Fix country/language selectors: Radio+flag-emoji values broke single-select (US never applied); use Dropdown with clean MX/US, en/es values

app.py

@@ -287,9 +287,12 @@ def build_demo() -> gr.Blocks:
         # Holds only a session-id string (never the un-pickleable Ledger object).
         session = gr.State(value=_new_session_id)
         with gr.Row():
-            country = gr.Radio(["🇲🇽 Mexico", "🇺🇸 USA"], value="🇲🇽 Mexico",
-                               label="Country / tax system",
-                               info="Mexico: RESICO · IVA · SAT. USA: Schedule C · SE tax · federal.")
+            # Use (label, value) tuples with PLAIN ascii values ("MX"/"US"). Flag-emoji
+            # values were breaking Gradio's option matching, so the selection never
+            # actually switched to USA. The emoji stays in the visible label only.
+            country = gr.Dropdown([("🇲🇽 Mexico", "MX"), ("🇺🇸 USA", "US")], value="MX",
+                                  label="Country / tax system", filterable=False,
+                                  info="Mexico: RESICO · IVA · SAT. USA: Schedule C · SE tax · federal.")
 
         with gr.Tabs():
             # ---- Capture ----
@@ -353,8 +356,8 @@ def build_demo() -> gr.Blocks:
                 with gr.Row():
                     a_year = gr.Dropdown(YEARS, value=2024, label="Year", scale=1)
                     a_month = gr.Dropdown(MONTHS, value=5, label="Month", scale=1)
-                    a_lang = gr.Radio(["English", "Español"], value="English",
-                                      label="Answer language", scale=1)
+                    a_lang = gr.Dropdown([("English", "en"), ("Español", "es")], value="en",
+                                         label="Answer language", filterable=False, scale=1)
                 a_q = gr.Textbox(label="Your question",
                                  placeholder="Which regime suits me? Can I deduct my laptop?")
                 a_go = gr.Button("Ask", variant="primary")

modal_app/serve_llm_modal.py

@@ -41,7 +41,11 @@ MINUTES = 60
 @app.function(
     image=vllm_image,
     gpu="A10G",
-    scaledown_window=10 * MINUTES,   # stay warm 10 min between questions
+    # Keep one GPU warm during judging → no cold start, ~10-30s answers.
+    # ⚠️ COSTS ~$26/day in Modal credits. To stop the cost after judging, set
+    # min_containers=0 (or delete this line) and redeploy.
+    min_containers=1,
+    scaledown_window=10 * MINUTES,
     timeout=20 * MINUTES,
     volumes={"/root/.cache/huggingface": hf_cache, "/root/.cache/vllm": vllm_cache},
     secrets=[hf_secret],
@@ -57,8 +61,11 @@ def serve():
         "--enable-auto-tool-choice",
         "--tool-call-parser", "hermes",
         "--reasoning-parser", "deepseek_r1",  # handles Qwen3's <think> tags
-        "--max-model-len", "16384",
+        "--max-model-len", "8192",            # plenty for our prompts; faster KV alloc
         "--gpu-memory-utilization", "0.92",
+        # Skip torch.compile + CUDA-graph capture — they added ~240s to cold start.
+        # Slightly slower per-token, but the model wakes up in ~1 min instead of ~5.
+        "--enforce-eager",
     ]
     if MODEL_REVISION:
         cmd += ["--revision", MODEL_REVISION]