models.py

# models.py

import gradio as gr
import spaces
import transformers_gradio

# Load each model separately
demo_qwen = gr.load(
    name="deepseek-ai/DeepSeek-R1-Distill-Qwen-32B",
    src=transformers_gradio.registry
)
demo_r1 = gr.load(
    name="deepseek-ai/DeepSeek-R1",
    src=transformers_gradio.registry
)
demo_zero = gr.load(
    name="deepseek-ai/DeepSeek-R1-Zero",
    src=transformers_gradio.registry
)

# Ensure they run on GPU and disable api_name
for demo in [demo_qwen, demo_r1, demo_zero]:
    demo.fn = spaces.GPU()(demo.fn)
    for fn in demo.fns.values():
        fn.api_name = False

# If you still want to keep a default single "demo" reference
# you can pick one. For example:
demo = demo_qwen

if __name__ == "__main__":
    # You could choose to launch one of them (or none, if controlling from app.py)
    demo_qwen.launch()