{
 "modeles": [
  {
   "nom": "Qwen2.5-0.5B",
   "source": "qwen25-05b-sft-q8.gguf",
   "positions": "rotary",
   "largeur": 896,
   "tetes": 14,
   "tetes_kv": 2,
   "taille_tete": 64,
   "couches": 24,
   "ctx": 32768,
   "base": 1000000.0,
   "dims_tournees": 64,
   "ffn": 4864,
   "octets_par_jeton": 12288,
   "cache_plein_go": 0.402653184,
   "cases": 360777252864,
   "onde": 4080185.126271958,
   "marge": 124.51736835546747,
   "table_positions": 0
  },
  {
   "nom": "Qwen2.5-1.5B",
   "source": "qwen25-15b-sft-q8.gguf",
   "positions": "rotary",
   "largeur": 1536,
   "tetes": 12,
   "tetes_kv": 2,
   "taille_tete": 128,
   "couches": 28,
   "ctx": 131072,
   "base": 1000000.0,
   "dims_tournees": 128,
   "ffn": 8960,
   "octets_par_jeton": 28672,
   "cache_plein_go": 3.758096384,
   "cases": 5772436045824,
   "onde": 5063255.794048396,
   "marge": 38.62957606543271,
   "table_positions": 0
  },
  {
   "nom": "Pythia-160M",
   "source": "config.json",
   "positions": "rotary",
   "largeur": 768,
   "tetes": 12,
   "tetes_kv": 12,
   "taille_tete": 64,
   "couches": 12,
   "ctx": 2048,
   "base": 10000,
   "dims_tournees": 16,
   "ffn": 3072,
   "octets_par_jeton": 36864,
   "cache_plein_go": 0.075497472,
   "cases": 603979776,
   "onde": 19869.176531592202,
   "marge": 9.701746353316505,
   "table_positions": 0
  },
  {
   "nom": "run-034 (le mien)",
   "source": "run-034-moe/train.py",
   "positions": "apprises",
   "largeur": 384,
   "tetes": 6,
   "tetes_kv": 6,
   "taille_tete": 64,
   "couches": 8,
   "ctx": 256,
   "base": null,
   "dims_tournees": null,
   "ffn": 1536,
   "octets_par_jeton": 12288,
   "cache_plein_go": 0.003145728,
   "cases": 3145728,
   "onde": null,
   "marge": null,
   "table_positions": 98304
  },
  {
   "nom": "run-062 (le mien)",
   "source": "run-062-tetes-largeur/train.py",
   "positions": "apprises",
   "largeur": 256,
   "tetes": 4,
   "tetes_kv": 4,
   "taille_tete": 64,
   "couches": 4,
   "ctx": 128,
   "base": null,
   "dims_tournees": null,
   "ffn": 1024,
   "octets_par_jeton": 4096,
   "cache_plein_go": 0.000524288,
   "cases": 262144,
   "onde": null,
   "marge": null,
   "table_positions": 32768
  }
 ],
 "gqa": {
  "avec": 12288,
  "sans": 86016,
  "facteur": 7.0,
  "go_avec": 0.402653184,
  "go_sans": 2.818572288
 },
 "note": "Valeurs lues dans les GGUF, le config.json de Pythia et mes train.py. Cache KV compté en 16 bits. Aucune n'est saisie à la main."
}