feat(perf): Optimize KV-Cache to Q4_K and fix Hermes Context bounds

backend/routers/gateway_proxy.py

@@ -98,11 +98,23 @@ async def _proxy(path: str, request: Request):
 
     client = request.app.state.gw_client  # geteilter Keep-Alive-Client (siehe app.py lifespan)
     if body.get("stream"):
+        req = client.build_request("POST", url, json=body, timeout=None)
+        r = await client.send(req, stream=True)
+        if r.status_code != 200:
+            await r.aread()
+            try:
+                resp_json = r.json()
+            except Exception:
+                resp_json = {"error": {"message": r.text, "type": "upstream_error"}}
+            return JSONResponse(resp_json, status_code=r.status_code, headers=routed)
+
         async def gen():
-            async with client.stream("POST", url, json=body, timeout=None) as r:
+            try:
                 async for chunk in r.aiter_raw():
                     record_stream_chunk(chunk, alias)
                     yield chunk
+            finally:
+                await r.aclose()
         return StreamingResponse(gen(), media_type="text/event-stream", headers=routed)
 
     r = await client.post(url, json=body, timeout=600.0)

deploy/llama-swap.config.yaml

@@ -8,7 +8,7 @@ models:
     # nicht mehr); KV Q8_0 macht das speicherneutral zu vorher (65k f16). Bench 2026-07-02:
     # Q8_0 kostet 0 t/s, MTP n-max 3 = +26% vs. ohne Spec.
     cmd: |
-      llama-server -m /srv/models/Qwen3.6-35B-A3B-MTP-GGUF/Qwen3.6-35B-A3B-UD-Q4_K_M.gguf --host 127.0.0.1 --port ${PORT} -c 131072 -ngl 999 -fa on --no-mmap --jinja --parallel 2 -cram 16384 -ctk q8_0 -ctv q8_0 --spec-type draft-mtp --spec-draft-n-max 3
+      llama-server -m /srv/models/Qwen3.6-35B-A3B-MTP-GGUF/Qwen3.6-35B-A3B-UD-Q4_K_M.gguf --host 127.0.0.1 --port ${PORT} -c 131072 -ngl 999 -fa on --no-mmap --jinja --parallel 2 -cram 16384 -ctk q4_k -ctv q4_k --spec-type draft-mtp --spec-draft-n-max 3
     ttl: 0
     aliases:
     - hermes