fix(mellea): restore vLLM probe with auth header + non-5xx check

app/mellea_validator.py

@@ -382,6 +382,38 @@ def reconcile_strict_streaming(
     _first_token_timeout = int(os.environ.get("RIPRAP_FIRST_TOKEN_TIMEOUT_S", "400"))
     _inter_token_timeout = int(os.environ.get("RIPRAP_TOKEN_TIMEOUT_S", "45"))
 
+    # When PRIMARY=vllm, RunPod cold-starts take ~250-360s (container boot +
+    # model load). The pod starts when the warmup request hits the proxy, but
+    # the proxy returns 503 immediately. LiteLLM would fall back to Ollama and
+    # fail before RunPod is ready. Poll /v1/models here (after specialists, but
+    # before generation) and wait — keepalives keep the SSE connection alive.
+    _vllm_base = os.environ.get("RIPRAP_LLM_BASE_URL", "").rstrip("/")
+    if os.environ.get("RIPRAP_LLM_PRIMARY", "ollama") == "vllm" and _vllm_base:
+        try:
+            import httpx as _httpx
+            _probe_url = f"{_vllm_base}/models"
+            _probe_key = os.environ.get("RIPRAP_LLM_API_KEY", "") or "EMPTY"
+            _probe_headers = {"Authorization": f"Bearer {_probe_key}"}
+            _probe_deadline = t0 + _first_token_timeout
+            log.info("mellea: polling vLLM readiness at %s", _probe_url)
+            while time.time() < _probe_deadline:
+                try:
+                    _r = _httpx.get(_probe_url, headers=_probe_headers, timeout=5.0)
+                    # Any non-503/502/504 means the service is UP (200 = ready,
+                    # 401 = auth-gated but alive, 404 = wrong path but alive).
+                    if _r.status_code not in (502, 503, 504):
+                        log.info("mellea: vLLM ready (status=%d, %.1fs elapsed)",
+                                 _r.status_code, time.time() - t0)
+                        break
+                except Exception as _pe:
+                    log.debug("mellea: vLLM probe: %r", _pe)
+                time.sleep(10)
+            else:
+                log.warning("mellea: vLLM not ready after %.1fs, proceeding anyway",
+                            time.time() - t0)
+        except ImportError:
+            log.warning("mellea: httpx not available, skipping vLLM probe")
+
     for attempt_idx in range(loop_budget):
         attempts = attempt_idx + 1
         # On reroll, append a tight feedback message naming what failed AND