Menu
Quick reference for src/vramgeist/calc.py — formulas, public functions, gotchas. Use before editing calc.py.
Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Keep input small. Use before any Read or Grep call when context budget tight.
Quick reference for src/vramgeist/gguf.py parser. Use before editing gguf.py or adding new metadata keys.
Pick next unfinished task from TASK_BOARD.md. Use at start of every session and after finishing one task.
One-page repo layout. Use to locate files without globbing. First skill to invoke after pick-task.
Required end-of-response format. Use before submitting any response that edited files.
Break a Tier 3 (🔴) task into 3-5 Tier-2-shaped sub-tasks before editing. Use whenever picked task is Tier 3.
Basé sur la classification professionnelle SOC
| name | calc-quickref |
| description | Quick reference for src/vramgeist/calc.py — formulas, public functions, gotchas. Use before editing calc.py. |
Avoid full re-read of calc.py (554 lines). Use this map.
| Function | Returns | Purpose |
|---|---|---|
calculate_vram_usage(model_mb, n_layers, n_gpu_layers, ctx, cfg) | float MB | Legacy approx |
calculate_ram_usage(...) | float MB | CPU side |
calculate_total_memory_usage(...) | (vram, ram) | Pair |
_per_token_kv_bytes(model_meta, bpe_override) | int bytes | GQA-aware KV/token |
_estimate_tps(ctx, n_layers, hidden, bpe, bw_gbps, beta, model_meta) | float | TPS via bandwidth |
fit_latency_model(measure_func, contexts, runs) | (a, b, r2) | Linear fit |
validate_context(measure_func, C, attempts, step) | (C, tries, ok) | Alloc backoff |
calculate_semantic_throughput_best_context(...) | dict | Score-based picker |
calculate_max_context(...) | int | Memory-bound max ctx |
Model_VRAM = model_mb * min(n_gpu/n_layers, 1.0)
Context_VRAM = ctx * hidden * 2 * bpe / 1024² (legacy, no GQA)
KV_per_token = 2 * n_kv_heads * head_dim * n_layers * bpe (preferred)
kv_bytes = ctx * per_token * batch * (1 + overhead_frac)
TPS = bw_bytes_per_sec / (beta * per_token)
score = tps * usefulness * mem_margin^memory_penalty_pow
usefulness = 1 / (1 + ctx / c_ref)
config.hidden_size (4096 default), config.bytes_per_element (2),
config.vram_overhead_mb (~500), config.vram_safety_margin (~0.9),
config.ram_overhead_mb, config.ram_safety_margin.
model_meta path differs from legacy path — branches on if model_meta:._per_token_kv_bytes falls back chain: explicit → infer from hidden/n_heads → head_dim=64 default.8388608. Returned ctx is rounded DOWN to one of them.estimate_free_vram_bytes may raise → fallback uses available_vram_mb * safety_margin.verify_alloc opt defaults True → calls measure_func; tests pass None.Pure functions. No I/O, no prints. Side-effect-free. If adding logic, keep that invariant.