Spaces:
Running
on
Zero
Running
on
Zero
Update app.py
Browse files
app.py
CHANGED
@@ -1,107 +1,243 @@
|
|
1 |
-
import
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2 |
import spaces
|
3 |
import torch
|
4 |
-
import
|
5 |
-
from
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
6 |
|
7 |
-
|
8 |
-
|
9 |
-
|
10 |
-
|
11 |
-
parser.add_argument("--do_sample", action="store_true")
|
12 |
-
# This allows ignoring unrecognized arguments, e.g., from Jupyter
|
13 |
-
return parser.parse_known_args()
|
14 |
|
15 |
-
|
16 |
-
"""Load model and tokenizer from Hugging Face."""
|
17 |
-
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
18 |
-
model = AutoModelForCausalLM.from_pretrained(
|
19 |
-
model_name,
|
20 |
-
torch_dtype=torch.bfloat16,
|
21 |
-
device_map="auto"
|
22 |
-
)
|
23 |
-
return model, tokenizer
|
24 |
|
25 |
-
|
26 |
-
|
27 |
-
|
28 |
-
|
29 |
-
|
30 |
-
|
31 |
-
|
32 |
-
|
33 |
-
|
34 |
-
return tokenizer.decode(output_tokens[0], skip_special_tokens=True)
|
35 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
36 |
|
37 |
-
def
|
38 |
-
|
39 |
-
|
40 |
-
|
41 |
-
|
42 |
-
|
43 |
-
|
44 |
-
|
45 |
-
|
46 |
-
|
47 |
-
chat_history is a list of (user_message, bot_reply) pairs.
|
48 |
-
"""
|
49 |
-
# Build a single text prompt from the conversation so far
|
50 |
-
prompt = ""
|
51 |
-
for (old_user_msg, old_bot_msg) in chat_history:
|
52 |
-
prompt += f"User: {old_user_msg}\nBot: {old_bot_msg}\n"
|
53 |
-
# Add the new user query
|
54 |
-
prompt += f"User: {user_message}\nBot:"
|
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 |
if __name__ == "__main__":
|
107 |
-
|
|
|
1 |
+
import os
|
2 |
+
import random
|
3 |
+
import uuid
|
4 |
+
import json
|
5 |
+
import time
|
6 |
+
import asyncio
|
7 |
+
from threading import Thread
|
8 |
+
|
9 |
+
import gradio as gr
|
10 |
import spaces
|
11 |
import torch
|
12 |
+
import numpy as np
|
13 |
+
from PIL import Image
|
14 |
+
import cv2
|
15 |
+
|
16 |
+
from transformers import (
|
17 |
+
AutoModelForCausalLM,
|
18 |
+
AutoTokenizer,
|
19 |
+
TextIteratorStreamer,
|
20 |
+
Qwen2VLForConditionalGeneration,
|
21 |
+
AutoProcessor,
|
22 |
+
)
|
23 |
+
from transformers.image_utils import load_image
|
24 |
|
25 |
+
# Constants for text generation
|
26 |
+
MAX_MAX_NEW_TOKENS = 2048
|
27 |
+
DEFAULT_MAX_NEW_TOKENS = 1024
|
28 |
+
MAX_INPUT_TOKEN_LENGTH = int(os.getenv("MAX_INPUT_TOKEN_LENGTH", "4096"))
|
|
|
|
|
|
|
29 |
|
30 |
+
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
31 |
|
32 |
+
# Load text-only model and tokenizer
|
33 |
+
model_id = "prithivMLmods/FastThink-0.5B-Tiny"
|
34 |
+
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
35 |
+
model = AutoModelForCausalLM.from_pretrained(
|
36 |
+
model_id,
|
37 |
+
device_map="auto",
|
38 |
+
torch_dtype=torch.bfloat16,
|
39 |
+
)
|
40 |
+
model.eval()
|
|
|
41 |
|
42 |
+
MODEL_ID = "prithivMLmods/Qwen2-VL-OCR-2B-Instruct"
|
43 |
+
processor = AutoProcessor.from_pretrained(MODEL_ID, trust_remote_code=True)
|
44 |
+
model_m = Qwen2VLForConditionalGeneration.from_pretrained(
|
45 |
+
MODEL_ID,
|
46 |
+
trust_remote_code=True,
|
47 |
+
torch_dtype=torch.float16
|
48 |
+
).to("cuda").eval()
|
49 |
|
50 |
+
def clean_chat_history(chat_history):
|
51 |
+
"""
|
52 |
+
Filter out any chat entries whose "content" is not a string.
|
53 |
+
This helps prevent errors when concatenating previous messages.
|
54 |
+
"""
|
55 |
+
cleaned = []
|
56 |
+
for msg in chat_history:
|
57 |
+
if isinstance(msg, dict) and isinstance(msg.get("content"), str):
|
58 |
+
cleaned.append(msg)
|
59 |
+
return cleaned
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
60 |
|
61 |
+
def downsample_video(video_path):
|
62 |
+
"""
|
63 |
+
Downsamples the video to 10 evenly spaced frames.
|
64 |
+
Each frame is returned as a PIL image along with its timestamp.
|
65 |
+
"""
|
66 |
+
vidcap = cv2.VideoCapture(video_path)
|
67 |
+
total_frames = int(vidcap.get(cv2.CAP_PROP_FRAME_COUNT))
|
68 |
+
fps = vidcap.get(cv2.CAP_PROP_FPS)
|
69 |
+
frames = []
|
70 |
+
# Sample 10 evenly spaced frames.
|
71 |
+
frame_indices = np.linspace(0, total_frames - 1, 10, dtype=int)
|
72 |
+
for i in frame_indices:
|
73 |
+
vidcap.set(cv2.CAP_PROP_POS_FRAMES, i)
|
74 |
+
success, image = vidcap.read()
|
75 |
+
if success:
|
76 |
+
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB) # Convert BGR to RGB
|
77 |
+
pil_image = Image.fromarray(image)
|
78 |
+
timestamp = round(i / fps, 2)
|
79 |
+
frames.append((pil_image, timestamp))
|
80 |
+
vidcap.release()
|
81 |
+
return frames
|
82 |
|
83 |
+
def progress_bar_html(label: str) -> str:
|
84 |
+
"""
|
85 |
+
Returns an HTML snippet for a thin progress bar with a label.
|
86 |
+
The progress bar is styled as a dark red animated bar.
|
87 |
+
"""
|
88 |
+
return f'''
|
89 |
+
<div style="display: flex; align-items: center;">
|
90 |
+
<span style="margin-right: 10px; font-size: 14px;">{label}</span>
|
91 |
+
<div style="width: 110px; height: 5px; background-color: #FFF0F5; border-radius: 2px; overflow: hidden;">
|
92 |
+
<div style="width: 100%; height: 100%; background-color: #FF69B4; animation: loading 1.5s linear infinite;"></div>
|
93 |
+
</div>
|
94 |
+
</div>
|
95 |
+
<style>
|
96 |
+
@keyframes loading {{
|
97 |
+
0% {{ transform: translateX(-100%); }}
|
98 |
+
100% {{ transform: translateX(100%); }}
|
99 |
+
}}
|
100 |
+
</style>
|
101 |
+
'''
|
102 |
|
103 |
+
@spaces.GPU(duration=60, enable_queue=True)
|
104 |
+
def generate(input_dict: dict, chat_history: list[dict],
|
105 |
+
max_new_tokens: int = 1024,
|
106 |
+
temperature: float = 0.6,
|
107 |
+
top_p: float = 0.9,
|
108 |
+
top_k: int = 50,
|
109 |
+
repetition_penalty: float = 1.2):
|
110 |
+
"""
|
111 |
+
Generates chatbot responses with support for multimodal input and video processing.
|
112 |
+
Special command:
|
113 |
+
- "@video-infer": triggers video processing using Qwen2VL.
|
114 |
+
"""
|
115 |
+
text = input_dict["text"]
|
116 |
+
files = input_dict.get("files", [])
|
117 |
+
lower_text = text.strip().lower()
|
118 |
|
119 |
+
# Branch for video processing with Qwen2VL.
|
120 |
+
if lower_text.startswith("@video-infer"):
|
121 |
+
prompt = text[len("@video-infer"):].strip()
|
122 |
+
if files:
|
123 |
+
# Assume the first file is a video.
|
124 |
+
video_path = files[0]
|
125 |
+
frames = downsample_video(video_path)
|
126 |
+
messages = [
|
127 |
+
{"role": "system", "content": [{"type": "text", "text": "You are a helpful assistant."}]},
|
128 |
+
{"role": "user", "content": [{"type": "text", "text": prompt}]}
|
129 |
+
]
|
130 |
+
# Append each frame with its timestamp.
|
131 |
+
for frame in frames:
|
132 |
+
image, timestamp = frame
|
133 |
+
image_path = f"video_frame_{uuid.uuid4().hex}.png"
|
134 |
+
image.save(image_path)
|
135 |
+
messages[1]["content"].append({"type": "text", "text": f"Frame {timestamp}:"})
|
136 |
+
messages[1]["content"].append({"type": "image", "url": image_path})
|
137 |
+
else:
|
138 |
+
messages = [
|
139 |
+
{"role": "system", "content": [{"type": "text", "text": "You are a helpful assistant."}]},
|
140 |
+
{"role": "user", "content": [{"type": "text", "text": prompt}]}
|
141 |
+
]
|
142 |
+
inputs = processor.apply_chat_template(
|
143 |
+
messages, tokenize=True, add_generation_prompt=True, return_dict=True, return_tensors="pt"
|
144 |
+
).to("cuda")
|
145 |
+
streamer = TextIteratorStreamer(processor, skip_prompt=True, skip_special_tokens=True)
|
146 |
+
generation_kwargs = {
|
147 |
+
**inputs,
|
148 |
+
"streamer": streamer,
|
149 |
+
"max_new_tokens": max_new_tokens,
|
150 |
+
"do_sample": True,
|
151 |
+
"temperature": temperature,
|
152 |
+
"top_p": top_p,
|
153 |
+
"top_k": top_k,
|
154 |
+
"repetition_penalty": repetition_penalty,
|
155 |
+
}
|
156 |
+
thread = Thread(target=model_m.generate, kwargs=generation_kwargs)
|
157 |
+
thread.start()
|
158 |
+
buffer = ""
|
159 |
+
yield progress_bar_html("Processing video with Qwen2VL")
|
160 |
+
for new_text in streamer:
|
161 |
+
buffer += new_text
|
162 |
+
buffer = buffer.replace("<|im_end|>", "")
|
163 |
+
time.sleep(0.01)
|
164 |
+
yield buffer
|
165 |
+
return
|
166 |
|
167 |
+
# Normal text or multimodal conversation processing.
|
168 |
+
if files:
|
169 |
+
if len(files) > 1:
|
170 |
+
images = [load_image(image) for image in files]
|
171 |
+
elif len(files) == 1:
|
172 |
+
images = [load_image(files[0])]
|
173 |
+
else:
|
174 |
+
images = []
|
175 |
+
messages = [{
|
176 |
+
"role": "user",
|
177 |
+
"content": [
|
178 |
+
*[{"type": "image", "image": image} for image in images],
|
179 |
+
{"type": "text", "text": text},
|
180 |
+
]
|
181 |
+
}]
|
182 |
+
prompt_full = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
|
183 |
+
inputs = processor(text=[prompt_full], images=images, return_tensors="pt", padding=True).to("cuda")
|
184 |
+
streamer = TextIteratorStreamer(processor, skip_prompt=True, skip_special_tokens=True)
|
185 |
+
generation_kwargs = {**inputs, "streamer": streamer, "max_new_tokens": max_new_tokens}
|
186 |
+
thread = Thread(target=model_m.generate, kwargs=generation_kwargs)
|
187 |
+
thread.start()
|
188 |
+
buffer = ""
|
189 |
+
yield progress_bar_html("Thinking...")
|
190 |
+
for new_text in streamer:
|
191 |
+
buffer += new_text
|
192 |
+
buffer = buffer.replace("<|im_end|>", "")
|
193 |
+
time.sleep(0.01)
|
194 |
+
yield buffer
|
195 |
+
else:
|
196 |
+
conversation = clean_chat_history(chat_history)
|
197 |
+
conversation.append({"role": "user", "content": text})
|
198 |
+
input_ids = tokenizer.apply_chat_template(conversation, add_generation_prompt=True, return_tensors="pt")
|
199 |
+
if input_ids.shape[1] > MAX_INPUT_TOKEN_LENGTH:
|
200 |
+
input_ids = input_ids[:, -MAX_INPUT_TOKEN_LENGTH:]
|
201 |
+
gr.Warning(f"Trimmed input from conversation as it was longer than {MAX_INPUT_TOKEN_LENGTH} tokens.")
|
202 |
+
input_ids = input_ids.to(model.device)
|
203 |
+
streamer = TextIteratorStreamer(tokenizer, timeout=20.0, skip_prompt=True, skip_special_tokens=True)
|
204 |
+
generation_kwargs = {
|
205 |
+
"input_ids": input_ids,
|
206 |
+
"streamer": streamer,
|
207 |
+
"max_new_tokens": max_new_tokens,
|
208 |
+
"do_sample": True,
|
209 |
+
"top_p": top_p,
|
210 |
+
"top_k": top_k,
|
211 |
+
"temperature": temperature,
|
212 |
+
"num_beams": 1,
|
213 |
+
"repetition_penalty": repetition_penalty,
|
214 |
+
}
|
215 |
+
t = Thread(target=model.generate, kwargs=generation_kwargs)
|
216 |
+
t.start()
|
217 |
+
outputs = []
|
218 |
+
yield progress_bar_html("Processing...")
|
219 |
+
for new_text in streamer:
|
220 |
+
outputs.append(new_text)
|
221 |
+
yield "".join(outputs)
|
222 |
+
final_response = "".join(outputs)
|
223 |
+
yield final_response
|
224 |
|
225 |
+
demo = gr.ChatInterface(
|
226 |
+
fn=generate,
|
227 |
+
additional_inputs=[
|
228 |
+
gr.Slider(label="Max new tokens", minimum=1, maximum=MAX_MAX_NEW_TOKENS, step=1, value=DEFAULT_MAX_NEW_TOKENS),
|
229 |
+
gr.Slider(label="Temperature", minimum=0.1, maximum=4.0, step=0.1, value=0.6),
|
230 |
+
gr.Slider(label="Top-p (nucleus sampling)", minimum=0.05, maximum=1.0, step=0.05, value=0.9),
|
231 |
+
gr.Slider(label="Top-k", minimum=1, maximum=1000, step=1, value=50),
|
232 |
+
gr.Slider(label="Repetition penalty", minimum=1.0, maximum=2.0, step=0.05, value=1.2),
|
233 |
+
],
|
234 |
+
cache_examples=False,
|
235 |
+
type="messages",
|
236 |
+
fill_height=True,
|
237 |
+
textbox=gr.MultimodalTextbox(label="Query Input", file_types=["image", "video"], file_count="multiple"),
|
238 |
+
stop_btn="Stop Generation",
|
239 |
+
multimodal=True,
|
240 |
+
)
|
241 |
|
242 |
if __name__ == "__main__":
|
243 |
+
demo.queue(max_size=20).launch(share=True)
|