Refactor app.py: Optimize model loading for ZeroGPU and adjust CSS injection for Gradio compatibility

app.py

@@ -111,11 +111,20 @@ except Exception as _e:
 
 model_id = "mistralai/Magistral-Small-2509"
 tokenizer = AutoTokenizer.from_pretrained(model_id, tokenizer_type="mistral", use_fast=False)
-model = Mistral3ForConditionalGeneration.from_pretrained(
+# On ZeroGPU, CUDA is emulated at module load and becomes a real GPU only inside
+# @spaces.GPU functions. The docs explicitly require model placement to happen at
+# the root module level (`.to("cuda")`); lazy moves inside @spaces.GPU are much
+# slower because tensor packing happens at startup.
+# https://huggingface.co/docs/hub/spaces-zerogpu#model-loading
+model = (
+    Mistral3ForConditionalGeneration.from_pretrained(
         model_id,
         torch_dtype=torch.bfloat16,
         low_cpu_mem_usage=True,
-    ).eval()
+    )
+    .to("cuda")
+    .eval()
+)
 
 
 # SYSTEM_PROMPT_TEXT = (
@@ -584,6 +593,7 @@ def _read_text(path: str) -> str:
         return ""
 
 APP_CSS = _read_text('static/style.css') + "\n#lat_box, #lng_box { display:none; }\n" + """
+#app-styles { display: none !important; }
 #next_btn { position: absolute; left: -9999px; }
 #lobby_group, #game_group{max-width:1024px;margin:24px auto;padding:16px;}
 #start_btn{height:48px;font-weight:700}
@@ -859,6 +869,11 @@ APP_BOOT_JS = """
 """.replace("__GMAPS_KEY__", GOOGLE_MAPS_API_KEY or '')
 
 with gr.Blocks(title="LLM GeoGuessr") as demo:
+    # Gradio 6 dropped `css=` on `Blocks(...)`, and `launch(css=...)` isn't
+    # always honoured under SSR. Injecting a <style> tag as a child of the app
+    # is the only place that's guaranteed to render in every version.
+    gr.HTML(f"<style>{APP_CSS}</style>", elem_id="app-styles")
+
     user_profile = gr.State()
 
     with gr.Row():
@@ -1344,8 +1359,4 @@ with gr.Blocks(title="LLM GeoGuessr") as demo:
 
 
 if __name__ == "__main__":
-    demo.queue().launch(
-        server_name="0.0.0.0",
-        server_port=7860,
-        css=APP_CSS,
-    )
+    demo.queue().launch(server_name="0.0.0.0", server_port=7860)