Faden 11: Gateway-Bild-Weiche - Requests mit Bild automatisch an VL-30B

Entscheid Weg A (vision-harness-verdikt): das MTP-Hirn (fast/hermes) bleibt
schnell und bildunfaehig; Bild-Requests routet das MC2-Gateway automatisch ans
Vision-Modell - kein manueller Modellwechsel, Lucys Flow bleibt.

router_logic.py: has_image(body) erkennt OpenAI-multimodalen content
(image_url/input_image/image); _text_of zieht jetzt nur Text-Parts fuer das
Komplexitaets-Routing (str() einer content-Liste haette die Zeichen-Schwelle
verfaelscht). VISION_CAPABLE = {vision, scout}.
routing_policy.py: neues UI-editierbares vision-Alias (Default env MC_ROUTE_VISION
= "vision"; leer = Weiche aus), darf wie coder_lite leer sein.
gateway_proxy.py _proxy: nach der Alias-Wahl - wenn ein Bild dabei ist und das
Ziel nicht bildfaehig - Override auf das vision-Alias (auch bei explizitem
model=hermes; genau Lucys Fall). Header x-mc-route-reason.

Verifiziert (Unit): Bild+hermes->vision, Bild+auto->vision, Text->fast,
Bild+scout->scout (schon bildfaehig), has_image korrekt.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
This commit is contained in:
Hitonabi
2026-07-13 22:42:08 +02:00
parent 913256e241
commit 5d02e7af91
3 changed files with 50 additions and 4 deletions
+12 -2
View File
@@ -5,7 +5,7 @@ from fastapi.responses import JSONResponse, StreamingResponse
from config import LLAMA_SWAP_URL from config import LLAMA_SWAP_URL
from services.gateway_stream import record_stream_chunk, record_usage from services.gateway_stream import record_stream_chunk, record_usage
from services.router_logic import choose_for_lane from services.router_logic import VISION_CAPABLE, choose_for_lane, has_image
from services.routing_policy import load_policy from services.routing_policy import load_policy
router = APIRouter(prefix="/v1") router = APIRouter(prefix="/v1")
@@ -99,8 +99,18 @@ async def _proxy(path: str, request: Request):
else: else:
alias = requested alias = requested
routed = {"x-mc-routed-to": requested} routed = {"x-mc-routed-to": requested}
# fast-Spur: Thinking aus für flotte Antworten (sofern Client es nicht selbst setzt).
pol = load_policy() pol = load_policy()
# Bild-Weiche (Faden 11): Requests mit Bild-Anhang automatisch ans Vision-Modell umleiten,
# sofern das Ziel nicht ohnehin bildfähig ist. Das MTP-Hirn (fast/hermes) kann keine Bilder —
# so bleibt Lucy schnell, ohne dass jemand manuell das Modell wechselt (Entscheid Weg A).
vision_alias = pol.get("vision")
if vision_alias and alias not in VISION_CAPABLE and has_image(body):
alias = vision_alias
body["model"] = alias
routed = {"x-mc-routed-to": alias, "x-mc-route-reason": "Bild erkannt → Vision-Modell",
"x-mc-lane": routed.get("x-mc-lane", "-")}
# fast-Spur: Thinking aus für flotte Antworten (sofern Client es nicht selbst setzt).
if pol["fast_no_think"] and alias == pol["fast"] and "chat_template_kwargs" not in body: if pol["fast_no_think"] and alias == pol["fast"] and "chat_template_kwargs" not in body:
body["chat_template_kwargs"] = {"enable_thinking": False} body["chat_template_kwargs"] = {"enable_thinking": False}
_inject_language(body, alias) _inject_language(body, alias)
+33 -1
View File
@@ -50,9 +50,41 @@ _CODE_HINT = re.compile(
) )
# Bild-Weiche (Faden 11): Aliase, die selbst Bilder koennen — die werden NIE umgeroutet.
VISION_CAPABLE = {"vision", "scout"}
_IMAGE_PART_TYPES = {"image_url", "input_image", "image"}
def has_image(body: dict) -> bool:
"""True, wenn irgendeine Nachricht einen Bild-Part enthaelt (OpenAI-multimodaler
content: eine Liste mit einem {"type": "image_url"|"input_image"|"image", ...}-Teil)."""
for m in body.get("messages") or []:
if not isinstance(m, dict):
continue
content = m.get("content")
if isinstance(content, list):
for part in content:
if isinstance(part, dict) and part.get("type") in _IMAGE_PART_TYPES:
return True
return False
def _text_of(body: dict) -> str: def _text_of(body: dict) -> str:
msgs = body.get("messages") or [] msgs = body.get("messages") or []
return "\n".join(str(m.get("content") or "") for m in msgs) # Multimodaler content ist eine Liste — nur die Text-Parts fuers Komplexitaets-Routing
# zusammenziehen (ein dict/Liste als str() wuerde die Zeichen-Schwelle verfaelschen).
out = []
for m in msgs:
if not isinstance(m, dict):
continue
c = m.get("content")
if isinstance(c, str):
out.append(c)
elif isinstance(c, list):
for part in c:
if isinstance(part, dict) and part.get("type") == "text":
out.append(str(part.get("text") or ""))
return "\n".join(out)
def _route_chat(text: str, n: int) -> tuple[str, str]: def _route_chat(text: str, n: int) -> tuple[str, str]:
+5 -1
View File
@@ -30,6 +30,9 @@ DEFAULTS: dict = {
"heavy": os.environ.get("MC_ROUTE_HEAVY", "heavy"), "heavy": os.environ.get("MC_ROUTE_HEAVY", "heavy"),
"coder": os.environ.get("MC_ROUTE_CODER", "coder"), "coder": os.environ.get("MC_ROUTE_CODER", "coder"),
"coder_lite": os.environ.get("MC_ROUTE_CODER_LITE", ""), "coder_lite": os.environ.get("MC_ROUTE_CODER_LITE", ""),
# Bild-Weiche (Faden 11): Requests mit Bild-Anhang werden automatisch hierhin geroutet
# (die MTP-Hirn-Config kann keine Bilder). "" schaltet die Weiche ab (Passthrough).
"vision": os.environ.get("MC_ROUTE_VISION", "vision"),
"heavy_chars": int(os.environ.get("MC_GATEWAY_HEAVY_CHARS", "8000")), "heavy_chars": int(os.environ.get("MC_GATEWAY_HEAVY_CHARS", "8000")),
"coding_escalate_chars": int(os.environ.get("MC_CODING_ESCALATE_CHARS", "120000")), "coding_escalate_chars": int(os.environ.get("MC_CODING_ESCALATE_CHARS", "120000")),
"fast_no_think": _env_bool("MC_FAST_NO_THINK", "1"), "fast_no_think": _env_bool("MC_FAST_NO_THINK", "1"),
@@ -41,6 +44,7 @@ FIELDS: list[dict] = [
{"key": "heavy", "label": "heavy-Alias (chat: lang/komplex)", "type": "str"}, {"key": "heavy", "label": "heavy-Alias (chat: lang/komplex)", "type": "str"},
{"key": "coder", "label": "coder-Alias (coding: stark / Eskalation)", "type": "str"}, {"key": "coder", "label": "coder-Alias (coding: stark / Eskalation)", "type": "str"},
{"key": "coder_lite", "label": "coder-lite-Alias (coding: schneller Default; leer = aus)", "type": "str"}, {"key": "coder_lite", "label": "coder-lite-Alias (coding: schneller Default; leer = aus)", "type": "str"},
{"key": "vision", "label": "vision-Alias (Bild-Weiche: Requests mit Bild; leer = aus)", "type": "str"},
{"key": "heavy_chars", "label": "chat → heavy ab N Zeichen", "type": "int", "min": 500, "max": 1_000_000}, {"key": "heavy_chars", "label": "chat → heavy ab N Zeichen", "type": "int", "min": 500, "max": 1_000_000},
{"key": "coding_escalate_chars", "label": "coding → starker Coder ab N Zeichen", "type": "int", "min": 1000, "max": 4_000_000}, {"key": "coding_escalate_chars", "label": "coding → starker Coder ab N Zeichen", "type": "int", "min": 1000, "max": 4_000_000},
{"key": "fast_no_think", "label": "fast-Spur: Thinking aus (flotte Antworten)", "type": "bool"}, {"key": "fast_no_think", "label": "fast-Spur: Thinking aus (flotte Antworten)", "type": "bool"},
@@ -77,7 +81,7 @@ def _coerce(patch: dict) -> dict:
out[k] = bool(v) out[k] = bool(v)
else: # str else: # str
sv = str(v).strip() sv = str(v).strip()
if k != "coder_lite" and not sv: if k not in ("coder_lite", "vision") and not sv:
raise ValueError(f"{k} darf nicht leer sein") raise ValueError(f"{k} darf nicht leer sein")
out[k] = sv out[k] = sv
return out return out