Feat: Discover zukunftssicher (Recency-Score) + Agent-Hirn (Hermes) im Modell-Manager
- discover.rank_runnable: Score aus Fit + Recency (lastModified, Halbwertszeit ~9 Mon) + Capability (params, log) + Popularity (downloads, log) -> neuere Generationen bevorzugt. - Agent-Hirn: neuer GET /api/agent/brain (aktuelles hermes-Modell + bestes NousResearch- Hermes-Update, versions-aware via Hermes-X.Y-Parsing). Cockpit zeigt "Agent-Hirn (Hermes)"-Karte mit aktuellem Brain + Aktualisieren-Button, wenn NousResearch eine neuere Generation hat (z.B. Hermes-4-14B -> Hermes-4.3-36B). Update installiert mit Rolle hermes. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
This commit is contained in:
Hitonabi
@@ -3,7 +3,7 @@
|
||||
from fastapi import APIRouter
|
||||
from pydantic import BaseModel
|
||||
|
||||
from services.agent import agent_status, update_brain_model
|
||||
from services.agent import agent_status, hermes_brain_info, update_brain_model
|
||||
|
||||
router = APIRouter(prefix="/api")
|
||||
|
||||
@@ -17,6 +17,12 @@ def status() -> dict:
|
||||
return agent_status()
|
||||
|
||||
|
||||
@router.get("/agent/brain")
|
||||
def brain_info() -> dict:
|
||||
"""Aktuelles Agent-Hirn (hermes) + bestes NousResearch-Hermes-Update."""
|
||||
return hermes_brain_info()
|
||||
|
||||
|
||||
@router.post("/agent/brain")
|
||||
def set_brain_model(body: BrainReq) -> dict:
|
||||
ok = update_brain_model(body.model)
|
||||
|
||||
@@ -6,14 +6,78 @@ Hermes' eigener Config verdrahtet (siehe docs/HERMES_SETUP.md).
|
||||
|
||||
import logging
|
||||
import os
|
||||
import re
|
||||
|
||||
import httpx
|
||||
import psutil
|
||||
|
||||
from config import ANYTHINGLLM_URL, HERMES_API_URL, HERMES_HOME, PC_EXECUTOR_URL
|
||||
|
||||
log = logging.getLogger(__name__)
|
||||
|
||||
|
||||
def _hermes_version(name: str) -> float | None:
|
||||
"""Versionszahl aus 'Hermes-4.3', 'Hermes-4', 'Nous-Hermes-2' → 4.3/4.0/2.0."""
|
||||
low = (name or "").lower()
|
||||
if "hermes" not in low:
|
||||
return None
|
||||
m = re.search(r"hermes[-_ ]?(\d+(?:\.\d+)?)", low)
|
||||
return float(m.group(1)) if m else None
|
||||
|
||||
|
||||
def hermes_brain_info() -> dict:
|
||||
"""Aktuelles Agent-Hirn (hermes-Rolle) + bestes verfügbares NousResearch-Hermes-Modell,
|
||||
das auf diese Hardware passt. Für den Modell-Manager: Brain sichtbar + updatebar,
|
||||
sobald NousResearch eine neuere Hermes-Generation veröffentlicht."""
|
||||
from services import discover, llamaswap
|
||||
from services.fit import evaluate_fit, extract_params_b
|
||||
|
||||
models = llamaswap.list_models()
|
||||
cur = next((m for m in models if m.get("role") == "hermes"), None)
|
||||
cur_ver = _hermes_version(cur["name"]) if cur else None
|
||||
cur_params = (cur.get("capabilities") or {}).get("params_b") if cur else None
|
||||
current = None
|
||||
if cur:
|
||||
current = {"name": cur["name"], "filename": cur.get("filename"),
|
||||
"params_b": cur_params, "quant": cur.get("quant"),
|
||||
"size_bytes": cur.get("size_bytes"), "version": cur_ver,
|
||||
"gguf_path": cur.get("gguf_path"), "incomplete": cur.get("incomplete")}
|
||||
|
||||
ram = psutil.virtual_memory().total / (1024 ** 3)
|
||||
best = None
|
||||
try:
|
||||
cands = []
|
||||
for r in discover._fetch_author_models("NousResearch"):
|
||||
rid = r.get("id", "")
|
||||
if "hermes" not in rid.lower():
|
||||
continue
|
||||
pb = extract_params_b(rid)
|
||||
fit = evaluate_fit(pb, "Q4_K_M", 8192, ram, name=rid)
|
||||
if fit["level"] == "too_tight":
|
||||
continue
|
||||
cands.append({"repo": rid, "name": rid.split("/")[-1],
|
||||
"version": _hermes_version(rid) or 0.0, "params_b": pb,
|
||||
"downloads": int(r.get("downloads") or 0), "fit": fit})
|
||||
# neueste Hermes-Version zuerst, dann größer/fähiger, dann beliebter
|
||||
cands.sort(key=lambda c: (c["version"], c["params_b"], c["downloads"]), reverse=True)
|
||||
best = cands[0] if cands else None
|
||||
except Exception:
|
||||
log.debug("hermes_brain_info: HF-Abfrage fehlgeschlagen", exc_info=True)
|
||||
|
||||
update = False
|
||||
if best is not None:
|
||||
if cur_ver is None:
|
||||
update = True
|
||||
elif best["version"] > cur_ver:
|
||||
update = True
|
||||
elif best["version"] == cur_ver and best["params_b"] > (cur_params or 0) * 1.05:
|
||||
update = True
|
||||
# gleiche Datei schon installiert? dann kein Update
|
||||
if current and best["repo"].split("/")[-1].lower() in (current["name"] or "").lower():
|
||||
update = False
|
||||
return {"current": current, "recommended": best, "update_available": update}
|
||||
|
||||
|
||||
def _reach(url: str, path: str = "") -> bool:
|
||||
try:
|
||||
with httpx.Client(timeout=3.0) as c:
|
||||
|
||||
@@ -9,8 +9,10 @@ spätere Auto-Setups nutzen ihn, damit sie nie auseinanderlaufen.
|
||||
"""
|
||||
|
||||
import json
|
||||
import math
|
||||
import os
|
||||
import time
|
||||
from datetime import datetime
|
||||
|
||||
import httpx
|
||||
|
||||
@@ -46,19 +48,41 @@ def _fetch_author_models(author: str) -> list:
|
||||
return []
|
||||
|
||||
|
||||
def _age_days(last_modified, now_ts: float) -> float:
|
||||
"""Alter eines HF-Modells in Tagen (lastModified ISO). Unbekannt → ~1.5 Jahre."""
|
||||
if not last_modified:
|
||||
return 540.0
|
||||
try:
|
||||
dt = datetime.fromisoformat(str(last_modified).replace("Z", "+00:00"))
|
||||
return max((now_ts - dt.timestamp()) / 86400.0, 0.0)
|
||||
except Exception:
|
||||
return 540.0
|
||||
|
||||
|
||||
def _score(m: dict, now_ts: float) -> float:
|
||||
"""Zukunftssicherer Rang-Score für DIESE Hardware. Kombiniert:
|
||||
- Fit: perfect dominiert (Bonus 3.0 > Summe der übrigen Terme → passt-komfortabel zuerst),
|
||||
- Recency: neuere Generationen bevorzugt (Halbwertszeit ~9 Monate über lastModified),
|
||||
- Capability: mehr Parameter (log-skaliert),
|
||||
- Popularity: Downloads (log-skaliert).
|
||||
So gewinnt bei vergleichbarer Größe die NEUERE Generation (z.B. Qwen3-Coder vor
|
||||
Qwen2.5-Coder), ohne dass kleine Populär-Modelle große verdrängen."""
|
||||
fit_bonus = 3.0 if m["fit"]["level"] == "perfect" else 0.0
|
||||
recency = 0.5 ** (_age_days(m.get("lastModified"), now_ts) / 270.0)
|
||||
cap = math.log2(max(float(m.get("params_b") or 1.0), 1.0) + 1.0) / 8.0
|
||||
pop = math.log10(float(m.get("downloads") or 0) + 1.0) / 7.0
|
||||
return fit_bonus + 1.2 * recency + 1.2 * cap + 0.5 * pop
|
||||
|
||||
|
||||
def rank_runnable(models: list[dict]) -> list[dict]:
|
||||
"""EINE Quelle der Wahrheit fürs Ranking lauffähiger Modelle für DIESE Hardware:
|
||||
1) nur was komfortabel passt (perfect vor marginal, too_tight fliegt raus),
|
||||
2) das FÄHIGSTE zuerst — mehr Parameter = mehr Können (bei MoE bleibt es dank
|
||||
aktiver-Param-Schätzung schnell),
|
||||
3) bei Gleichstand das meistgeladene.
|
||||
So bevorzugt die 128-GB-Box große (MoE-)Modelle statt kleiner Populär-Modelle —
|
||||
und „Modelle finden" schlägt nie ein Downgrade vor (z.B. 35B-A3B → 4B)."""
|
||||
"""EINE Quelle der Wahrheit fürs Ranking lauffähiger Modelle für DIESE Hardware.
|
||||
Nur was passt (too_tight fliegt raus), dann nach `_score` (Fit + Recency + Capability
|
||||
+ Popularity). Bevorzugt neuere, fähige Modelle → zukunftssicher; „Modelle finden"
|
||||
schlägt nie ein Downgrade vor (Downgrade-Sperre zusätzlich in maintenance)."""
|
||||
now_ts = time.time()
|
||||
return sorted(
|
||||
[m for m in models if m["fit"]["level"] != "too_tight"],
|
||||
key=lambda m: (_FIT_ORDER[m["fit"]["level"]],
|
||||
-float(m.get("params_b") or 0.0),
|
||||
-int(m.get("downloads") or 0)),
|
||||
key=lambda m: -_score(m, now_ts),
|
||||
)
|
||||
|
||||
|
||||
|
||||
Reference in New Issue
Block a user