feat(cookbook): autonome Modell-Entdeckung aus vertrauenswuerdigen Quellen
"Aktuell beste Modelle fuer dein System": fragt vertrauenswuerdige HF-Orgs (unsloth/bartowski/ggml-org/lmstudio-community) LIVE ab, kategorisiert die Treffer (vision/coder/reasoning/agent/scout), filtert per hw_math auf das, was auf die Hardware passt, und cached das Ergebnis (TTL 12 h, lazy + Knopf "Aktualisieren"). Damit bleibt das Cookbook von selbst aktuell, ohne dass Modelle hartkodiert werden. - sources.py: TRUSTED_AUTHORS + CATEGORIES + SKIP_TOKENS (reine Daten). - config.py: DISCOVER_CACHE_PATH (persistent neben den Modellen, uebersteht Deploys) + DISCOVER_TTL. - cookbook.py: /api/cookbook/discover (force-Param), refresh_discover, Bestandsabgleich (_model_installed -> "schon installiert als X") und ehrliche Voraussetzungen je Modell (Vision->mmproj/jinja, unquantisiert, zu gross/knapp). - cookbook.js: Sektion mit Kategorien, Fit-Ampel, Downloads, Hinweisen, 1-Klick-Installieren bzw. "installiert"-Markierung; "Aktualisieren"-Knopf. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
This commit is contained in:
Hitonabi
+150
-1
@@ -6,6 +6,7 @@ import httpx
|
||||
import json
|
||||
import os
|
||||
import re
|
||||
import time
|
||||
from fastapi import APIRouter, Depends, HTTPException
|
||||
from pydantic import BaseModel
|
||||
import psutil
|
||||
@@ -14,10 +15,12 @@ from ruamel.yaml.scalarstring import LiteralScalarString
|
||||
|
||||
from auth import auth
|
||||
from hw_math import evaluate_fit, max_ctx_for
|
||||
from config import MODELS_DIR, CMD_TEMPLATE, DEFAULT_TTL, HF_DOWNLOAD_ENV, USER_RECIPES_PATH, hf_bin
|
||||
from config import (MODELS_DIR, CMD_TEMPLATE, DEFAULT_TTL, HF_DOWNLOAD_ENV, USER_RECIPES_PATH,
|
||||
DISCOVER_CACHE_PATH, DISCOVER_TTL, hf_bin)
|
||||
from llamaswap import read_config, write_config
|
||||
from jobengine import start_job, JOBS, attach_download_progress
|
||||
from recipes import RECIPES, UPGRADES
|
||||
from sources import TRUSTED_AUTHORS, CATEGORIES, SKIP_TOKENS
|
||||
|
||||
router = APIRouter(prefix="/api/cookbook", dependencies=[Depends(auth)])
|
||||
|
||||
@@ -286,6 +289,152 @@ def _pick_gguf(repo: str, quant: str = "Q4_K_M") -> str | None:
|
||||
return (pref or nosplit or ggufs)[0]
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Automatische Modell-Entdeckung ("aktuell beste Modelle", live von HF + Cache).
|
||||
# ---------------------------------------------------------------------------
|
||||
def _categorize(repo_id: str) -> str:
|
||||
low = repo_id.lower()
|
||||
for cat in CATEGORIES:
|
||||
if any(k in low for k in cat["kw"]):
|
||||
return cat["role"]
|
||||
return "scout" # Fallback: Allrounder
|
||||
|
||||
|
||||
def _installed_index() -> dict:
|
||||
"""alias -> kleingeschriebene cmd-Zeile der installierten Modelle (fuer Bestandsabgleich)."""
|
||||
return {alias: str(spec.get("cmd", "")).lower()
|
||||
for alias, spec in (read_config().get("models") or {}).items()}
|
||||
|
||||
|
||||
def _model_installed(repo: str, idx: dict) -> str | None:
|
||||
"""Liefert den Alias, unter dem dieses Repo schon installiert ist — sonst None."""
|
||||
base = repo.split("/")[-1].lower()
|
||||
# GGUF-Repos heissen meist '<modell>-GGUF' -> auch ohne Suffix vergleichen
|
||||
stem = base[:-5] if base.endswith("-gguf") else base
|
||||
for alias, cmd in idx.items():
|
||||
if base in cmd or (stem and stem in cmd):
|
||||
return alias
|
||||
return None
|
||||
|
||||
|
||||
def _model_requirements(role: str, quant: str, fit: dict) -> list[str]:
|
||||
"""Ehrliche Voraussetzungen/Hinweise pro Modell (Klartext fuer Anfaenger)."""
|
||||
reqs = []
|
||||
if role == "vision":
|
||||
reqs.append("Bild-Modell: Projektor (mmproj) + --jinja werden beim Einpflegen automatisch ergänzt.")
|
||||
if quant.upper() in ("FP16", "BF16", "F16", "F32"):
|
||||
reqs.append("Unquantisiert — sehr groß. Eine Q4/Q5-Variante ist meist die bessere Wahl.")
|
||||
if fit["level"] == "too_tight":
|
||||
reqs.append("Größer als dein Speicher — nur mit kleinerem Quant realistisch.")
|
||||
elif fit["level"] == "marginal":
|
||||
reqs.append("Läuft, wird aber knapp — andere Programme währenddessen besser schließen.")
|
||||
return reqs
|
||||
|
||||
|
||||
def _fetch_author_models(author: str) -> list:
|
||||
url = (f"https://huggingface.co/api/models?author={author}"
|
||||
f"&filter=gguf&sort=downloads&direction=-1&limit=40")
|
||||
try:
|
||||
with httpx.Client(timeout=12.0) as c:
|
||||
data = c.get(url).json()
|
||||
return data if isinstance(data, list) else []
|
||||
except Exception: # noqa: BLE001
|
||||
return []
|
||||
|
||||
|
||||
def refresh_discover(ram_gb: float) -> dict:
|
||||
"""Quellen live abfragen, kategorisieren, nach Hardware-Fit + Beliebtheit ranken
|
||||
und cachen. Wirft nur, wenn KEINE einzige Quelle erreichbar war."""
|
||||
raw, seen, ok = [], set(), 0
|
||||
for author in TRUSTED_AUTHORS:
|
||||
models = _fetch_author_models(author)
|
||||
if models:
|
||||
ok += 1
|
||||
for m in models:
|
||||
rid = m.get("id")
|
||||
if not rid or rid in seen:
|
||||
continue
|
||||
seen.add(rid)
|
||||
raw.append(m)
|
||||
if ok == 0:
|
||||
raise RuntimeError("Keine Quelle erreichbar.")
|
||||
|
||||
by_cat = {c["role"]: [] for c in CATEGORIES}
|
||||
for m in raw:
|
||||
rid = m["id"]
|
||||
low = rid.lower()
|
||||
if any(tok in low for tok in SKIP_TOKENS):
|
||||
continue
|
||||
role = _categorize(rid)
|
||||
params_b = extract_params_b(rid)
|
||||
quant = "Q4_K_M" # Referenz-Quant fuer die Fit-Einschaetzung
|
||||
fit = evaluate_fit(params_b, quant, 8192, ram_gb)
|
||||
by_cat[role].append({
|
||||
"name": rid.split("/")[-1], "author": rid.split("/")[0], "repo": rid,
|
||||
"role": role, "params_b": params_b, "quant": quant,
|
||||
"downloads": int(m.get("downloads") or 0), "likes": int(m.get("likes") or 0),
|
||||
"lastModified": m.get("lastModified"),
|
||||
"fit": fit, "optimal_ctx": max_ctx_for(params_b, quant, ram_gb),
|
||||
})
|
||||
|
||||
cats = []
|
||||
for c in CATEGORIES:
|
||||
items = by_cat[c["role"]]
|
||||
# Passendes zuerst (perfect/marginal vor too_tight), dann nach Downloads.
|
||||
items.sort(key=lambda x: (x["fit"]["level"] != "too_tight", x["downloads"]), reverse=True)
|
||||
top = items[:4]
|
||||
if top:
|
||||
cats.append({"role": c["role"], "title": c["title"], "icon": c["icon"], "models": top})
|
||||
|
||||
data = {"updated": time.time(), "categories": cats}
|
||||
try:
|
||||
DISCOVER_CACHE_PATH.parent.mkdir(parents=True, exist_ok=True)
|
||||
tmp = DISCOVER_CACHE_PATH.with_name(DISCOVER_CACHE_PATH.name + ".tmp")
|
||||
tmp.write_text(json.dumps(data, ensure_ascii=False, indent=2), encoding="utf-8")
|
||||
os.replace(tmp, DISCOVER_CACHE_PATH)
|
||||
except Exception: # noqa: BLE001
|
||||
pass # Cache ist nur Beschleunigung — Entdeckung funktioniert auch ohne Schreibrecht
|
||||
return data
|
||||
|
||||
|
||||
def _load_discover() -> dict | None:
|
||||
try:
|
||||
if DISCOVER_CACHE_PATH.exists():
|
||||
return json.loads(DISCOVER_CACHE_PATH.read_text(encoding="utf-8"))
|
||||
except Exception: # noqa: BLE001
|
||||
pass
|
||||
return None
|
||||
|
||||
|
||||
@router.get("/discover")
|
||||
def discover(force: bool = False):
|
||||
"""Aktuell beste Modelle je Kategorie — live aus den Quellen (gecacht, TTL).
|
||||
Markiert pro Modell, ob es schon installiert ist, und nennt Voraussetzungen."""
|
||||
ram_gb = psutil.virtual_memory().total / (1024 ** 3)
|
||||
cached = _load_discover()
|
||||
fresh = bool(cached) and (time.time() - cached.get("updated", 0) < DISCOVER_TTL)
|
||||
if fresh and not force:
|
||||
data = cached
|
||||
else:
|
||||
try:
|
||||
data = refresh_discover(ram_gb)
|
||||
except Exception: # noqa: BLE001
|
||||
if cached:
|
||||
data = cached # Quellen down -> alter Cache ist besser als nichts
|
||||
else:
|
||||
raise HTTPException(502, "Modell-Quellen gerade nicht erreichbar — später erneut versuchen.")
|
||||
# Bestand + Voraussetzungen frisch dazurechnen (aendern sich unabhaengig vom Cache).
|
||||
idx = _installed_index()
|
||||
out = []
|
||||
for c in data["categories"]:
|
||||
ms = [{**m, "installed": _model_installed(m["repo"], idx),
|
||||
"requirements": _model_requirements(m["role"], m.get("quant", "Q4_K_M"), m["fit"])}
|
||||
for m in c["models"]]
|
||||
out.append({**c, "models": ms})
|
||||
return {"updated": data["updated"], "categories": out,
|
||||
"sys_ram_gb": round(ram_gb, 1), "stale": not fresh}
|
||||
|
||||
|
||||
def compute_upgrades(ram_gb):
|
||||
"""Liste relevanter Modell-Upgrades für die installierten Modelle (UPGRADES-Map)."""
|
||||
installed = (read_config().get("models") or {})
|
||||
|
||||
Reference in New Issue
Block a user