Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
|
@@ -3,8 +3,8 @@ import subprocess
|
|
| 3 |
from huggingface_hub import hf_hub_download
|
| 4 |
|
| 5 |
# 1. Install llama-cpp-python in runtime (not via requirements.txt)
|
| 6 |
-
# Important: remove `llama-cpp-python` from requirements.txt
|
| 7 |
-
#
|
| 8 |
subprocess.run("pip install -q 'llama_cpp_python==0.3.15'", shell=True, check=False)
|
| 9 |
from llama_cpp import Llama
|
| 10 |
|
|
@@ -52,7 +52,7 @@ def _extract_text_from_content(content):
|
|
| 52 |
In Gradio 6 ChatInterface, history uses the messages format.
|
| 53 |
content can be:
|
| 54 |
- a string
|
| 55 |
-
- a list of blocks: [{
|
| 56 |
We convert it into a simple string.
|
| 57 |
"""
|
| 58 |
if isinstance(content, list):
|
|
@@ -89,7 +89,7 @@ def build_prompt(message, history, style):
|
|
| 89 |
prompt_parts.append(f"System: {system_prompt}\n")
|
| 90 |
prompt_parts.append("Conversation:\n")
|
| 91 |
|
| 92 |
-
# history is a list of dicts: {
|
| 93 |
for msg in history or []:
|
| 94 |
role = msg.get("role")
|
| 95 |
content = _extract_text_from_content(msg.get("content", ""))
|
|
@@ -187,10 +187,9 @@ style_radio = gr.Radio(
|
|
| 187 |
label="Answer style",
|
| 188 |
)
|
| 189 |
|
| 190 |
-
# 5. Christmas theme CSS
|
| 191 |
-
|
| 192 |
-
|
| 193 |
-
christmas_css = """
|
| 194 |
body {
|
| 195 |
background: radial-gradient(circle at top, #1b1c2b 0, #050611 55%, #000000 100%);
|
| 196 |
color: #fdf6e3;
|
|
@@ -368,53 +367,43 @@ input[type="range"] {
|
|
| 368 |
background: rgba(255, 255, 255, 0.5);
|
| 369 |
border-radius: 999px;
|
| 370 |
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 371 |
"""
|
| 372 |
|
| 373 |
|
| 374 |
-
|
| 375 |
-
|
| 376 |
-
|
| 377 |
-
|
| 378 |
-
|
| 379 |
-
|
| 380 |
-
|
| 381 |
-
|
| 382 |
-
|
| 383 |
-
|
| 384 |
-
|
| 385 |
-
|
| 386 |
-
|
| 387 |
-
</p>
|
| 388 |
-
</div>
|
| 389 |
-
</div>
|
| 390 |
-
<div class="hero-keyline"></div>
|
| 391 |
-
</div>
|
| 392 |
-
"""
|
| 393 |
-
)
|
| 394 |
-
|
| 395 |
-
gr.Markdown(
|
| 396 |
-
"π **Tip:** Try switching between short answers and step by step reasoning, "
|
| 397 |
-
"and play with temperature and top p to see how the model behaves."
|
| 398 |
-
)
|
| 399 |
-
|
| 400 |
-
gr.ChatInterface(
|
| 401 |
-
fn=chat_fn,
|
| 402 |
-
title="Lab 2 β Fine tuned GGUF model",
|
| 403 |
-
description=(
|
| 404 |
-
"Chat with our fine tuned Llama based model, converted to GGUF and "
|
| 405 |
-
"loaded via llama.cpp from `Jeppcode/ScalableLab2`.\n\n"
|
| 406 |
-
"Use the controls in the accordion below like a DJ board to tweak "
|
| 407 |
-
"response length, randomness and style."
|
| 408 |
-
),
|
| 409 |
-
additional_inputs=[
|
| 410 |
-
max_new_tokens_slider,
|
| 411 |
-
temperature_slider,
|
| 412 |
-
top_p_slider,
|
| 413 |
-
repetition_penalty_slider,
|
| 414 |
-
style_radio,
|
| 415 |
-
],
|
| 416 |
-
additional_inputs_accordion="Generation controls",
|
| 417 |
-
)
|
| 418 |
|
| 419 |
if __name__ == "__main__":
|
| 420 |
demo.launch()
|
|
|
|
| 3 |
from huggingface_hub import hf_hub_download
|
| 4 |
|
| 5 |
# 1. Install llama-cpp-python in runtime (not via requirements.txt)
|
| 6 |
+
# Important: remove `llama-cpp-python` from requirements.txt,
|
| 7 |
+
# otherwise Spaces may try to build from source and get stuck.
|
| 8 |
subprocess.run("pip install -q 'llama_cpp_python==0.3.15'", shell=True, check=False)
|
| 9 |
from llama_cpp import Llama
|
| 10 |
|
|
|
|
| 52 |
In Gradio 6 ChatInterface, history uses the messages format.
|
| 53 |
content can be:
|
| 54 |
- a string
|
| 55 |
+
- a list of blocks: [{'type': 'text', 'text': '...'} , ...]
|
| 56 |
We convert it into a simple string.
|
| 57 |
"""
|
| 58 |
if isinstance(content, list):
|
|
|
|
| 89 |
prompt_parts.append(f"System: {system_prompt}\n")
|
| 90 |
prompt_parts.append("Conversation:\n")
|
| 91 |
|
| 92 |
+
# history is a list of dicts: {'role': 'user'/'assistant'/'system', 'content': ...}
|
| 93 |
for msg in history or []:
|
| 94 |
role = msg.get("role")
|
| 95 |
content = _extract_text_from_content(msg.get("content", ""))
|
|
|
|
| 187 |
label="Answer style",
|
| 188 |
)
|
| 189 |
|
| 190 |
+
# 5. Christmas theme: inject CSS + hero directly into description
|
| 191 |
+
christmas_style_and_hero = """
|
| 192 |
+
<style>
|
|
|
|
| 193 |
body {
|
| 194 |
background: radial-gradient(circle at top, #1b1c2b 0, #050611 55%, #000000 100%);
|
| 195 |
color: #fdf6e3;
|
|
|
|
| 367 |
background: rgba(255, 255, 255, 0.5);
|
| 368 |
border-radius: 999px;
|
| 369 |
}
|
| 370 |
+
</style>
|
| 371 |
+
|
| 372 |
+
<div class="hero">
|
| 373 |
+
<div class="hero-inner">
|
| 374 |
+
<div class="hero-badge"></div>
|
| 375 |
+
<div class="hero-text">
|
| 376 |
+
<h1>Scalable Lab 2 Christmas Chat</h1>
|
| 377 |
+
<p>
|
| 378 |
+
Talk to our fine tuned Llama based model, wrapped as a compact GGUF
|
| 379 |
+
and running on CPU. Use the controls in the accordion below to tune
|
| 380 |
+
response length, randomness and style like a Christmas DJ for language models.
|
| 381 |
+
</p>
|
| 382 |
+
</div>
|
| 383 |
+
</div>
|
| 384 |
+
<div class="hero-keyline"></div>
|
| 385 |
+
</div>
|
| 386 |
+
|
| 387 |
+
<p>
|
| 388 |
+
π <strong>Tip:</strong> Try switching between short answers and step by step reasoning,
|
| 389 |
+
and play with temperature and top p to see how the model behaves.
|
| 390 |
+
</p>
|
| 391 |
"""
|
| 392 |
|
| 393 |
|
| 394 |
+
demo = gr.ChatInterface(
|
| 395 |
+
fn=chat_fn,
|
| 396 |
+
title="Lab 2 β Fine-tuned GGUF model",
|
| 397 |
+
description=christmas_style_and_hero,
|
| 398 |
+
additional_inputs=[
|
| 399 |
+
max_new_tokens_slider,
|
| 400 |
+
temperature_slider,
|
| 401 |
+
top_p_slider,
|
| 402 |
+
repetition_penalty_slider,
|
| 403 |
+
style_radio,
|
| 404 |
+
],
|
| 405 |
+
additional_inputs_accordion="Generation controls",
|
| 406 |
+
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 407 |
|
| 408 |
if __name__ == "__main__":
|
| 409 |
demo.launch()
|