Hitonabi
6d529e3f79
max_ctx_in_budget schaetzte den KV-Cache LINEAR mit den Params, waehrend Footprint/Fit sqrt rechnen (kalibriert an Hermes-14B@128K~19GB). Folge: fuer grosse Modelle viel zu konservativ -> setup_aware_ctx schlug z.B. fuer heavy-122B 8192 vor, obwohl 32768 real passt. Jetzt exakte Inverse der Footprint-Formel (sqrt*0.84) -> heavy bekommt ~49k statt 8k, keine faelschlichen Reduktionen mehr. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
107 lines
4.4 KiB
Python
107 lines
4.4 KiB
Python
"""
|
||
Hardware-Fit-Mathe (VRAM/RAM, tps-Schätzung) für APUs mit Unified Memory
|
||
(Bosgame M5 / Strix Halo). Portiert aus Mission Control v1 (hw_math.py).
|
||
"""
|
||
|
||
import re
|
||
|
||
# Bytes pro Parameter je GGUF-Quant (Annahme).
|
||
QUANT_BYTES_PER_PARAM = {
|
||
"Q2_K": 0.35, "Q3_K_S": 0.38, "Q3_K_M": 0.42, "Q3_K_L": 0.45,
|
||
"Q4_0": 0.50, "Q4_1": 0.55, "Q4_K_S": 0.50, "Q4_K_M": 0.55,
|
||
"Q5_0": 0.62, "Q5_1": 0.68, "Q5_K_S": 0.62, "Q5_K_M": 0.65,
|
||
"Q6_K": 0.75, "Q8_0": 1.00, "F16": 2.00, "BF16": 2.00,
|
||
"MXFP4": 0.55, "FP8": 1.05, "AWQ": 0.55,
|
||
}
|
||
|
||
|
||
def estimate_memory_gb(params_b: float, quant: str, ctx: int) -> float:
|
||
"""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)
|
||
weights = params_b * bpp
|
||
context_vram = (ctx / 8192) * (max(params_b, 7) / 7) ** 0.5 * 0.84
|
||
return weights + context_vram
|
||
|
||
|
||
def extract_active_params_b(name: str) -> float | None:
|
||
"""Aktive Parameter bei MoE ('30B-A3B' → 3.0). None bei Dense."""
|
||
m = re.search(r"(?<![a-zA-Z])a(\d+(?:\.\d+)?)b\b", name.lower())
|
||
return float(m.group(1)) if m else None
|
||
|
||
|
||
def estimate_speed(req_gb: float, sys_ram_gb: float, moe_active_ratio: float = 1.0) -> float:
|
||
"""Geschätzte t/s anhand der ~273 GB/s Bandbreite der APU.
|
||
moe_active_ratio = aktive/gesamt Params; < 1 bei MoE."""
|
||
bw = 273 if sys_ram_gb > 8 else 70
|
||
if req_gb <= 0:
|
||
return 0.0
|
||
raw_tps = (bw / req_gb) * 0.55
|
||
if moe_active_ratio < 0.8:
|
||
raw_tps *= (1.0 / moe_active_ratio) ** 0.5
|
||
return raw_tps
|
||
|
||
|
||
def evaluate_fit(params_b: float, quant: str, ctx: int, sys_ram_gb: float, name: str = "") -> dict:
|
||
"""Fit für ein Shared-Memory-System (APU). name → MoE-Erkennung (optional)."""
|
||
req_gb = estimate_memory_gb(params_b, quant, ctx)
|
||
active_b = extract_active_params_b(name) if name else None
|
||
moe_ratio = (active_b / params_b) if (active_b and params_b > 0) else 1.0
|
||
tps = estimate_speed(req_gb, sys_ram_gb, moe_ratio)
|
||
usable_ram = max(sys_ram_gb - 4.0, 0)
|
||
if req_gb > usable_ram:
|
||
fit_level, text = "too_tight", "Zu groß (OOM)"
|
||
elif req_gb > usable_ram * 0.8:
|
||
fit_level, text = "marginal", "Könnte knapp werden"
|
||
else:
|
||
fit_level, text = "perfect", "Passt perfekt"
|
||
return {"level": fit_level, "text": text, "req_gb": round(req_gb, 1), "tps": round(tps, 0)}
|
||
|
||
|
||
def extract_params_b(name: str) -> float:
|
||
"""Parametergröße (Mrd.) aus Repo-/Dateiname. 8x7B (MoE) → 56."""
|
||
moe = re.search(r"(\d+)x(\d+(?:\.\d+)?)[bB]", name)
|
||
if moe:
|
||
return float(moe.group(1)) * float(moe.group(2))
|
||
m = re.search(r"(\d+(?:\.\d+)?)[bB](?![a-zA-Z])", name)
|
||
return float(m.group(1)) if m else 7.0
|
||
|
||
|
||
_NICE_CTX = [2048, 4096, 8192, 16384, 32768, 49152, 65536, 98304, 131072]
|
||
|
||
|
||
def max_ctx_in_budget(params_b: float, quant: str, budget_gb: float) -> int:
|
||
"""Größter 'schöner' Kontext, dessen Gewichte + KV in budget_gb passen.
|
||
Budget-basierter Kern → wird von der setup-bewussten ctx-Vergabe
|
||
(services.budget) mit dem ECHTEN freien Budget gefüttert."""
|
||
bpp = QUANT_BYTES_PER_PARAM.get(quant.upper(), 0.65)
|
||
weights = params_b * bpp
|
||
ctx_budget = budget_gb - weights
|
||
if ctx_budget <= 0:
|
||
return 2048
|
||
# KV pro 8k — EXAKTE Inverse von estimate_memory_gb (sqrt, kalibriert an
|
||
# Hermes-14B@128K≈19GB). Vorher linear → für große Modelle viel zu konservativ.
|
||
per_8k = (max(params_b, 7) / 7) ** 0.5 * 0.84
|
||
raw_ctx = (ctx_budget / per_8k) * 8192
|
||
best = _NICE_CTX[0]
|
||
for c in _NICE_CTX:
|
||
if c <= raw_ctx:
|
||
best = c
|
||
return best
|
||
|
||
|
||
def max_ctx_for(params_b: float, quant: str, sys_ram_gb: float) -> int:
|
||
"""Roh-Obergrenze: größter Kontext für dieses Modell ALLEIN gegen den
|
||
Gesamt-RAM (80 % nutzbar). Ignoriert bewusst das übrige Setup —
|
||
setup-bewusst rechnet services.budget.setup_aware_ctx."""
|
||
return max_ctx_in_budget(params_b, quant, max(sys_ram_gb - 4.0, 0) * 0.8)
|
||
|
||
|
||
def recommend_ctx(params_b: float, quant: str, sys_ram_gb: float) -> dict:
|
||
ctx = max_ctx_for(params_b, quant, sys_ram_gb)
|
||
k = ctx // 1024
|
||
return {"ctx": ctx, "k": k,
|
||
"note": f"Bis ~{k}k Kontext passt komfortabel auf deine Hardware ({round(sys_ram_gb)} GB)."}
|