Add final Qwen3.8 runtime and uncensored profile

router/ai_profile_router.py

@@ -2,7 +2,7 @@
 """AI Profile Router – OpenAI-kompatibler Proxy vor llama.cpp.
 
 Leitet OpenAI-kompatible Requests transparent an den lokalen llama.cpp-Server
-weiter (Streaming, Tool Calls, JSON) und schaltet zwischen vier festen
+weiter (Streaming, Tool Calls, JSON) und schaltet zwischen fünf festen
 Profilen um:
 
     Profil   Kontext
@@ -11,9 +11,11 @@ Profilen um:
     medium   160000
     large    192000
     ultra    262144
+    uncensored 80000
 
-Virtuelle Modelle:  qwen-fast, qwen-medium, qwen-large, qwen-ultra
-Kommandos:          POST /fast, /medium, /large, /ultra (Profilwechsel)
+Virtuelle Modelle:  qwen-fast, qwen-medium, qwen-large, qwen-ultra,
+                    qwen-uncensored
+Kommandos:          POST /fast, /medium, /large, /ultra, /uncensored
                     GET  /status                 (Zustand)
 
 Bildgenerierung (FLUX.2 [klein] 4B Base):
@@ -496,7 +498,7 @@ def _wait_ready(profile: str, deadline: float) -> None:
     while True:
         status = upstream_status()
         if (status["reachable"] and status.get("model")
-                and status.get("ctx") == expected_ctx
+                and _context_matches(expected_ctx, status.get("ctx"))
                 and (not expected_model or status.get("model") == expected_model)):
             log.info("llama.cpp bereit: Profil=%s Modell=%s ctx=%s",
                      profile, status.get("model"), status.get("ctx"))
@@ -514,11 +516,22 @@ def _profile_is_ready(profile: str, status: dict | None = None) -> bool:
     status = status or upstream_status()
     expected_model = EXPECTED_MODELS.get(profile)
     return bool(status.get("reachable") and status.get("model")
-                and status.get("ctx") == PROFILES[profile]
+                and _context_matches(PROFILES[profile], status.get("ctx"))
                 and (not expected_model
                      or status.get("model") == expected_model))
 
 
+def _context_matches(expected: int, reported: object) -> bool:
+    """Allow llama.cpp's small MTP/speculative context overhead.
+
+    Recent builds can report a slot context slightly above the requested
+    ``--ctx-size`` (currently 128 tokens with the tested Qwen MTP setup).
+    The alias check and this narrow bound still reject neighbouring profiles.
+    """
+    return (isinstance(reported, int)
+            and expected <= reported <= expected + 1024)
+
+
 def switch_profile(profile: str, implicit: bool = False) -> None:
     """Stellt sicher, dass das Profil aktiv ist, und wartet bis es geladen ist.
 
@@ -1184,11 +1197,11 @@ class Handler(BaseHTTPRequestHandler):
                 self._images_list()
             elif path.startswith("/images/") and self.command == "GET":
                 self._image_serve(path[len("/images/"):])
-            elif (path in ("/fast", "/medium", "/large", "/ultra")
+            elif (path in ("/fast", "/medium", "/large", "/ultra", "/uncensored")
                   and (self.command == "POST"
                        or (self.command == "GET" and ALLOW_LEGACY_GET_SWITCH))):
                 self._switch(path[1:])
-            elif (path in ("/fast", "/medium", "/large", "/ultra")
+            elif (path in ("/fast", "/medium", "/large", "/ultra", "/uncensored")
                   and self.command == "GET"):
                 self._send_error(405, "Profilwechsel erfordert POST",
                                  "invalid_request_error", "method_not_allowed")