ZeroGPU: xlarge slice for 54.7GB model, duration 120, theme on launch()

app.py

@@ -70,7 +70,10 @@ def split_thinking(text):
     return [gr.ChatMessage(role="assistant", content=text)]
 
 
-@spaces.GPU(duration=150)
+# size='xlarge': this is a 27B at bf16 (~54.7 GB packed), which does not fit the default
+# ZeroGPU slice -- the symptom was "GPU task aborted" with nothing in the Space logs.
+# duration is kept at 120s, the usual ZeroGPU ceiling; 150 was over it.
+@spaces.GPU(duration=120, size="xlarge")
 def respond(message: dict, history: list, temperature: float, top_p: float, top_k: int):
     """Chat with Qwen3.8-27B-Uncensored. Accepts text and images, streams the reply.
 
@@ -123,7 +126,7 @@ EXAMPLES = [
               "subscription service that mails artisanal ice cubes.", "files": []}],
 ]
 
-with gr.Blocks(title="Qwen3.8-27B-Uncensored", theme=gr.themes.Soft()) as demo:
+with gr.Blocks(title="Qwen3.8-27B-Uncensored") as demo:
     gr.Markdown(
         "# Qwen3.8-27B-Uncensored\n"
         "Chat with the GGUF-quantized uncensored Qwen3.8-27B. Text and images. "
@@ -149,5 +152,5 @@ with gr.Blocks(title="Qwen3.8-27B-Uncensored", theme=gr.themes.Soft()) as demo:
     )
 
 if __name__ == "__main__":
-    # theme belongs to Blocks, not launch(): passing it here is a TypeError at startup.
-    demo.launch(mcp_server=True)
+    # Gradio 6 moved theme from the Blocks constructor to launch().
+    demo.launch(mcp_server=True, theme=gr.themes.Soft())