scripts/litellm-health-check.py

@@ -55,9 +55,12 @@ def probe_http(url, method="GET", bearer_token=None, data=None, timeout=10, foll
     try:
         rc, stdout, stderr = run_command(cmd, timeout)
         if rc != 0:
-            # Determine failure kind from curl exit code
+            # Check if this is a timeout from run_command (rc=1, stderr="TIMEOUT")
+            if rc == 1 and stderr == "TIMEOUT":
+                return (000, "timeout after " + str(timeout) + "s")
+            # Otherwise, determine failure kind from curl exit code
             # curl exit codes: 28=timeout, 7=refused, 6=dns, 35=ssl, 52=empty
-            if rc == 28:
+            elif rc == 28:
                 return (000, "timeout after " + str(timeout) + "s")
             elif rc == 7:
                 return (000, "connection refused")
@@ -73,6 +76,20 @@ def probe_http(url, method="GET", bearer_token=None, data=None, timeout=10, foll
     except subprocess.TimeoutExpired:
         return (000, "timeout after " + str(timeout) + "s")
 
+def check_host_health(host_ip):
+    """Check if the GPU host's llama-chat-api health endpoint is reachable
+    
+    Returns: (healthy: bool, detail: str)
+    """
+    code, _ = probe_http("http://" + host_ip + ":8080/health", timeout=10)
+    if code == 200:
+        return True, "host healthy (200)"
+    elif code == 000:
+        return False, "host unreachable (timeout or refused)"
+    else:
+        return False, "host unhealthy (HTTP " + str(code) + ")"
+
+
 def get_response_body(url, method="POST", bearer_token=None, data=None, timeout=30):
     """Get response body for 401/403 credential faults (truncated to 200 chars)"""
     cmd = "curl -s -m " + str(timeout)
@@ -115,43 +132,53 @@ def check_model_probes():
     
     results = []
     
+    # Host health mapping: model -> host IP
+    model_hosts = {
+        "gpu-dense": "192.168.68.8",  # RTX 3090
+        "gpu-vision": "192.168.68.110",  # RTX 5070
+        "strix-moe": "192.168.68.15"  # Strix Halo
+    }
+    
     for model in ["gpu-dense", "gpu-vision", "strix-moe"]:
-        # Single-host aliases: 30s initial timeout, retry once at 45s on failure
+        host_ip = model_hosts[model]
-        # gpu-dense (RTX 3090) may need long warmup/prefill or concurrent generation hold
+        # Single-host aliases: 30s initial timeout, retry once at 90s on failure
+        # Worst-case prefill ~76s, so 90s retry ensures we cover it
         code, failure_kind = probe_http("http://" + BACKEND_HOST + "/litellm/v1/chat/completions",
                                        method="POST",
                                        bearer_token=monitor_key,
                                        data='{"model":"' + model + '","messages":[{"role":"user","content":"health ' + str(random.randint(1000, 9999)) + '"}],"max_tokens":4}',
                                        timeout=30)
         
-        first_kind = None  # Track first attempt's failure kind
+        first_kind = None
         if code == 000 and failure_kind:
-            # Retry once with longer timeout (45s) before declaring failure
             first_kind = failure_kind
             time.sleep(1)
             code, failure_kind = probe_http("http://" + BACKEND_HOST + "/litellm/v1/chat/completions",
                                            method="POST",
                                            bearer_token=monitor_key,
                                            data='{"model":"' + model + '","messages":[{"role":"user","content":"health ' + str(random.randint(1000, 9999)) + '"}],"max_tokens":4}',
-                                           timeout=45)
+                                           timeout=90)
         
         if code == 000 and failure_kind:
-            # Both attempts failed - report both kinds
+            # Both attempts failed - check host health to distinguish busy from down
-            if first_kind:
+            host_healthy, host_detail = check_host_health(host_ip)
-                results.append((model, False, "probe-failed: " + model + " " + first_kind + " then " + failure_kind + " (2 attempts)"))
+            if host_healthy:
+                results.append((model, False, "busy (completion timed out after retry; host healthy " + host_detail + ")"))
             else:
-                results.append((model, False, "probe-failed: " + model + " " + failure_kind))
+                # Host unreachable - report both kinds
+                if first_kind:
+                    results.append((model, False, "probe-failed: " + model + " " + first_kind + " then " + failure_kind + " (2 attempts)"))
+                else:
+                    results.append((model, False, "probe-failed: " + model + " " + failure_kind))
         elif code == 200:
             results.append((model, True, str(code) + " (target: " + BACKEND_HOST + "/litellm/v1/chat/completions, model=" + model + ")"))
         elif code in (401, 403):
-            # Credential fault - capture body and key alias
             body = get_response_body("http://" + BACKEND_HOST + "/litellm/v1/chat/completions",
                                     method="POST",
                                     bearer_token=monitor_key,
                                     data='{"model":"' + model + '","messages":[{"role":"user","content":"health"}],"max_tokens":4}',
                                     timeout=10)
-            # Resolve key alias
+            alias = "monitor-20260813"
-            alias = "monitor-20260813"  # Known from /etc/litellm-monitor.env on CT 116
             results.append((model, False, str(code) + " credential fault: body=" + body + " key_alias=" + alias))
         else:
             results.append((model, False, str(code) + " (target: " + BACKEND_HOST + "/litellm/v1/chat/completions, model=" + model + ")"))