Spaces:
Running
on
Zero
Running
on
Zero
Upload 2 files
Browse files
app.py
CHANGED
@@ -1,167 +1,289 @@
|
|
|
|
|
|
1 |
import gradio as gr
|
2 |
-
from
|
3 |
-
from transformers.image_utils import load_image
|
4 |
-
from threading import Thread
|
5 |
-
import time
|
6 |
-
import torch
|
7 |
import spaces
|
8 |
-
import cv2
|
9 |
-
import numpy as np
|
10 |
-
from PIL import Image
|
11 |
|
12 |
-
|
13 |
-
"""
|
14 |
-
Returns an HTML snippet for a thin progress bar with a label.
|
15 |
-
The progress bar is styled as a dark animated bar.
|
16 |
-
"""
|
17 |
-
return f'''
|
18 |
-
<div style="display: flex; align-items: center;">
|
19 |
-
<span style="margin-right: 10px; font-size: 14px;">{label}</span>
|
20 |
-
<div style="width: 110px; height: 5px; background-color: #9370DB; border-radius: 2px; overflow: hidden;">
|
21 |
-
<div style="width: 100%; height: 100%; background-color: #4B0082; animation: loading 1.5s linear infinite;"></div>
|
22 |
-
</div>
|
23 |
-
</div>
|
24 |
-
<style>
|
25 |
-
@keyframes loading {{
|
26 |
-
0% {{ transform: translateX(-100%); }}
|
27 |
-
100% {{ transform: translateX(100%); }}
|
28 |
-
}}
|
29 |
-
</style>
|
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 |
-
# MODEL_ID = "XiaomiMiMo/MiMo-VL-7B-RL"
|
58 |
-
MODEL_ID = "XiaomiMiMo/MiMo-VL-7B-RL-2508"
|
59 |
-
processor = AutoProcessor.from_pretrained(MODEL_ID, trust_remote_code=True)
|
60 |
-
model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
|
61 |
-
MODEL_ID,
|
62 |
-
trust_remote_code=True,
|
63 |
-
torch_dtype=torch.bfloat16
|
64 |
-
).to("cuda").eval()
|
65 |
|
66 |
-
@spaces.GPU
|
67 |
-
def
|
68 |
-
|
69 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
70 |
|
71 |
-
if text.strip().lower().startswith("@video-infer"):
|
72 |
-
# Remove the tag from the query.
|
73 |
-
text = text[len("@video-infer"):].strip()
|
74 |
-
if not files:
|
75 |
-
yield "⚠️ Please upload a video file along with your `@video-infer` query."
|
76 |
-
return
|
77 |
-
# Assume the first file is a video.
|
78 |
-
video_path = files[0]
|
79 |
-
frames = downsample_video(video_path)
|
80 |
-
if not frames:
|
81 |
-
yield "⚠️ Could not process the video (no frames were read)."
|
82 |
-
return
|
83 |
-
# Build messages: start with the text prompt.
|
84 |
-
messages = [
|
85 |
-
{
|
86 |
-
"role": "user",
|
87 |
-
"content": [{"type": "text", "text": text}]
|
88 |
-
}
|
89 |
-
]
|
90 |
-
# Append each frame with a timestamp label.
|
91 |
-
for image, timestamp in frames:
|
92 |
-
messages[0]["content"].append({"type": "text", "text": f"Frame {timestamp}:"})
|
93 |
-
messages[0]["content"].append({"type": "image", "image": image})
|
94 |
-
# Collect only the images from the frames.
|
95 |
-
video_images = [image for image, _ in frames]
|
96 |
-
# Prepare the prompt.
|
97 |
-
prompt = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
|
98 |
-
inputs = processor(
|
99 |
-
text=[prompt],
|
100 |
-
images=video_images,
|
101 |
-
return_tensors="pt",
|
102 |
-
padding=True,
|
103 |
-
).to("cuda")
|
104 |
-
# Set up streaming generation.
|
105 |
-
streamer = TextIteratorStreamer(processor.tokenizer, skip_prompt=True, skip_special_tokens=True)
|
106 |
-
generation_kwargs = dict(inputs, streamer=streamer, max_new_tokens=2048)
|
107 |
-
thread = Thread(target=model.generate, kwargs=generation_kwargs)
|
108 |
-
thread.start()
|
109 |
-
buffer = ""
|
110 |
-
yield progress_bar_html("Processing video with MiMo-VL-7B-RL Model")
|
111 |
-
for new_text in streamer:
|
112 |
-
buffer += new_text
|
113 |
-
time.sleep(0.01)
|
114 |
-
yield buffer
|
115 |
-
return
|
116 |
|
117 |
-
|
118 |
-
|
119 |
-
|
120 |
-
|
121 |
-
|
122 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
123 |
|
124 |
-
if text == "" and not images:
|
125 |
-
yield "⚠️ Please enter a question and/or upload image(s)."
|
126 |
-
return
|
127 |
-
if text == "" and images:
|
128 |
-
yield "⚠️ Please enter a text prompt along with the image(s)."
|
129 |
-
return
|
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 |
-
demo.launch(debug=True)
|
|
|
1 |
+
# modified from https://github.com/XiaomiMiMo/MiMo-VL/tree/main/app.py
|
2 |
+
import os
|
3 |
import gradio as gr
|
4 |
+
from infer import MiMoVLInfer
|
|
|
|
|
|
|
|
|
5 |
import spaces
|
|
|
|
|
|
|
6 |
|
7 |
+
infer = MiMoVLInfer(checkpoint_path=os.environ.get('CKPT_PATH'))
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
8 |
|
9 |
+
label_translations = {
|
10 |
+
"gr_chatinterface_ofl": {
|
11 |
+
"English": "Chatbot",
|
12 |
+
},
|
13 |
+
"gr_chatinterface_ol": {
|
14 |
+
"English": "Chatbot",
|
15 |
+
},
|
16 |
+
"gr_tab_ol": {
|
17 |
+
"English": "Online",
|
18 |
+
},
|
19 |
+
"gr_tab_ofl": {
|
20 |
+
"English": "Offline",
|
21 |
+
},
|
22 |
+
"gr_temperature": {
|
23 |
+
"English": "Temperature",
|
24 |
+
},
|
25 |
+
"gr_webcam_image": {
|
26 |
+
"English": "🤳 Open Webcam",
|
27 |
+
},
|
28 |
+
"gr_webcam_images": {
|
29 |
+
"English": "📹 Recorded Frames",
|
30 |
+
},
|
31 |
+
"gr_chatinterface_ofl.textbox.placeholder": {
|
32 |
+
"English":
|
33 |
+
"Ask me anything. You can also drop in images and .mp4 videos.",
|
34 |
+
},
|
35 |
+
"gr_chatinterface_ol.textbox.placeholder": {
|
36 |
+
"English": "Ask me anything...",
|
37 |
+
}
|
38 |
+
}
|
39 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
40 |
|
41 |
+
@spaces.GPU(duration=120) # bump if your requests take >60s
|
42 |
+
def offline_chat(gr_inputs: dict, gr_history: list, infer_history: list, temperature: float):
|
43 |
+
infer.to_device("cuda")
|
44 |
+
try:
|
45 |
+
yield [{"role": "assistant", "content": "⏳ Reserving GPU & preparing inference…"}], infer_history
|
46 |
+
for response_text, infer_history in infer(inputs=gr_inputs,
|
47 |
+
history=infer_history,
|
48 |
+
temperature=temperature):
|
49 |
+
if response_text.startswith('<think>') and '</think>' not in response_text:
|
50 |
+
reasoning_text = response_text.lstrip('<think>')
|
51 |
+
response_message = [{
|
52 |
+
"role": "assistant",
|
53 |
+
"content": reasoning_text,
|
54 |
+
'metadata': {'title': '🤔 Thinking'}
|
55 |
+
}]
|
56 |
+
yield response_message, infer_history
|
57 |
+
elif '<think>' in response_text and '</think>' in response_text:
|
58 |
+
reasoning_text, response_text2 = response_text.split('</think>', 1)
|
59 |
+
reasoning_text = reasoning_text.lstrip('<think>')
|
60 |
+
response_message = [{
|
61 |
+
"role": "assistant",
|
62 |
+
"content": reasoning_text,
|
63 |
+
'metadata': {'title': '🤔 Thinking'}
|
64 |
+
}, {
|
65 |
+
"role": "assistant",
|
66 |
+
"content": response_text2
|
67 |
+
}]
|
68 |
+
yield response_message, infer_history
|
69 |
+
else:
|
70 |
+
yield [{"role": "assistant", "content": response_text}], infer_history
|
71 |
+
finally:
|
72 |
+
infer.to_device("cpu")
|
73 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
74 |
|
75 |
+
@spaces.GPU(duration=120)
|
76 |
+
def online_record_chat(text: str, gr_history: list, gr_webcam_images: list, gr_counter: int,
|
77 |
+
infer_history: list, temperature: float):
|
78 |
+
infer.to_device("cuda")
|
79 |
+
try:
|
80 |
+
if not gr_webcam_images:
|
81 |
+
gr_webcam_images = []
|
82 |
+
gr_webcam_images = gr_webcam_images[gr_counter:]
|
83 |
+
inputs = {'text': text, 'files': [webp for webp, _ in gr_webcam_images]}
|
84 |
+
# send an immediate chunk
|
85 |
+
yield f'received {len(gr_webcam_images)} new frames, processing…', gr_counter + len(gr_webcam_images), infer_history
|
86 |
+
for response_message, infer_history in offline_chat(
|
87 |
+
inputs, gr_history, infer_history, temperature):
|
88 |
+
yield response_message, gr.skip(), infer_history
|
89 |
+
finally:
|
90 |
+
infer.to_device("cpu")
|
91 |
|
|
|
|
|
|
|
|
|
|
|
|
|
92 |
|
93 |
+
with gr.Blocks() as demo:
|
94 |
+
gr.Markdown("""<center><font size=8>MiMo-7b-VL</center>""")
|
95 |
+
with gr.Column():
|
96 |
+
# gr_title = gr.Markdown('# MiMo-VL')
|
97 |
+
|
98 |
+
with gr.Row():
|
99 |
+
gr_lang_selector = gr.Dropdown(choices=["English"],
|
100 |
+
value="English",
|
101 |
+
label="🌐 Interface",
|
102 |
+
interactive=True,
|
103 |
+
min_width=250,
|
104 |
+
scale=0)
|
105 |
+
with gr.Tabs():
|
106 |
+
with gr.Tab("Offline") as gr_tab_ofl:
|
107 |
+
gr_infer_history = gr.State([])
|
108 |
+
gr_temperature_hidden = gr.Slider(minimum=0.0,
|
109 |
+
maximum=2.0,
|
110 |
+
step=0.1,
|
111 |
+
value=1.0,
|
112 |
+
interactive=True,
|
113 |
+
visible=False)
|
114 |
+
gr_chatinterface_ofl = gr.ChatInterface(
|
115 |
+
fn=offline_chat,
|
116 |
+
type="messages",
|
117 |
+
multimodal=True,
|
118 |
+
chatbot=gr.Chatbot(height=800),
|
119 |
+
textbox=gr.MultimodalTextbox(
|
120 |
+
file_count="multiple",
|
121 |
+
file_types=["image", ".mp4"],
|
122 |
+
sources=["upload"],
|
123 |
+
stop_btn=True,
|
124 |
+
placeholder=label_translations[
|
125 |
+
'gr_chatinterface_ofl.textbox.placeholder']['English'],
|
126 |
+
),
|
127 |
+
additional_inputs=[
|
128 |
+
gr_infer_history, gr_temperature_hidden
|
129 |
+
],
|
130 |
+
additional_outputs=[gr_infer_history],
|
131 |
+
)
|
132 |
+
gr.on(triggers=[gr_chatinterface_ofl.chatbot.clear],
|
133 |
+
fn=lambda: [],
|
134 |
+
outputs=[gr_infer_history])
|
135 |
+
with gr.Row():
|
136 |
+
with gr.Column(scale=1, min_width=200):
|
137 |
+
gr_temperature_ofl = gr.Slider(
|
138 |
+
minimum=0.0,
|
139 |
+
maximum=2.0,
|
140 |
+
step=0.1,
|
141 |
+
value=0.4,
|
142 |
+
label=label_translations['gr_temperature']['English'],
|
143 |
+
interactive=True)
|
144 |
+
gr_temperature_ofl.change(lambda x: x,
|
145 |
+
inputs=gr_temperature_ofl,
|
146 |
+
outputs=gr_temperature_hidden)
|
147 |
+
with gr.Column(scale=8):
|
148 |
+
with gr.Column(visible=True) as gr_examples_en:
|
149 |
+
gr.Examples(
|
150 |
+
examples=[
|
151 |
+
{
|
152 |
+
"text": "Who are you?",
|
153 |
+
"files": []
|
154 |
+
},
|
155 |
+
{
|
156 |
+
"text": "OCR and return markdown",
|
157 |
+
"files": ["examples/24-25-pl.png"]
|
158 |
+
},
|
159 |
+
{
|
160 |
+
"text":
|
161 |
+
"""describe the video""",
|
162 |
+
"files":
|
163 |
+
["examples/hitting_baseball.mp4"]
|
164 |
+
},
|
165 |
+
{
|
166 |
+
"text":
|
167 |
+
"For the model ranked first on WebSRC, what is its score on MathVision?",
|
168 |
+
"files": [
|
169 |
+
"examples/mimovl_gui.png",
|
170 |
+
"examples/mimovl_reason.png"
|
171 |
+
]
|
172 |
+
},
|
173 |
+
],
|
174 |
+
inputs=[gr_chatinterface_ofl.textbox],
|
175 |
+
)
|
176 |
+
with gr.Tab("Online") as gr_tab_ol:
|
177 |
+
with gr.Row():
|
178 |
+
with gr.Column(scale=1):
|
179 |
+
gr_infer_history = gr.State([])
|
180 |
+
gr_temperature_hidden = gr.Slider(minimum=0.0,
|
181 |
+
maximum=2.0,
|
182 |
+
step=0.1,
|
183 |
+
value=1.0,
|
184 |
+
interactive=True,
|
185 |
+
visible=False)
|
186 |
+
with gr.Row():
|
187 |
+
with gr.Column(scale=1):
|
188 |
+
gr_webcam_image = gr.Image(
|
189 |
+
label=label_translations['gr_webcam_image']
|
190 |
+
['English'],
|
191 |
+
sources="webcam",
|
192 |
+
height=250,
|
193 |
+
type='filepath')
|
194 |
+
gr_webcam_images = gr.Gallery(
|
195 |
+
label=label_translations['gr_webcam_images']
|
196 |
+
['English'],
|
197 |
+
show_label=True,
|
198 |
+
format='webp',
|
199 |
+
columns=1,
|
200 |
+
height=250,
|
201 |
+
preview=True,
|
202 |
+
interactive=False)
|
203 |
+
gr_counter = gr.Number(value=0, visible=False)
|
204 |
+
with gr.Column(scale=3):
|
205 |
+
gr_chatinterface_ol = gr.ChatInterface(
|
206 |
+
fn=online_record_chat,
|
207 |
+
type="messages",
|
208 |
+
multimodal=False,
|
209 |
+
chatbot=gr.Chatbot(height=800),
|
210 |
+
textbox=gr.
|
211 |
+
Textbox(placeholder=label_translations[
|
212 |
+
'gr_chatinterface_ol.textbox.placeholder']
|
213 |
+
['English'],
|
214 |
+
submit_btn=True,
|
215 |
+
stop_btn=True),
|
216 |
+
additional_inputs=[
|
217 |
+
gr_webcam_images, gr_counter,
|
218 |
+
gr_infer_history, gr_temperature_hidden
|
219 |
+
],
|
220 |
+
additional_outputs=[
|
221 |
+
gr_counter, gr_infer_history
|
222 |
+
],
|
223 |
+
)
|
224 |
|
225 |
+
def cache_webcam(recorded_image: str,
|
226 |
+
recorded_images: list):
|
227 |
+
if not recorded_images:
|
228 |
+
recorded_images = []
|
229 |
+
return recorded_images + [recorded_image]
|
230 |
+
|
231 |
+
gr_webcam_image.stream(
|
232 |
+
fn=cache_webcam,
|
233 |
+
inputs=[gr_webcam_image, gr_webcam_images],
|
234 |
+
outputs=[gr_webcam_images],
|
235 |
+
stream_every=1,
|
236 |
+
concurrency_limit=30,
|
237 |
+
)
|
238 |
+
with gr.Row():
|
239 |
+
gr_temperature_ol = gr.Slider(
|
240 |
+
minimum=0.0,
|
241 |
+
maximum=2.0,
|
242 |
+
step=0.1,
|
243 |
+
value=0.4,
|
244 |
+
label=label_translations['gr_temperature']
|
245 |
+
['English'],
|
246 |
+
interactive=True)
|
247 |
+
gr_temperature_ol.change(
|
248 |
+
lambda x: x,
|
249 |
+
inputs=gr_temperature_ol,
|
250 |
+
outputs=gr_temperature_hidden)
|
251 |
+
|
252 |
+
def update_lang(lang: str):
|
253 |
+
return (
|
254 |
+
gr.update(label=label_translations['gr_chatinterface_ofl'][lang]),
|
255 |
+
gr.update(label=label_translations['gr_chatinterface_ol'][lang]),
|
256 |
+
gr.update(placeholder=label_translations[
|
257 |
+
'gr_chatinterface_ofl.textbox.placeholder'][lang]),
|
258 |
+
gr.update(placeholder=label_translations[
|
259 |
+
'gr_chatinterface_ol.textbox.placeholder'][lang]),
|
260 |
+
gr.update(label=label_translations['gr_tab_ofl'][lang]),
|
261 |
+
gr.update(label=label_translations['gr_tab_ol'][lang]),
|
262 |
+
gr.update(label=label_translations['gr_temperature'][lang]),
|
263 |
+
gr.update(label=label_translations['gr_temperature'][lang]),
|
264 |
+
gr.update(visible=lang == 'English'),
|
265 |
+
gr.update(visible=lang != 'English'),
|
266 |
+
gr.update(label=label_translations['gr_webcam_image'][lang]),
|
267 |
+
gr.update(label=label_translations['gr_webcam_images'][lang]),
|
268 |
+
)
|
269 |
+
|
270 |
+
gr_lang_selector.change(fn=update_lang,
|
271 |
+
inputs=[gr_lang_selector],
|
272 |
+
outputs=[
|
273 |
+
gr_chatinterface_ofl.chatbot,
|
274 |
+
gr_chatinterface_ol.chatbot,
|
275 |
+
gr_chatinterface_ofl.textbox,
|
276 |
+
gr_chatinterface_ol.textbox,
|
277 |
+
gr_tab_ofl,
|
278 |
+
gr_tab_ol,
|
279 |
+
gr_temperature_ofl,
|
280 |
+
gr_temperature_ol,
|
281 |
+
gr_examples_en,
|
282 |
+
gr_webcam_image,
|
283 |
+
gr_webcam_images,
|
284 |
+
])
|
285 |
+
demo.queue(default_concurrency_limit=2, max_size=50)
|
286 |
+
|
287 |
+
if __name__ == "__main__":
|
288 |
+
demo.launch()
|
289 |
|
|
infer.py
CHANGED
@@ -1,4 +1,6 @@
|
|
1 |
-
# modified from https://github.com/
|
|
|
|
|
2 |
from transformers import AutoProcessor, Qwen2_5_VLForConditionalGeneration, TextIteratorStreamer
|
3 |
from transformers.generation.stopping_criteria import EosTokenCriteria, StoppingCriteriaList
|
4 |
from qwen_vl_utils import process_vision_info
|
@@ -6,67 +8,73 @@ from threading import Thread
|
|
6 |
|
7 |
|
8 |
class MiMoVLInfer:
|
9 |
-
def __init__(self, checkpoint_path,
|
|
|
10 |
self.model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
|
11 |
-
checkpoint_path,
|
12 |
-
|
13 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
14 |
|
15 |
def __call__(self, inputs: dict, history: list = [], temperature: float = 1.0):
|
16 |
messages = self.construct_messages(inputs)
|
17 |
updated_history = history + messages
|
18 |
text = self.processor.apply_chat_template(updated_history, tokenize=False, add_generation_prompt=True)
|
19 |
image_inputs, video_inputs = process_vision_info(updated_history)
|
|
|
20 |
model_inputs = self.processor(
|
21 |
text=[text], images=image_inputs, videos=video_inputs, padding=True, return_tensors='pt'
|
22 |
).to(self.model.device)
|
|
|
23 |
tokenizer = self.processor.tokenizer
|
24 |
-
streamer = TextIteratorStreamer(tokenizer, timeout=
|
|
|
|
|
25 |
gen_kwargs = {
|
26 |
-
'max_new_tokens':
|
|
|
|
|
|
|
27 |
'streamer': streamer,
|
28 |
'stopping_criteria': StoppingCriteriaList([EosTokenCriteria(eos_token_id=self.model.config.eos_token_id)]),
|
29 |
'pad_token_id': self.model.config.eos_token_id,
|
30 |
**model_inputs
|
31 |
}
|
32 |
-
|
|
|
33 |
thread.start()
|
34 |
partial_response = ""
|
35 |
for new_text in streamer:
|
36 |
partial_response += new_text
|
37 |
yield partial_response, updated_history + [{
|
38 |
'role': 'assistant',
|
39 |
-
'content': [{
|
40 |
-
'type': 'text',
|
41 |
-
'text': partial_response
|
42 |
-
}]
|
43 |
}]
|
44 |
|
45 |
def _is_video_file(self, filename):
|
46 |
-
|
47 |
-
|
48 |
|
49 |
def construct_messages(self, inputs: dict) -> list:
|
50 |
content = []
|
51 |
-
for
|
52 |
if self._is_video_file(path):
|
53 |
-
content.append({
|
54 |
-
"type": "video",
|
55 |
-
"video": f'file://{path}'
|
56 |
-
})
|
57 |
else:
|
58 |
-
content.append({
|
59 |
-
"type": "image",
|
60 |
-
"image": f'file://{path}'
|
61 |
-
})
|
62 |
query = inputs.get('text', '')
|
63 |
if query:
|
64 |
-
content.append({
|
65 |
-
|
66 |
-
"text": query,
|
67 |
-
})
|
68 |
-
messages = [{
|
69 |
-
"role": "user",
|
70 |
-
"content": content,
|
71 |
-
}]
|
72 |
-
return messages
|
|
|
1 |
+
# modified from https://github.com/XiaomiMiMo/MiMo-VL/tree/main/infer.py
|
2 |
+
import os
|
3 |
+
import torch
|
4 |
from transformers import AutoProcessor, Qwen2_5_VLForConditionalGeneration, TextIteratorStreamer
|
5 |
from transformers.generation.stopping_criteria import EosTokenCriteria, StoppingCriteriaList
|
6 |
from qwen_vl_utils import process_vision_info
|
|
|
8 |
|
9 |
|
10 |
class MiMoVLInfer:
|
11 |
+
def __init__(self, checkpoint_path, **kwargs):
|
12 |
+
dtype = torch.float16
|
13 |
self.model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
|
14 |
+
checkpoint_path,
|
15 |
+
torch_dtype=dtype,
|
16 |
+
device_map={"": "cpu"},
|
17 |
+
attn_implementation="eager",
|
18 |
+
trust_remote_code=True,
|
19 |
+
).eval()
|
20 |
+
self.processor = AutoProcessor.from_pretrained(checkpoint_path, trust_remote_code=True)
|
21 |
+
self._on_cuda = False
|
22 |
+
|
23 |
+
def to_device(self, device: str):
|
24 |
+
if device == "cuda" and not self._on_cuda:
|
25 |
+
self.model.to("cuda")
|
26 |
+
self._on_cuda = True
|
27 |
+
elif device == "cpu" and self._on_cuda:
|
28 |
+
self.model.to("cpu")
|
29 |
+
self._on_cuda = False
|
30 |
|
31 |
def __call__(self, inputs: dict, history: list = [], temperature: float = 1.0):
|
32 |
messages = self.construct_messages(inputs)
|
33 |
updated_history = history + messages
|
34 |
text = self.processor.apply_chat_template(updated_history, tokenize=False, add_generation_prompt=True)
|
35 |
image_inputs, video_inputs = process_vision_info(updated_history)
|
36 |
+
|
37 |
model_inputs = self.processor(
|
38 |
text=[text], images=image_inputs, videos=video_inputs, padding=True, return_tensors='pt'
|
39 |
).to(self.model.device)
|
40 |
+
|
41 |
tokenizer = self.processor.tokenizer
|
42 |
+
streamer = TextIteratorStreamer(tokenizer, timeout=60.0, skip_prompt=True, skip_special_tokens=True)
|
43 |
+
|
44 |
+
max_new = int(os.getenv("MAX_NEW_TOKENS", "1024"))
|
45 |
gen_kwargs = {
|
46 |
+
'max_new_tokens': max_new,
|
47 |
+
'do_sample': True,
|
48 |
+
'temperature': max(0.0, float(temperature)),
|
49 |
+
'top_p': 0.95,
|
50 |
'streamer': streamer,
|
51 |
'stopping_criteria': StoppingCriteriaList([EosTokenCriteria(eos_token_id=self.model.config.eos_token_id)]),
|
52 |
'pad_token_id': self.model.config.eos_token_id,
|
53 |
**model_inputs
|
54 |
}
|
55 |
+
|
56 |
+
thread = Thread(target=self.model.generate, kwargs=gen_kwargs, daemon=True)
|
57 |
thread.start()
|
58 |
partial_response = ""
|
59 |
for new_text in streamer:
|
60 |
partial_response += new_text
|
61 |
yield partial_response, updated_history + [{
|
62 |
'role': 'assistant',
|
63 |
+
'content': [{'type': 'text', 'text': partial_response}]
|
|
|
|
|
|
|
64 |
}]
|
65 |
|
66 |
def _is_video_file(self, filename):
|
67 |
+
return any(filename.lower().endswith(ext) for ext in
|
68 |
+
['.mp4', '.avi', '.mkv', '.mov', '.wmv', '.flv', '.webm', '.mpeg'])
|
69 |
|
70 |
def construct_messages(self, inputs: dict) -> list:
|
71 |
content = []
|
72 |
+
for path in inputs.get('files', []):
|
73 |
if self._is_video_file(path):
|
74 |
+
content.append({"type": "video", "video": f'file://{path}'})
|
|
|
|
|
|
|
75 |
else:
|
76 |
+
content.append({"type": "image", "image": f'file://{path}'})
|
|
|
|
|
|
|
77 |
query = inputs.get('text', '')
|
78 |
if query:
|
79 |
+
content.append({"type": "text", "text": query})
|
80 |
+
return [{"role": "user", "content": content}]
|
|
|
|
|
|
|
|
|
|
|
|
|
|