Fix: KV-Schaetzung kalibriert + Brain-Fit-Check brain-spezifisch
- fit.estimate_memory_gb: KV-Cache jetzt sqrt-skaliert (nicht linear mit Gesamt-Params),
kalibriert an Hermes-14B@128K ~19GB KV -> realistische Footprints (vorher massive Ueberschaetzung).
- agent.hermes_brain_info Budget: prueft jetzt Brain (immer resident) + groesstes on-demand-Modell
<= GTT-Budget (fast/vision duerfen verdraengt werden) -> brain-spezifische, aussagekraeftige Warnung.
- Cockpit: Budget-Zeile + Confirm-Warnung entsprechend ("Brain ~X GB + groesstes on-demand ~Y GB").
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
This commit is contained in:
Hitonabi
@@ -107,21 +107,26 @@ def hermes_brain_info() -> dict:
|
|||||||
|
|
||||||
cur_name = cur["name"] if cur else None
|
cur_name = cur["name"] if cur else None
|
||||||
brain_ctx = int((cur.get("ctx") if cur else None) or 32768)
|
brain_ctx = int((cur.get("ctx") if cur else None) or 32768)
|
||||||
# Always-Warm = persist-Modelle, das Brain durch die Empfehlung ersetzt
|
|
||||||
warm = sum(_foot(m) for m in models if m["name"] in persist and m["name"] != cur_name)
|
|
||||||
if best:
|
if best:
|
||||||
warm += estimate_memory_gb(float(best["params_b"]), "Q4_K_M", brain_ctx)
|
brain_gb = estimate_memory_gb(float(best["params_b"]), "Q4_K_M", brain_ctx)
|
||||||
elif cur:
|
elif cur:
|
||||||
warm += _foot(cur)
|
brain_gb = _foot(cur)
|
||||||
|
else:
|
||||||
|
brain_gb = 0.0
|
||||||
|
# voller Always-Warm-Footprint (alle persist, Brain=Empfehlung) — nur Info
|
||||||
|
warm = brain_gb + sum(_foot(m) for m in models
|
||||||
|
if m["name"] in persist and m["name"] != cur_name)
|
||||||
largest_od = max((_foot(m) for m in models if m["name"] not in persist), default=0.0)
|
largest_od = max((_foot(m) for m in models if m["name"] not in persist), default=0.0)
|
||||||
gtt = _gtt_budget_gb()
|
gtt = _gtt_budget_gb()
|
||||||
free_after = gtt - warm
|
# Brain muss immer resident sein → passt Brain + größtes on-demand zusammen?
|
||||||
|
# (fast/vision dürfen beim Laden eines großen Modells verdrängt werden.)
|
||||||
budget = {
|
budget = {
|
||||||
"gtt_gb": gtt,
|
"gtt_gb": gtt,
|
||||||
|
"brain_gb": round(brain_gb, 1),
|
||||||
"warm_projected_gb": round(warm, 1),
|
"warm_projected_gb": round(warm, 1),
|
||||||
"free_after_gb": round(free_after, 1),
|
|
||||||
"largest_ondemand_gb": round(largest_od, 1),
|
"largest_ondemand_gb": round(largest_od, 1),
|
||||||
"fits": free_after >= largest_od,
|
"fits": (brain_gb + largest_od) <= gtt,
|
||||||
|
"free_after_gb": round(gtt - brain_gb - largest_od, 1),
|
||||||
}
|
}
|
||||||
except Exception:
|
except Exception:
|
||||||
log.debug("hermes_brain_info: Budget-Berechnung fehlgeschlagen", exc_info=True)
|
log.debug("hermes_brain_info: Budget-Berechnung fehlgeschlagen", exc_info=True)
|
||||||
|
|||||||
@@ -15,10 +15,13 @@ QUANT_BYTES_PER_PARAM = {
|
|||||||
|
|
||||||
|
|
||||||
def estimate_memory_gb(params_b: float, quant: str, ctx: int) -> float:
|
def estimate_memory_gb(params_b: float, quant: str, ctx: int) -> float:
|
||||||
"""Geschätzter Speicherbedarf in GB (Gewichte + Kontext-KV)."""
|
"""Geschätzter Speicherbedarf in GB (Gewichte + Kontext-KV).
|
||||||
|
KV-Cache skaliert NICHT linear mit den Gesamt-Parametern (er hängt an
|
||||||
|
Layern × KV-Heads, gedämpft durch GQA) → sqrt-Skalierung, kalibriert am
|
||||||
|
gemessenen Punkt Hermes-4-14B @ 128K ≈ 19 GB KV."""
|
||||||
bpp = QUANT_BYTES_PER_PARAM.get(quant.upper(), 0.65)
|
bpp = QUANT_BYTES_PER_PARAM.get(quant.upper(), 0.65)
|
||||||
weights = params_b * bpp
|
weights = params_b * bpp
|
||||||
context_vram = (ctx / 8192) * (max(params_b, 7) / 7) * 0.8
|
context_vram = (ctx / 8192) * (max(params_b, 7) / 7) ** 0.5 * 0.84
|
||||||
return weights + context_vram
|
return weights + context_vram
|
||||||
|
|
||||||
|
|
||||||
|
|||||||
+19
-19
File diff suppressed because one or more lines are too long
Vendored
+1
-1
@@ -7,7 +7,7 @@
|
|||||||
<link rel="manifest" href="/manifest.webmanifest" />
|
<link rel="manifest" href="/manifest.webmanifest" />
|
||||||
<link rel="icon" type="image/svg+xml" href="/favicon.svg" />
|
<link rel="icon" type="image/svg+xml" href="/favicon.svg" />
|
||||||
<title>Mission Control 2.0</title>
|
<title>Mission Control 2.0</title>
|
||||||
<script type="module" crossorigin src="/assets/index-eiqRvBvu.js"></script>
|
<script type="module" crossorigin src="/assets/index-lvQamTQ3.js"></script>
|
||||||
<link rel="stylesheet" crossorigin href="/assets/index-D4NZ6il2.css">
|
<link rel="stylesheet" crossorigin href="/assets/index-D4NZ6il2.css">
|
||||||
</head>
|
</head>
|
||||||
<body>
|
<body>
|
||||||
|
|||||||
@@ -317,10 +317,11 @@ export interface HermesBrainCandidate {
|
|||||||
|
|
||||||
export interface HermesBrainBudget {
|
export interface HermesBrainBudget {
|
||||||
gtt_gb: number
|
gtt_gb: number
|
||||||
warm_projected_gb: number // Always-On-Footprint mit dem empfohlenen Brain
|
brain_gb: number // Footprint des (empfohlenen) Brains, das immer resident bleibt
|
||||||
free_after_gb: number
|
warm_projected_gb: number // alle persist-Modelle warm (Info)
|
||||||
largest_ondemand_gb: number // größtes on-demand-Modell (z.B. heavy)
|
free_after_gb: number // frei nach Brain + größtem on-demand
|
||||||
fits: boolean // passt das größte on-demand daneben?
|
largest_ondemand_gb: number // größtes on-demand-Modell (z.B. heavy/coder)
|
||||||
|
fits: boolean // passt Brain + größtes on-demand zusammen ins Budget?
|
||||||
}
|
}
|
||||||
|
|
||||||
export interface HermesBrainResp {
|
export interface HermesBrainResp {
|
||||||
|
|||||||
@@ -179,7 +179,7 @@ export function Cockpit() {
|
|||||||
async function handleBrainUpdate(repo: string) {
|
async function handleBrainUpdate(repo: string) {
|
||||||
const b = brain?.budget
|
const b = brain?.budget
|
||||||
const warn = b && !b.fits
|
const warn = b && !b.fits
|
||||||
? `\n\n⚠ Speicher-Warnung: Als Always-On würde dieses Brain ~${b.warm_projected_gb} GB belegen (Budget ${b.gtt_gb} GB) → nur ${b.free_after_gb} GB frei. Das größte on-demand-Modell (~${b.largest_ondemand_gb} GB) passt dann NICHT mehr gleichzeitig daneben. Erwäge ein kleineres Brain oder weniger Kontext.`
|
? `\n\n⚠ Speicher-Warnung: Dieses Brain (~${b.brain_gb} GB) muss immer resident sein. Zusammen mit dem größten on-demand-Modell (~${b.largest_ondemand_gb} GB) sprengt das das Budget (${b.gtt_gb} GB) → heavy/coder würden das Brain verdrängen. Erwäge ein kleineres Brain oder weniger Kontext.`
|
||||||
: ""
|
: ""
|
||||||
showConfirm(
|
showConfirm(
|
||||||
"Agent-Hirn aktualisieren?",
|
"Agent-Hirn aktualisieren?",
|
||||||
@@ -709,11 +709,11 @@ export function Cockpit() {
|
|||||||
brain.budget.fits ? "text-muted-foreground/80" : "text-red-400 font-semibold")}>
|
brain.budget.fits ? "text-muted-foreground/80" : "text-red-400 font-semibold")}>
|
||||||
<HardDrive className="h-3 w-3 shrink-0 mt-0.5" />
|
<HardDrive className="h-3 w-3 shrink-0 mt-0.5" />
|
||||||
<span>
|
<span>
|
||||||
{brain.update_available ? "Mit diesem Brain als Always-On: " : "Always-On-Budget: "}
|
Always-On-Brain ~{brain.budget.brain_gb} GB + größtes on-demand (~{brain.budget.largest_ondemand_gb} GB)
|
||||||
~{brain.budget.warm_projected_gb} / {brain.budget.gtt_gb} GB belegt → {brain.budget.free_after_gb} GB frei
|
= {(brain.budget.brain_gb + brain.budget.largest_ondemand_gb).toFixed(1)} / {brain.budget.gtt_gb} GB
|
||||||
{brain.budget.fits
|
{brain.budget.fits
|
||||||
? ` · größtes on-demand-Modell (~${brain.budget.largest_ondemand_gb} GB) passt daneben ✓`
|
? " · passt gleichzeitig ✓"
|
||||||
: ` · größtes on-demand-Modell (~${brain.budget.largest_ondemand_gb} GB) passt NICHT mehr daneben ⚠ — kleineres Brain/weniger Kontext`}
|
: " · passt NICHT gleichzeitig ⚠ — heavy/coder würden das Brain verdrängen. Kleineres Brain oder weniger Kontext."}
|
||||||
</span>
|
</span>
|
||||||
</div>
|
</div>
|
||||||
)}
|
)}
|
||||||
|
|||||||
Reference in New Issue
Block a user