if __name__ == '__main__':
    main()