feat: lower memory dedupe threshold for more aggressive cleaning
This commit is contained in:
root
+179
-179
@@ -1,179 +1,179 @@
|
||||
"""
|
||||
Automatische Modell-Entdeckung ("aktuell beste Modelle"): fragt vertrauenswürdige
|
||||
HF-Orgs live ab, kategorisiert per Stichwort, rankt nach Hardware-Fit + Beliebtheit
|
||||
und cached. Portiert aus Mission Control v1 (cookbook.py-Discover).
|
||||
|
||||
Wichtig (Greenfield-Fix gegen v1): EIN gemeinsamer Ranking-Helfer `rank_runnable`
|
||||
ist die Quelle der Wahrheit — sowohl die „beste Empfehlung" je Kategorie als auch
|
||||
spätere Auto-Setups nutzen ihn, damit sie nie auseinanderlaufen.
|
||||
"""
|
||||
|
||||
import json
|
||||
import math
|
||||
import os
|
||||
import time
|
||||
from datetime import datetime
|
||||
|
||||
import httpx
|
||||
|
||||
import logging
|
||||
|
||||
from config import DISCOVER_CACHE_PATH, DISCOVER_TTL
|
||||
from services import catalog
|
||||
from services.caps import capabilities
|
||||
from services.fit import evaluate_fit, extract_params_b, max_ctx_for
|
||||
from services.sources import CATEGORIES, SKIP_TOKENS, TRUSTED_AUTHORS
|
||||
|
||||
log = logging.getLogger(__name__)
|
||||
|
||||
_FIT_ORDER = {"perfect": 0, "marginal": 1, "too_tight": 2}
|
||||
|
||||
|
||||
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"
|
||||
|
||||
|
||||
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:
|
||||
log.debug("discover: Abfrage für Autor %s fehlgeschlagen", author, exc_info=True)
|
||||
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.
|
||||
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: -_score(m, now_ts),
|
||||
)
|
||||
|
||||
|
||||
def refresh_discover(ram_gb: float) -> dict:
|
||||
"""Quellen live abfragen, kategorisieren, ranken, cachen. Wirft nur, wenn KEINE
|
||||
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 and not catalog.entries():
|
||||
raise RuntimeError("Keine Quelle erreichbar.")
|
||||
|
||||
by_cat: dict[str, list] = {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 für die Fit-Einschätzung
|
||||
fit = evaluate_fit(params_b, quant, 8192, ram_gb, name=rid)
|
||||
tags = [str(t) for t in (m.get("tags") or [])]
|
||||
by_cat[role].append({
|
||||
"name": rid.split("/")[-1], "author": rid.split("/")[0], "repo": rid,
|
||||
"role": role, "params_b": params_b, "quant": quant, "tags": tags,
|
||||
"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),
|
||||
"caps": capabilities(name=rid, hf={"tags": tags}),
|
||||
})
|
||||
|
||||
cats = []
|
||||
for c in CATEGORIES:
|
||||
role = c["role"]
|
||||
# 1) KATALOG zuerst (kuratierte, korrekte Metadaten, MoE-bewusst gerankt) —
|
||||
# macht die Empfehlung präzise statt Namens-Raterei.
|
||||
cat_entries = sorted(catalog.entries_for_role(role),
|
||||
key=lambda e: -catalog.stack_score(e, ram_gb))
|
||||
cat_models = [m for m in (catalog.to_model_dict(e, ram_gb) for e in cat_entries)
|
||||
if m["fit"]["level"] != "too_tight"]
|
||||
# 2) HF-Dynamik als Ergänzung (nicht-kuratierte Funde), dedupliziert.
|
||||
hf_ranked = rank_runnable(by_cat[role])
|
||||
seen = {catalog._norm(m["repo"]) for m in cat_models}
|
||||
extra = [h for h in hf_ranked if catalog._norm(h["repo"]) not in seen]
|
||||
combined = cat_models + extra
|
||||
if combined:
|
||||
cats.append({
|
||||
"role": role, "title": c["title"], "icon": c["icon"],
|
||||
"models": combined[:6],
|
||||
# Empfehlung = bester KURATIERTER Eintrag, sonst beste HF-Fundstelle.
|
||||
"recommended": (cat_models[0]["repo"] if cat_models
|
||||
else (hf_ranked[0]["repo"] if hf_ranked else None)),
|
||||
})
|
||||
|
||||
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:
|
||||
log.debug("discover: Cache-Schreiben fehlgeschlagen (nur Beschleunigung)", exc_info=True)
|
||||
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:
|
||||
log.debug("discover: Cache-Lesen fehlgeschlagen", exc_info=True)
|
||||
return None
|
||||
|
||||
|
||||
def safe_discover(ram_gb: float) -> dict | None:
|
||||
"""Aus Cache (wenn frisch) oder live; wirft nie — None wenn nichts da."""
|
||||
cached = load_discover()
|
||||
if cached and (time.time() - cached.get("updated", 0) < DISCOVER_TTL):
|
||||
return cached
|
||||
try:
|
||||
return refresh_discover(ram_gb)
|
||||
except Exception:
|
||||
log.warning("discover: Live-Refresh fehlgeschlagen, nutze Cache", exc_info=True)
|
||||
return cached
|
||||
"""
|
||||
Automatische Modell-Entdeckung ("aktuell beste Modelle"): fragt vertrauenswürdige
|
||||
HF-Orgs live ab, kategorisiert per Stichwort, rankt nach Hardware-Fit + Beliebtheit
|
||||
und cached. Portiert aus Mission Control v1 (cookbook.py-Discover).
|
||||
|
||||
Wichtig (Greenfield-Fix gegen v1): EIN gemeinsamer Ranking-Helfer `rank_runnable`
|
||||
ist die Quelle der Wahrheit — sowohl die „beste Empfehlung" je Kategorie als auch
|
||||
spätere Auto-Setups nutzen ihn, damit sie nie auseinanderlaufen.
|
||||
"""
|
||||
|
||||
import json
|
||||
import math
|
||||
import os
|
||||
import time
|
||||
from datetime import datetime
|
||||
|
||||
import httpx
|
||||
|
||||
import logging
|
||||
|
||||
from config import DISCOVER_CACHE_PATH, DISCOVER_TTL
|
||||
from services import catalog
|
||||
from services.caps import capabilities
|
||||
from services.fit import evaluate_fit, extract_params_b, max_ctx_for
|
||||
from services.sources import CATEGORIES, SKIP_TOKENS, TRUSTED_AUTHORS
|
||||
|
||||
log = logging.getLogger(__name__)
|
||||
|
||||
_FIT_ORDER = {"perfect": 0, "marginal": 1, "too_tight": 2}
|
||||
|
||||
|
||||
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"
|
||||
|
||||
|
||||
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:
|
||||
log.debug("discover: Abfrage für Autor %s fehlgeschlagen", author, exc_info=True)
|
||||
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.
|
||||
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: -_score(m, now_ts),
|
||||
)
|
||||
|
||||
|
||||
def refresh_discover(ram_gb: float) -> dict:
|
||||
"""Quellen live abfragen, kategorisieren, ranken, cachen. Wirft nur, wenn KEINE
|
||||
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 and not catalog.entries():
|
||||
raise RuntimeError("Keine Quelle erreichbar.")
|
||||
|
||||
by_cat: dict[str, list] = {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 für die Fit-Einschätzung
|
||||
fit = evaluate_fit(params_b, quant, 8192, ram_gb, name=rid)
|
||||
tags = [str(t) for t in (m.get("tags") or [])]
|
||||
by_cat[role].append({
|
||||
"name": rid.split("/")[-1], "author": rid.split("/")[0], "repo": rid,
|
||||
"role": role, "params_b": params_b, "quant": quant, "tags": tags,
|
||||
"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),
|
||||
"caps": capabilities(name=rid, hf={"tags": tags}),
|
||||
})
|
||||
|
||||
cats = []
|
||||
for c in CATEGORIES:
|
||||
role = c["role"]
|
||||
# 1) KATALOG zuerst (kuratierte, korrekte Metadaten, MoE-bewusst gerankt) —
|
||||
# macht die Empfehlung präzise statt Namens-Raterei.
|
||||
cat_entries = sorted(catalog.entries_for_role(role),
|
||||
key=lambda e: -catalog.stack_score(e, ram_gb))
|
||||
cat_models = [m for m in (catalog.to_model_dict(e, ram_gb) for e in cat_entries)
|
||||
if m["fit"]["level"] != "too_tight"]
|
||||
# 2) HF-Dynamik als Ergänzung (nicht-kuratierte Funde), dedupliziert.
|
||||
hf_ranked = rank_runnable(by_cat[role])
|
||||
seen = {catalog._norm(m["repo"]) for m in cat_models}
|
||||
extra = [h for h in hf_ranked if catalog._norm(h["repo"]) not in seen]
|
||||
combined = cat_models + extra
|
||||
if combined:
|
||||
cats.append({
|
||||
"role": role, "title": c["title"], "icon": c["icon"],
|
||||
"models": combined[:6],
|
||||
# Empfehlung = bester KURATIERTER Eintrag, sonst beste HF-Fundstelle.
|
||||
"recommended": (cat_models[0]["repo"] if cat_models
|
||||
else (hf_ranked[0]["repo"] if hf_ranked else None)),
|
||||
})
|
||||
|
||||
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:
|
||||
log.debug("discover: Cache-Schreiben fehlgeschlagen (nur Beschleunigung)", exc_info=True)
|
||||
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:
|
||||
log.debug("discover: Cache-Lesen fehlgeschlagen", exc_info=True)
|
||||
return None
|
||||
|
||||
|
||||
def safe_discover(ram_gb: float) -> dict | None:
|
||||
"""Aus Cache (wenn frisch) oder live; wirft nie — None wenn nichts da."""
|
||||
cached = load_discover()
|
||||
if cached and (time.time() - cached.get("updated", 0) < DISCOVER_TTL):
|
||||
return cached
|
||||
try:
|
||||
return refresh_discover(ram_gb)
|
||||
except Exception:
|
||||
log.warning("discover: Live-Refresh fehlgeschlagen, nutze Cache", exc_info=True)
|
||||
return cached
|
||||
|
||||
Reference in New Issue
Block a user