Feat: Vulkan/RADV-Engine + vocab-gepruefte Spec-Drafts + Provisioning/Sync

Engine-Cutover ROCm/HIP -> Vulkan/RADV (gfx1151): +12-22% tg auf MoE (llama-bench
verifiziert, fast 53->65 t/s). ROCm-Build bleibt als Rollback unter /opt/llamacpp.

- Backend: vocab-aware Speculative Decoding. services/gguf_meta.py liest den
  Tokenizer-Fingerprint (model/pre/n_vocab) direkt aus dem GGUF-Header (ohne Modell-Load);
  register_model + migrate_config haengen nur VOCAB-KOMPATIBLE Drafts an (inkl. --spec-type,
  das in dieser llama.cpp-Generation noetig ist). Neue Endpoints /api/models/drafts + /{id}/draft.
- Frontend: idiotensichere Spec-Draft-UI (SpecDraftModal) - nur kompatible Drafts waehlbar,
  inkompatible gesperrt mit Begruendung; SPEC/SPEC?-Badge nach echtem Aktiv-Status; Rolle in AddModel.
- maintenance.py: Engine-Update-Quelle -> ggml-org/llama.cpp (Build-Nummer-Vergleich),
  ENGINE_PATH=/opt/llamacpp-vulkan.
- Startup-Warmup der brains (deploy/warmup.sh, self-detaching ExecStartPost) + deploy/provision-engine.sh.
- Cleanup: tote LiteLLM gateway/config.yaml + alle Referenzen (config.py/backup.py/backup.sh) entfernt;
  README + docs/memory aktualisiert.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
This commit is contained in:
Hitonabi
2026-06-27 01:54:29 +02:00
parent 530d77ff1b
commit c48e583790
23 changed files with 1128 additions and 508 deletions
+31 -5
View File
@@ -7,6 +7,7 @@ erweiterte sudoers (siehe docs/BEDIENUNG.md). Lange Ops laufen als jobengine-Job
"""
import os
import re
import subprocess
import time
from datetime import datetime
@@ -21,10 +22,30 @@ SYSTEM_SERVICES = {"llama-swap"}
USER_SERVICES = {"mission-control-2", "hermes-gateway", "hermes-dashboard", "hermes-webui"}
ENGINE_UPDATE_CMD = os.environ.get("MC_ENGINE_UPDATE_CMD", "")
ENGINE_PATH = os.environ.get("MC_ENGINE_PATH", "/opt/llamacpp")
# Engine = offizieller Vulkan-Build von ggml-org/llama.cpp (RADV auf Strix Halo).
ENGINE_PATH = os.environ.get("MC_ENGINE_PATH", "/opt/llamacpp-vulkan")
ENGINE_REPO = os.environ.get("MC_ENGINE_REPO", "ggml-org/llama.cpp")
_engine_cache = {"ts": 0.0, "avail": False}
def _installed_engine_build() -> int | None:
"""Build-Nummer der installierten llama-server-Binary (z.B. 9821), oder None.
Vulkan-Build braucht LD_LIBRARY_PATH=ENGINE_PATH zum Start von --version."""
bin_path = os.path.join(ENGINE_PATH, "llama-server")
if not os.path.exists(bin_path):
return None
try:
env = dict(os.environ, LD_LIBRARY_PATH=ENGINE_PATH)
out = subprocess.run([bin_path, "--version"], capture_output=True, text=True,
timeout=20, env=env)
txt = (out.stderr or "") + (out.stdout or "")
if (m := re.search(r"build:\s*\S+\s*\((\d+)\)", txt)) or (m := re.search(r"\bb(\d{3,})\b", txt)):
return int(m.group(1))
except Exception:
return None
return None
def _ram_gb() -> float:
return psutil.virtual_memory().total / (1024 ** 3)
@@ -45,11 +66,16 @@ def _engine_update_available() -> bool:
return _engine_cache["avail"]
avail = False
try:
rel = httpx.get("https://api.github.com/repos/lemonade-sdk/llamacpp-rocm/releases/latest",
rel = httpx.get(f"https://api.github.com/repos/{ENGINE_REPO}/releases/latest",
timeout=6, headers={"User-Agent": "MissionControl2"}).json()
pub = datetime.fromisoformat(rel["published_at"].replace("Z", "+00:00")).timestamp()
inst = os.path.getmtime(ENGINE_PATH)
avail = pub > inst + 86400
tag = str(rel.get("tag_name", ""))
latest = int(m.group(1)) if (m := re.search(r"(\d{3,})", tag)) else None
installed = _installed_engine_build()
if latest is not None and installed is not None:
avail = latest > installed # präziser Build-Nummer-Vergleich
else: # Fallback: Release-Datum vs. Engine-mtime
pub = datetime.fromisoformat(rel["published_at"].replace("Z", "+00:00")).timestamp()
avail = pub > os.path.getmtime(ENGINE_PATH) + 86400
except Exception:
avail = False
_engine_cache.update(ts=now, avail=avail)