Michael and Claude Opus 5.5
9aa426a389
Moteur : trois réglages d'expert sur demande — marge VRAM par carte (FIT_TARGET), seuil d'offload des experts, branches Q/K/V parallèles
...
Sur deux cartes inégales (5060 Ti + 3060) ou un MoE aux experts sur CPU, il
reste quelques leviers de placement que llama.cpp expose mais que Loki ne
savait pas poser. Aucun ne touche au modèle (poids, contexte, cache,
échantillonnage) ; tous restent éteints tant que le preset ne les demande pas,
et le gain se décide à la mesure.
- FIT_TARGET=1024,3072 → --fit-target : marge libre par carte pour --fit.
Validée (entiers, 8 valeurs au plus, jamais transmise brute : stoull ferait
mourir le moteur en boucle), remontée à 1024 Mio au minimum. Seulement si
l'aide la liste, sans -fitt ni LLAMA_ARG_FIT_TARGET déjà posés, avec un
contexte chiffré (CTX=0 laisserait fit réduire le contexte) et si --fit
tournera vraiment : NGL chiffré, --tensor-split, -ot, --n-cpu-moe, -cmoe,
--fit off ou -sm row le coupent, et on le dit au lieu de ne rien faire.
- OP_OFFLOAD_MIN_BATCH=N → GGML_OP_OFFLOAD_MIN_BATCH : les lots de moins de N
jetons restent sur CPU pour les poids en RAM. Utile pour des brouillons
ngram de 32 jetons ou plus ; MTP vérifie déjà sous 32.
- CUDA_GRAPH_OPT=on → GGML_CUDA_GRAPH_OPT=1 : expérimental, sortie identique en
amont, gain de 0 à quelques % ; refusé si -sm row/tensor ou un -ot sur
l'attention pouvait séparer Q, K et V d'une couche. Le lancement rappelle
« off puis redémarrer » en cas de GGML_ASSERT.
- une variable déjà dans l'environnement n'est jamais touchée ; elles ne vont
qu'à llama-server via « loki serve ».
- tensorOverride, extrait de pipelineBlocker, sert aussi à la garde de --fit.
- --backend-sampling écarté : pas identique au bit près, et inactif sur les
requêtes à outils (grammaire) qui font l'essentiel du trafic.
- clés documentées dans le squelette de « loki edit » ; tests de table.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com >
2026-10-03 22:53:44 +02:00
..
2026-10-02 23:30:00 +02:00
2026-10-03 22:43:04 +02:00
2026-08-14 21:33:15 +00:00
2026-08-20 09:31:59 +00:00
2026-08-14 21:37:51 +00:00
2026-08-16 11:02:35 +00:00
2026-08-14 21:37:51 +00:00
2026-08-20 09:31:59 +00:00
2026-08-16 20:43:47 +00:00
2026-08-18 21:58:04 +02:00
2026-08-16 20:43:47 +00:00
2026-10-02 23:13:56 +02:00
2026-10-02 23:13:56 +02:00
2026-08-14 21:33:15 +00:00
2026-08-20 10:32:44 +00:00
2026-08-20 10:32:44 +00:00
2026-08-14 21:33:15 +00:00
2026-08-14 21:33:15 +00:00
2026-08-19 08:28:53 +02:00
2026-08-14 21:33:15 +00:00
2026-08-29 17:53:14 +02:00
2026-09-12 20:20:33 +00:00
2026-09-12 20:20:33 +00:00
2026-09-12 20:20:33 +00:00
2026-08-14 21:33:15 +00:00
2026-09-12 20:20:33 +00:00
2026-08-14 21:33:15 +00:00
2026-08-14 21:33:15 +00:00
2026-08-16 20:43:47 +00:00
2026-10-03 22:43:04 +02:00
2026-10-03 22:43:04 +02:00
2026-08-20 09:31:59 +00:00
2026-10-03 22:37:59 +02:00
2026-10-03 22:53:44 +02:00
2026-10-03 22:53:44 +02:00
2026-10-03 22:27:04 +02:00
2026-10-03 22:47:38 +02:00
2026-10-03 22:53:44 +02:00
2026-10-03 22:37:59 +02:00
2026-10-03 22:37:59 +02:00
2026-10-03 22:53:44 +02:00
2026-09-22 13:59:10 +00:00
2026-09-22 13:59:10 +00:00
2026-09-22 13:19:37 +00:00
2026-08-14 21:33:15 +00:00
2026-08-15 11:54:52 +00:00
2026-10-02 23:23:21 +02:00
2026-10-02 23:38:07 +02:00
2026-10-02 23:38:07 +02:00
2026-09-12 20:25:48 +00:00
2026-10-02 23:07:46 +02:00
2026-10-02 23:36:51 +02:00
2026-08-16 15:33:55 +00:00
2026-09-22 13:59:10 +00:00
2026-10-02 23:04:30 +02:00
2026-10-02 23:04:30 +02:00
2026-08-30 15:47:10 +02:00
2026-08-30 15:47:10 +02:00
2026-10-02 23:11:40 +02:00
2026-10-02 23:10:56 +02:00
2026-10-02 23:10:56 +02:00
2026-10-02 23:08:46 +02:00
2026-08-14 21:33:15 +00:00
2026-10-02 23:04:51 +02:00
2026-10-02 23:04:51 +02:00
2026-08-30 15:47:10 +02:00
2026-10-02 23:05:08 +02:00
2026-09-22 13:33:39 +00:00
2026-10-02 23:11:40 +02:00
2026-09-07 22:34:33 +02:00
2026-09-22 13:59:10 +00:00
2026-09-07 22:33:55 +02:00
2026-10-02 23:13:56 +02:00
2026-09-22 13:33:39 +00:00
2026-09-22 13:33:39 +00:00
2026-10-02 23:08:46 +02:00
2026-10-02 23:34:42 +02:00
2026-10-02 23:34:42 +02:00
2026-08-14 21:33:15 +00:00
2026-08-14 21:33:15 +00:00
2026-10-02 23:34:42 +02:00
2026-09-12 21:05:19 +00:00
2026-09-22 13:19:37 +00:00
2026-08-16 15:33:55 +00:00
2026-08-20 09:31:59 +00:00
2026-08-14 21:33:15 +00:00
2026-08-14 21:47:57 +00:00
2026-10-02 23:32:29 +02:00
2026-10-02 23:32:29 +02:00
2026-08-18 21:58:57 +02:00
2026-10-02 23:30:00 +02:00
2026-10-02 23:38:07 +02:00
2026-10-02 23:36:51 +02:00
2026-10-02 23:36:51 +02:00
2026-08-20 09:31:59 +00:00
2026-08-18 21:58:57 +02:00
2026-08-20 09:31:59 +00:00
2026-08-18 21:58:57 +02:00
2026-08-20 09:31:59 +00:00
2026-10-02 23:40:38 +02:00
2026-10-02 23:31:33 +02:00
2026-10-02 23:31:33 +02:00
2026-08-18 21:58:57 +02:00
2026-09-22 14:08:01 +00:00
2026-10-02 23:36:01 +02:00
2026-10-02 23:36:01 +02:00
2026-08-20 09:31:59 +00:00
2026-08-18 21:58:57 +02:00
2026-10-02 23:30:00 +02:00
2026-10-02 23:30:00 +02:00
2026-10-02 23:36:51 +02:00
2026-09-22 13:19:37 +00:00
2026-09-22 13:19:37 +00:00
2026-09-22 13:19:37 +00:00
2026-09-22 13:19:37 +00:00
2026-09-07 22:34:57 +02:00
2026-09-07 22:34:57 +02:00
2026-09-07 22:34:57 +02:00
2026-09-07 22:34:57 +02:00
2026-09-07 22:34:57 +02:00
2026-09-07 22:34:57 +02:00
2026-09-07 22:34:57 +02:00
2026-08-14 21:33:15 +00:00
2026-10-02 23:04:51 +02:00
2026-10-02 23:13:56 +02:00
2026-08-18 08:56:32 +00:00
2026-10-02 23:36:01 +02:00
2026-10-02 23:40:38 +02:00
2026-08-20 08:49:09 +00:00
2026-08-20 08:49:09 +00:00
2026-08-16 19:45:43 +00:00
2026-08-16 19:45:43 +00:00
2026-08-14 21:33:15 +00:00
2026-10-02 23:18:35 +02:00
2026-09-12 21:05:19 +00:00
2026-10-02 23:18:35 +02:00
2026-08-15 14:36:04 +00:00
2026-08-14 21:33:15 +00:00
2026-08-14 21:33:15 +00:00
2026-08-18 21:58:57 +02:00
2026-10-02 23:34:42 +02:00
2026-08-18 10:53:09 +00:00
2026-08-17 10:06:14 +02:00
2026-08-17 10:06:14 +02:00
2026-08-14 21:33:15 +00:00
2026-10-02 23:08:46 +02:00
2026-08-14 21:33:15 +00:00
2026-09-22 13:59:10 +00:00
2026-09-22 13:59:10 +00:00
2026-09-22 13:59:10 +00:00
2026-09-22 13:59:10 +00:00
2026-09-07 22:34:33 +02:00
2026-09-07 22:34:33 +02:00
2026-09-22 13:59:10 +00:00
2026-10-02 23:27:43 +02:00
2026-09-22 13:59:10 +00:00
2026-10-02 23:05:08 +02:00
2026-09-22 13:59:10 +00:00
2026-09-22 13:59:10 +00:00
2026-10-02 23:27:43 +02:00
2026-09-22 13:59:10 +00:00
2026-09-22 13:59:10 +00:00
2026-08-14 21:33:15 +00:00
2026-08-14 21:33:15 +00:00
2026-08-14 21:33:15 +00:00
2026-08-14 21:33:15 +00:00
2026-08-14 21:33:15 +00:00
2026-09-07 22:33:55 +02:00
2026-08-14 21:37:51 +00:00
2026-10-02 23:04:51 +02:00
2026-08-20 09:31:59 +00:00
2026-09-07 22:34:33 +02:00
2026-10-02 23:05:08 +02:00
2026-09-22 13:37:16 +00:00
2026-08-14 21:33:15 +00:00
2026-08-14 21:33:15 +00:00
2026-08-14 21:33:15 +00:00
2026-08-16 19:45:43 +00:00
2026-08-14 21:33:15 +00:00
2026-08-14 21:33:15 +00:00
2026-10-02 23:41:37 +02:00
2026-08-14 21:33:15 +00:00
2026-08-14 21:33:15 +00:00
2026-08-14 21:33:15 +00:00
2026-09-07 22:34:33 +02:00
2026-08-16 18:15:24 +00:00
2026-08-14 21:33:15 +00:00
2026-08-14 21:33:15 +00:00
2026-08-14 21:33:15 +00:00
2026-08-14 21:33:15 +00:00
2026-08-18 21:58:04 +02:00
2026-08-19 08:28:53 +02:00
2026-08-19 08:28:53 +02:00
2026-08-14 21:33:15 +00:00
2026-08-14 21:33:15 +00:00
2026-08-14 21:33:15 +00:00
2026-08-14 21:33:15 +00:00
2026-08-14 21:33:15 +00:00
2026-08-14 21:33:15 +00:00
2026-08-14 21:33:15 +00:00
2026-08-14 21:33:15 +00:00
2026-08-14 21:33:15 +00:00
2026-08-14 21:33:15 +00:00
2026-08-14 21:33:15 +00:00
2026-08-20 22:13:16 +02:00
2026-08-18 23:19:21 +02:00
2026-08-14 21:33:15 +00:00
2026-08-14 21:33:15 +00:00
2026-09-22 13:33:39 +00:00
2026-08-18 21:58:57 +02:00
2026-08-18 21:58:57 +02:00
2026-08-14 21:33:15 +00:00
2026-08-14 21:33:15 +00:00
2026-08-14 21:37:51 +00:00
2026-08-14 21:37:51 +00:00
2026-08-14 21:37:51 +00:00
2026-08-14 21:37:51 +00:00
2026-10-03 22:53:44 +02:00
2026-08-14 21:33:15 +00:00
2026-08-14 21:33:15 +00:00
2026-08-14 21:33:15 +00:00
2026-08-14 21:33:15 +00:00
2026-08-14 21:33:15 +00:00
2026-08-14 21:33:15 +00:00
2026-08-14 21:33:15 +00:00
2026-08-14 21:33:15 +00:00
2026-08-15 11:07:45 +00:00
2026-08-20 09:31:59 +00:00
2026-10-02 23:25:04 +02:00
2026-10-02 23:25:04 +02:00
2026-10-02 23:25:04 +02:00
2026-08-20 09:31:59 +00:00
2026-09-22 13:33:39 +00:00
2026-08-20 09:31:59 +00:00
2026-10-02 23:25:04 +02:00
2026-09-22 13:33:39 +00:00
2026-10-02 23:10:56 +02:00
2026-10-02 23:10:56 +02:00
2026-10-02 23:40:38 +02:00
2026-09-07 22:34:33 +02:00
2026-09-22 13:59:10 +00:00
2026-08-14 21:33:15 +00:00
2026-10-02 23:27:43 +02:00
2026-10-02 23:04:05 +02:00
2026-10-02 23:04:05 +02:00
2026-08-14 21:33:15 +00:00
2026-10-02 23:25:04 +02:00
2026-08-18 08:35:51 +00:00
2026-08-14 21:33:15 +00:00
2026-08-30 15:47:10 +02:00
2026-08-16 20:43:47 +00:00
2026-10-02 23:13:56 +02:00
2026-08-14 21:33:15 +00:00
2026-08-16 15:33:55 +00:00
2026-08-18 22:48:11 +02:00
2026-08-18 22:48:11 +02:00
2026-08-16 15:33:55 +00:00
2026-10-02 23:05:34 +02:00
2026-10-02 23:05:34 +02:00
2026-08-20 10:32:44 +00:00
2026-08-14 21:33:15 +00:00
2026-08-14 21:33:15 +00:00
2026-08-16 20:43:47 +00:00
2026-08-15 11:35:40 +00:00
2026-09-07 22:34:33 +02:00
2026-08-16 19:45:43 +00:00
2026-08-16 19:45:43 +00:00
2026-09-22 13:37:16 +00:00
2026-10-02 23:08:46 +02:00
2026-10-02 23:21:18 +02:00
2026-09-07 22:34:33 +02:00
2026-09-07 22:34:33 +02:00
2026-09-22 13:33:39 +00:00
2026-09-07 22:34:33 +02:00
2026-09-07 22:34:57 +02:00
2026-09-07 22:34:57 +02:00
2026-09-22 13:19:37 +00:00
2026-09-22 13:19:37 +00:00
2026-08-16 15:33:55 +00:00
2026-10-02 23:13:56 +02:00
2026-08-29 17:53:14 +02:00
2026-09-07 22:34:57 +02:00