Spaces:
Running
on
Zero
Running
on
Zero
Update app.py
Browse files
app.py
CHANGED
@@ -21,11 +21,15 @@ LICENSE = """
|
|
21 |
---.
|
22 |
"""
|
23 |
|
|
|
|
|
24 |
def load_model(model_id):
|
25 |
-
|
26 |
-
|
27 |
-
|
28 |
-
|
|
|
|
|
29 |
|
30 |
if not torch.cuda.is_available():
|
31 |
DESCRIPTION += "\n<p>Running on CPU 🥶 This demo does not work on CPU.</p>"
|
@@ -86,7 +90,7 @@ chat_interface = gr.ChatInterface(
|
|
86 |
fn=generate,
|
87 |
additional_inputs=[
|
88 |
gr.Textbox(label="System prompt", lines=6),
|
89 |
-
gr.Textbox(label="Model ID",
|
90 |
gr.Slider(
|
91 |
label="Max new tokens",
|
92 |
minimum=1,
|
@@ -137,4 +141,4 @@ with gr.Blocks(css="style.css") as demo:
|
|
137 |
gr.Markdown(LICENSE)
|
138 |
|
139 |
if __name__ == "__main__":
|
140 |
-
demo.queue(max_size=20).launch()
|
|
|
21 |
---.
|
22 |
"""
|
23 |
|
24 |
+
models_cache = {}
|
25 |
+
|
26 |
def load_model(model_id):
|
27 |
+
if model_id not in models_cache:
|
28 |
+
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto", load_in_4bit=True)
|
29 |
+
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
30 |
+
tokenizer.use_default_system_prompt = False
|
31 |
+
models_cache[model_id] = (model, tokenizer)
|
32 |
+
return models_cache[model_id]
|
33 |
|
34 |
if not torch.cuda.is_available():
|
35 |
DESCRIPTION += "\n<p>Running on CPU 🥶 This demo does not work on CPU.</p>"
|
|
|
90 |
fn=generate,
|
91 |
additional_inputs=[
|
92 |
gr.Textbox(label="System prompt", lines=6),
|
93 |
+
gr.Textbox(label="Model ID", default="Nekochu/Luminia-13B-v3"),
|
94 |
gr.Slider(
|
95 |
label="Max new tokens",
|
96 |
minimum=1,
|
|
|
141 |
gr.Markdown(LICENSE)
|
142 |
|
143 |
if __name__ == "__main__":
|
144 |
+
demo.queue(max_size=20).launch()
|