Spaces:
Running
on
Zero
Running
on
Zero
Update app.py
Browse files
app.py
CHANGED
@@ -23,9 +23,9 @@ from transformers import (
|
|
23 |
from transformers.image_utils import load_image
|
24 |
from diffusers import StableDiffusionXLPipeline, EulerAncestralDiscreteScheduler
|
25 |
|
26 |
-
|
27 |
DESCRIPTION = """
|
28 |
# QwQ Edge 💬
|
|
|
29 |
"""
|
30 |
|
31 |
css = '''
|
@@ -40,6 +40,34 @@ h1 {
|
|
40 |
background: #1565c0;
|
41 |
border-radius: 100vh;
|
42 |
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
43 |
'''
|
44 |
|
45 |
MAX_MAX_NEW_TOKENS = 2048
|
@@ -63,7 +91,7 @@ TTS_VOICES = [
|
|
63 |
"en-US-GuyNeural", # @tts2
|
64 |
]
|
65 |
|
66 |
-
MODEL_ID = "prithivMLmods/Qwen2-VL-OCR-2B-Instruct"
|
67 |
processor = AutoProcessor.from_pretrained(MODEL_ID, trust_remote_code=True)
|
68 |
model_m = Qwen2VLForConditionalGeneration.from_pretrained(
|
69 |
MODEL_ID,
|
@@ -78,24 +106,20 @@ async def text_to_speech(text: str, voice: str, output_file="output.mp3"):
|
|
78 |
return output_file
|
79 |
|
80 |
def clean_chat_history(chat_history):
|
81 |
-
"""
|
82 |
-
Filter out any chat entries whose "content" is not a string.
|
83 |
-
This helps prevent errors when concatenating previous messages.
|
84 |
-
"""
|
85 |
cleaned = []
|
86 |
for msg in chat_history:
|
87 |
if isinstance(msg, dict) and isinstance(msg.get("content"), str):
|
88 |
cleaned.append(msg)
|
89 |
return cleaned
|
90 |
|
91 |
-
#
|
92 |
-
MODEL_ID_SD = os.getenv("MODEL_VAL_PATH")
|
93 |
MAX_IMAGE_SIZE = int(os.getenv("MAX_IMAGE_SIZE", "4096"))
|
94 |
USE_TORCH_COMPILE = os.getenv("USE_TORCH_COMPILE", "0") == "1"
|
95 |
ENABLE_CPU_OFFLOAD = os.getenv("ENABLE_CPU_OFFLOAD", "0") == "1"
|
96 |
-
BATCH_SIZE = int(os.getenv("BATCH_SIZE", "1"))
|
97 |
|
98 |
-
# Load the SDXL pipeline
|
99 |
sd_pipe = StableDiffusionXLPipeline.from_pretrained(
|
100 |
MODEL_ID_SD,
|
101 |
torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
|
@@ -104,22 +128,19 @@ sd_pipe = StableDiffusionXLPipeline.from_pretrained(
|
|
104 |
).to(device)
|
105 |
sd_pipe.scheduler = EulerAncestralDiscreteScheduler.from_config(sd_pipe.scheduler.config)
|
106 |
|
107 |
-
# Ensure that the text encoder is in half-precision if using CUDA.
|
108 |
if torch.cuda.is_available():
|
109 |
sd_pipe.text_encoder = sd_pipe.text_encoder.half()
|
110 |
|
111 |
-
# Optional: compile the model for speedup if enabled
|
112 |
if USE_TORCH_COMPILE:
|
113 |
sd_pipe.compile()
|
114 |
|
115 |
-
# Optional: offload parts of the model to CPU if needed
|
116 |
if ENABLE_CPU_OFFLOAD:
|
117 |
sd_pipe.enable_model_cpu_offload()
|
118 |
|
119 |
MAX_SEED = np.iinfo(np.int32).max
|
120 |
|
121 |
def save_image(img: Image.Image) -> str:
|
122 |
-
"""Save a PIL image with a unique filename and return the path
|
123 |
unique_name = str(uuid.uuid4()) + ".png"
|
124 |
img.save(unique_name)
|
125 |
return unique_name
|
@@ -144,7 +165,7 @@ def generate_image_fn(
|
|
144 |
num_images: int = 1,
|
145 |
progress=gr.Progress(track_tqdm=True),
|
146 |
):
|
147 |
-
"""Generate images using the SDXL pipeline
|
148 |
seed = int(randomize_seed_fn(seed, randomize_seed))
|
149 |
generator = torch.Generator(device=device).manual_seed(seed)
|
150 |
|
@@ -162,13 +183,11 @@ def generate_image_fn(
|
|
162 |
options["use_resolution_binning"] = True
|
163 |
|
164 |
images = []
|
165 |
-
# Process in batches
|
166 |
for i in range(0, num_images, BATCH_SIZE):
|
167 |
batch_options = options.copy()
|
168 |
batch_options["prompt"] = options["prompt"][i:i+BATCH_SIZE]
|
169 |
if "negative_prompt" in batch_options and batch_options["negative_prompt"] is not None:
|
170 |
batch_options["negative_prompt"] = options["negative_prompt"][i:i+BATCH_SIZE]
|
171 |
-
# Wrap the pipeline call in autocast if using CUDA
|
172 |
if device.type == "cuda":
|
173 |
with torch.autocast("cuda", dtype=torch.float16):
|
174 |
outputs = sd_pipe(**batch_options)
|
@@ -197,35 +216,14 @@ def generate(
|
|
197 |
text = input_dict["text"]
|
198 |
files = input_dict.get("files", [])
|
199 |
|
200 |
-
# Define an HTML template for the animated progress bar.
|
201 |
-
# The bar is a thin 5px line in light green with a simple opacity animation.
|
202 |
-
progress_bar_html = """
|
203 |
-
<div style="display: flex; align-items: center;">
|
204 |
-
<span>{message}</span>
|
205 |
-
<div style="flex-grow: 1; margin-left: 10px;">
|
206 |
-
<div class="progress-bar"></div>
|
207 |
-
</div>
|
208 |
-
</div>
|
209 |
-
<style>
|
210 |
-
.progress-bar {{
|
211 |
-
width: 100%;
|
212 |
-
height: 5px;
|
213 |
-
background: lightgreen;
|
214 |
-
animation: progressAnim 2s infinite;
|
215 |
-
}}
|
216 |
-
@keyframes progressAnim {{
|
217 |
-
0% {{ opacity: 0.5; }}
|
218 |
-
50% {{ opacity: 1; }}
|
219 |
-
100% {{ opacity: 0.5; }}
|
220 |
-
}}
|
221 |
-
</style>
|
222 |
-
"""
|
223 |
-
|
224 |
if text.strip().lower().startswith("@image"):
|
225 |
-
# Remove the "@image" tag and use the rest as prompt.
|
226 |
prompt = text[len("@image"):].strip()
|
227 |
-
#
|
228 |
-
yield gr.HTML(
|
|
|
|
|
|
|
|
|
229 |
image_paths, used_seed = generate_image_fn(
|
230 |
prompt=prompt,
|
231 |
negative_prompt="",
|
@@ -239,9 +237,9 @@ def generate(
|
|
239 |
use_resolution_binning=True,
|
240 |
num_images=1,
|
241 |
)
|
242 |
-
#
|
243 |
yield gr.Image(image_paths[0])
|
244 |
-
return
|
245 |
|
246 |
tts_prefix = "@tts"
|
247 |
is_tts = any(text.strip().lower().startswith(f"{tts_prefix}{i}") for i in range(1, 3))
|
@@ -250,11 +248,9 @@ def generate(
|
|
250 |
if is_tts and voice_index:
|
251 |
voice = TTS_VOICES[voice_index - 1]
|
252 |
text = text.replace(f"{tts_prefix}{voice_index}", "").strip()
|
253 |
-
# Clear previous chat history for a fresh TTS request.
|
254 |
conversation = [{"role": "user", "content": text}]
|
255 |
else:
|
256 |
voice = None
|
257 |
-
# Remove any stray @tts tags and build the conversation history.
|
258 |
text = text.replace(tts_prefix, "").strip()
|
259 |
conversation = clean_chat_history(chat_history)
|
260 |
conversation.append({"role": "user", "content": text})
|
@@ -280,21 +276,18 @@ def generate(
|
|
280 |
thread = Thread(target=model_m.generate, kwargs=generation_kwargs)
|
281 |
thread.start()
|
282 |
|
283 |
-
#
|
284 |
-
yield gr.HTML(
|
|
|
|
|
|
|
285 |
buffer = ""
|
286 |
for new_text in streamer:
|
287 |
buffer += new_text
|
288 |
buffer = buffer.replace("<|im_end|>", "")
|
289 |
time.sleep(0.01)
|
290 |
-
#
|
291 |
-
|
292 |
-
<div style="display: flex; flex-direction: column;">
|
293 |
-
{progress_bar_html.format(message="Thinking...")}
|
294 |
-
<div style="margin-top: 10px;">{buffer}</div>
|
295 |
-
</div>
|
296 |
-
"""
|
297 |
-
yield gr.HTML(combined_html)
|
298 |
else:
|
299 |
input_ids = tokenizer.apply_chat_template(conversation, add_generation_prompt=True, return_tensors="pt")
|
300 |
if input_ids.shape[1] > MAX_INPUT_TOKEN_LENGTH:
|
@@ -316,23 +309,18 @@ def generate(
|
|
316 |
t = Thread(target=model.generate, kwargs=generation_kwargs)
|
317 |
t.start()
|
318 |
|
319 |
-
#
|
320 |
-
yield gr.HTML(
|
321 |
-
|
|
|
|
|
|
|
322 |
for new_text in streamer:
|
323 |
-
|
324 |
-
|
325 |
-
|
326 |
-
{progress_bar_html.format(message="Thinking...")}
|
327 |
-
<div style="margin-top: 10px;">{''.join(outputs)}</div>
|
328 |
-
</div>
|
329 |
-
"""
|
330 |
-
yield gr.HTML(combined_html)
|
331 |
-
final_response = "".join(outputs)
|
332 |
-
# Final response: progress bar is removed and only the generated text is shown.
|
333 |
-
yield final_response
|
334 |
|
335 |
-
|
336 |
if is_tts and voice:
|
337 |
output_file = asyncio.run(text_to_speech(final_response, voice))
|
338 |
yield gr.Audio(output_file, autoplay=True)
|
|
|
23 |
from transformers.image_utils import load_image
|
24 |
from diffusers import StableDiffusionXLPipeline, EulerAncestralDiscreteScheduler
|
25 |
|
|
|
26 |
DESCRIPTION = """
|
27 |
# QwQ Edge 💬
|
28 |
+
**Note:** During image generation, a progress bar will appear both at the top of the interface and within the chat. For text generation, a loading animation will display until the response begins.
|
29 |
"""
|
30 |
|
31 |
css = '''
|
|
|
40 |
background: #1565c0;
|
41 |
border-radius: 100vh;
|
42 |
}
|
43 |
+
|
44 |
+
/* Custom styling for progress bars within chat */
|
45 |
+
.progress-bar-container {
|
46 |
+
width: 100%;
|
47 |
+
margin-top: 5px;
|
48 |
+
}
|
49 |
+
|
50 |
+
.progress-bar {
|
51 |
+
width: 100%;
|
52 |
+
height: 4px;
|
53 |
+
background-color: #e0e0e0;
|
54 |
+
border-radius: 2px;
|
55 |
+
}
|
56 |
+
|
57 |
+
.progress-bar::-webkit-progress-bar {
|
58 |
+
background-color: #e0e0e0;
|
59 |
+
border-radius: 2px;
|
60 |
+
}
|
61 |
+
|
62 |
+
.progress-bar::-webkit-progress-value {
|
63 |
+
background-color: #90ee90; /* Light green */
|
64 |
+
border-radius: 2px;
|
65 |
+
}
|
66 |
+
|
67 |
+
.progress-bar::-moz-progress-bar {
|
68 |
+
background-color: #90ee90; /* Light green */
|
69 |
+
border-radius: 2px;
|
70 |
+
}
|
71 |
'''
|
72 |
|
73 |
MAX_MAX_NEW_TOKENS = 2048
|
|
|
91 |
"en-US-GuyNeural", # @tts2
|
92 |
]
|
93 |
|
94 |
+
MODEL_ID = "prithivMLmods/Qwen2-VL-OCR-2B-Instruct"
|
95 |
processor = AutoProcessor.from_pretrained(MODEL_ID, trust_remote_code=True)
|
96 |
model_m = Qwen2VLForConditionalGeneration.from_pretrained(
|
97 |
MODEL_ID,
|
|
|
106 |
return output_file
|
107 |
|
108 |
def clean_chat_history(chat_history):
|
109 |
+
"""Filter out non-string content to prevent concatenation errors"""
|
|
|
|
|
|
|
110 |
cleaned = []
|
111 |
for msg in chat_history:
|
112 |
if isinstance(msg, dict) and isinstance(msg.get("content"), str):
|
113 |
cleaned.append(msg)
|
114 |
return cleaned
|
115 |
|
116 |
+
# Stable Diffusion XL setup
|
117 |
+
MODEL_ID_SD = os.getenv("MODEL_VAL_PATH")
|
118 |
MAX_IMAGE_SIZE = int(os.getenv("MAX_IMAGE_SIZE", "4096"))
|
119 |
USE_TORCH_COMPILE = os.getenv("USE_TORCH_COMPILE", "0") == "1"
|
120 |
ENABLE_CPU_OFFLOAD = os.getenv("ENABLE_CPU_OFFLOAD", "0") == "1"
|
121 |
+
BATCH_SIZE = int(os.getenv("BATCH_SIZE", "1"))
|
122 |
|
|
|
123 |
sd_pipe = StableDiffusionXLPipeline.from_pretrained(
|
124 |
MODEL_ID_SD,
|
125 |
torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
|
|
|
128 |
).to(device)
|
129 |
sd_pipe.scheduler = EulerAncestralDiscreteScheduler.from_config(sd_pipe.scheduler.config)
|
130 |
|
|
|
131 |
if torch.cuda.is_available():
|
132 |
sd_pipe.text_encoder = sd_pipe.text_encoder.half()
|
133 |
|
|
|
134 |
if USE_TORCH_COMPILE:
|
135 |
sd_pipe.compile()
|
136 |
|
|
|
137 |
if ENABLE_CPU_OFFLOAD:
|
138 |
sd_pipe.enable_model_cpu_offload()
|
139 |
|
140 |
MAX_SEED = np.iinfo(np.int32).max
|
141 |
|
142 |
def save_image(img: Image.Image) -> str:
|
143 |
+
"""Save a PIL image with a unique filename and return the path"""
|
144 |
unique_name = str(uuid.uuid4()) + ".png"
|
145 |
img.save(unique_name)
|
146 |
return unique_name
|
|
|
165 |
num_images: int = 1,
|
166 |
progress=gr.Progress(track_tqdm=True),
|
167 |
):
|
168 |
+
"""Generate images using the SDXL pipeline"""
|
169 |
seed = int(randomize_seed_fn(seed, randomize_seed))
|
170 |
generator = torch.Generator(device=device).manual_seed(seed)
|
171 |
|
|
|
183 |
options["use_resolution_binning"] = True
|
184 |
|
185 |
images = []
|
|
|
186 |
for i in range(0, num_images, BATCH_SIZE):
|
187 |
batch_options = options.copy()
|
188 |
batch_options["prompt"] = options["prompt"][i:i+BATCH_SIZE]
|
189 |
if "negative_prompt" in batch_options and batch_options["negative_prompt"] is not None:
|
190 |
batch_options["negative_prompt"] = options["negative_prompt"][i:i+BATCH_SIZE]
|
|
|
191 |
if device.type == "cuda":
|
192 |
with torch.autocast("cuda", dtype=torch.float16):
|
193 |
outputs = sd_pipe(**batch_options)
|
|
|
216 |
text = input_dict["text"]
|
217 |
files = input_dict.get("files", [])
|
218 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
219 |
if text.strip().lower().startswith("@image"):
|
|
|
220 |
prompt = text[len("@image"):].strip()
|
221 |
+
# Initial message with progress bar at 0%
|
222 |
+
yield gr.HTML(
|
223 |
+
'<div>Generating Image...</div>'
|
224 |
+
'<progress class="progress-bar" value="0" max="100" '
|
225 |
+
'style="width:100%; height:4px; background-color:#e0e0e0;"></progress>'
|
226 |
+
)
|
227 |
image_paths, used_seed = generate_image_fn(
|
228 |
prompt=prompt,
|
229 |
negative_prompt="",
|
|
|
237 |
use_resolution_binning=True,
|
238 |
num_images=1,
|
239 |
)
|
240 |
+
# Final message with the image, progress bar at 100%
|
241 |
yield gr.Image(image_paths[0])
|
242 |
+
return
|
243 |
|
244 |
tts_prefix = "@tts"
|
245 |
is_tts = any(text.strip().lower().startswith(f"{tts_prefix}{i}") for i in range(1, 3))
|
|
|
248 |
if is_tts and voice_index:
|
249 |
voice = TTS_VOICES[voice_index - 1]
|
250 |
text = text.replace(f"{tts_prefix}{voice_index}", "").strip()
|
|
|
251 |
conversation = [{"role": "user", "content": text}]
|
252 |
else:
|
253 |
voice = None
|
|
|
254 |
text = text.replace(tts_prefix, "").strip()
|
255 |
conversation = clean_chat_history(chat_history)
|
256 |
conversation.append({"role": "user", "content": text})
|
|
|
276 |
thread = Thread(target=model_m.generate, kwargs=generation_kwargs)
|
277 |
thread.start()
|
278 |
|
279 |
+
# Initial loading bar (indeterminate animation via CSS)
|
280 |
+
yield gr.HTML(
|
281 |
+
'<div>Generating response...</div>'
|
282 |
+
'<progress class="progress-bar" style="width:100%; height:4px; background-color:#e0e0e0;"></progress>'
|
283 |
+
)
|
284 |
buffer = ""
|
285 |
for new_text in streamer:
|
286 |
buffer += new_text
|
287 |
buffer = buffer.replace("<|im_end|>", "")
|
288 |
time.sleep(0.01)
|
289 |
+
# Yield only the text, replacing the loading bar
|
290 |
+
yield buffer
|
|
|
|
|
|
|
|
|
|
|
|
|
291 |
else:
|
292 |
input_ids = tokenizer.apply_chat_template(conversation, add_generation_prompt=True, return_tensors="pt")
|
293 |
if input_ids.shape[1] > MAX_INPUT_TOKEN_LENGTH:
|
|
|
309 |
t = Thread(target=model.generate, kwargs=generation_kwargs)
|
310 |
t.start()
|
311 |
|
312 |
+
# Initial loading bar
|
313 |
+
yield gr.HTML(
|
314 |
+
'<div>Generating response...</div>'
|
315 |
+
'<progress class="progress-bar" style="width:100%; height:4px; background-color:#e0e0e0;"></progress>'
|
316 |
+
)
|
317 |
+
buffer = ""
|
318 |
for new_text in streamer:
|
319 |
+
buffer += new_text
|
320 |
+
# Yield only the text, replacing the loading bar
|
321 |
+
yield buffer
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
322 |
|
323 |
+
final_response = buffer
|
324 |
if is_tts and voice:
|
325 |
output_file = asyncio.run(text_to_speech(final_response, voice))
|
326 |
yield gr.Audio(output_file, autoplay=True)
|