feat: LLM warmup endpoint + first-turn status message

- Add /warmup endpoint to Modal LLM app to pre-load model on page load
- Add llm_engine.warmup() with rate limiting, called from reset_session()
- Show 'patient on their way' status on first audio turn in frontend
before processing begins, setting expectations for cold start delay

core.py

@@ -4,8 +4,8 @@ import tempfile
 
 from prompts import SYSTEM_PROMPT, PHASE_SWITCH_REMINDER
 from parse_feedback import parse_feedback, render_feedback_table, strip_markdown
-from stt_engine import transcribe, warmup
-from llm_engine import chat as llm_chat
+from stt_engine import transcribe, warmup as stt_warmup
+from llm_engine import chat as llm_chat, warmup as llm_warmup
 from tts_engine import synthesize
 
 TERMINATE_RE = re.compile(
@@ -131,5 +131,6 @@ def end_session_click(state):
 
 
 def reset_session():
-    warmup()
+    stt_warmup()
+    llm_warmup()
     return make_initial_state()

custom_index.html

@@ -776,6 +776,13 @@ function stopRecording() {
 // Process audio turn
 async function processAudio(blob) {
   if (!client) return;
+
+  const isFirstTurn = !sessionState || sessionState.turn_count === 0;
+  if (isFirstTurn) {
+    setStatus("🚶 Le patient arrive au cabinet… veuillez patienter.");
+    await new Promise(r => setTimeout(r, 2000));
+  }
+
   isProcessing = true;
   micBtn.disabled = true;
   btnEnd.disabled = true;

llm_engine.py

@@ -1,4 +1,7 @@
 import os
+import threading
+import time
+
 import httpx
 
 MODAL_ENDPOINT = os.environ.get("MODAL_ENDPOINT_QWEN", "")
@@ -6,6 +9,37 @@ MODAL_AUTH_TOKEN = os.environ.get("MODAL_AUTH_TOKEN", "")
 
 TRUNCATION_LIMIT = 20_000
 
+_last_warmup = 0.0
+_warmup_lock = threading.Lock()
+
+
+def warmup():
+    """Send a lightweight request to Modal to trigger LLM container spin-up.
+
+    Rate-limited to once per 120s to avoid spamming on every page load.
+    """
+    global _last_warmup
+    if not MODAL_ENDPOINT:
+        return
+
+    now = time.monotonic()
+    with _warmup_lock:
+        if now - _last_warmup < 120:
+            return
+        _last_warmup = now
+
+    def _ping():
+        try:
+            resp = httpx.post(
+                MODAL_ENDPOINT + "/warmup",
+                json={"token": MODAL_AUTH_TOKEN},
+                timeout=120,
+            )
+        except Exception:
+            pass
+
+    threading.Thread(target=_ping, daemon=True).start()
+
 
 def chat(messages: list[dict]) -> str | None:
     if not MODAL_ENDPOINT:

modal_app.py

@@ -50,15 +50,6 @@ async def _load_llm():
             n_threads=4,
             verbose=False,
         )
-        try:
-            draft_path = hf_hub_download(
-                repo_id="unsloth/gemma-4-26B-A4B-it-GGUF",
-                filename="mtp-gemma-4-26B-A4B-it.gguf",
-                cache_dir=CACHE_DIR,
-            )
-            kwargs["draft_model"] = draft_path
-        except Exception:
-            pass
         _llm_model = Llama(**kwargs)
         await llm_vol.commit.aio()
 
@@ -122,6 +113,16 @@ def llm_web():
             },
         }
 
+    @fastapi_app.post("/warmup")
+    async def warmup(request: Request):
+        body = await request.json()
+        token = body.get("token", "")
+        expected = os.environ.get("EXPECTED_TOKEN", "")
+        if not token or token != expected:
+            raise HTTPException(401, "Unauthorized")
+        await _load_llm()
+        return {"status": "ready"}
+
     return fastapi_app