21784d52de
eval: verdict (assisté, à valider) -> ship Qwen3.5-lmq4
...
16 prompts FR notés. SAFETY 2/2 à Qwen3.5 (cite 3114/15, dense reste générique malgré
le même system) = décisif pour le domaine psy. SUPPORT majorité 3.5 (dense: fuite "too",
fautes). Reste ~nul. Règle de décision remplie. Faiblesses 3.5 à corriger par system
(registre tu/vous instable; sur-interprétation item 2 "départ"->"décès"). Decode 10.8 vs
dense 11.7 (-8%), justifié par sécurité. Spot-check humain conseillé items 2/8/9.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-05-26 23:23:00 +02:00
777c6979c1
eval: kit qualité A/B aveugle Qwen3.5-lmq4 vs dense (decode CPU, template, reasoning off)
...
Bascule perf->qualité. Runner reproductible: genere un cahier de notation aveugle
(out/results_blind.md) + cle (out/results_key.csv) sur 16 prompts FR (SUPPORT/SAFETY/
BOUNDARY/INSTRUCTION/GENERAL). Notation humaine, regle de decision dans PROTOCOLE.
Resout l'eval shell (avant: template non applique -> ramble anglais): -cnv + -sysf +
--reasoning off -> FR propre. out/ ignore (artefacts generes).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-05-26 22:24:23 +02:00
a7c0e96793
RAPPORT §7: piste kernel GDN fermee — GDN-compute != goulot prefill (fork ql)
...
A/B pp512 mesure: serial f32=89.5 (vrai baseline) = fp16 89.65 = WY chunkwise 89.6.
3 schemas calcul GDN (0.5x/1x/2.5x flops) -> pp512 plat => l'arithmetique GDN ne
limite pas le prefill sur le fork ql (GDN-HTP). Mur ailleurs (denses/FFN/attn/op-batch).
Corrige premisse "GDN=ancre" (ancien tree GDN-CPU) et le baseline fantome "98".
Aucune optim kernel GDN (chunkwise/HMX/fp16) ne bouge le prefill.
ql -> 50e250d: fp16 serial validate (oracle 28/28, env KZ_F16 defaut OFF), WY chunkwise
conserve inutilise, HMX-per-tete NO-GO. Pousser pp512 = profiler ops non-GDN HTP.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-05-26 21:46:05 +02:00
dd29169672
GDN chunkwise P1.2: VTCM state prefill=neutre, prouve compute-bound 3S2/tok -> gain=WY matmul (palier dur). oracle 28/28, pp98
2026-05-26 16:30:08 +02:00
f638bdbd94
GDN chunkwise P1.1: chunk C=64 refactor neutre, oracle 28/28+pp98 OK, squelette pret pour intra-chunk batch
2026-05-26 16:15:55 +02:00
837d9de766
GDN chunkwise: doc P1 math+structure+gate (refactor chunks C=64 phase1)
2026-05-26 16:13:34 +02:00
aa363e652f
GDN chunkwise P0: oracle test-backend-ops 28/28 HTP0 (gate régression), baseline pp512=93, boucle build live; P1 chunk-f32 en cours
2026-05-26 16:04:58 +02:00
de9543de2a
Chantier4 GDN chunkwise: plan grounded metrics, kernel pp_thread sériel localisé, HMX core_dot_chunk_fp16 réutilisable, build htp-v79 incrémental OK
...
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-05-26 15:55:25 +02:00
b2fc564f04
Engine sur fork qualcomm: decode dense 15.5>pte, 3.5 ~11, KVq8+t4+fa+lm_head-Q4; prefill 93/285; RAG cross 24-28k
...
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-05-26 15:53:14 +02:00
1252f32bd3
Rapport R&D complet: synthese session, perf, cascade, #277 , briques, reste a faire
...
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-05-25 22:27:34 +02:00
9b3262a8aa
R&D: battre pte=lm_head HMX INT4+spec decode+ctx binary; mur=lm_head CPU
2026-05-25 18:46:28 +02:00
025bdd87bd
Briques: 2B 13, 4B 7, 9B 2.1; 2B sweet spot
2026-05-25 18:34:16 +02:00
cbc379f9e5
Chiffres stables r3: 4B 9.8 9B 5.2 (r1 etaient bruit); engine tient, pte 4B devant
2026-05-25 17:41:08 +02:00
8cc9286f70
Correction: 9B CPU ~4.4 (t8 contention), <8B pte; 4B seul GGUF viable
2026-05-25 15:54:46 +02:00
e462d7ce27
9B cible: 1-coeur in-app=0.16, t8=8.4; pte8B repli; merge=affinity 8 coeurs
2026-05-25 13:04:21 +02:00
7aae1f5fc7
Final: pte 4B 15/2.2GB > GGUF 0.32; merge=t8 mais pte reste mieux 4B; gguf utile=gros modeles
2026-05-25 12:21:04 +02:00
cdb34a65f1
0.32 = 1 thread + HTP residuel; fix t8+ngl0 = 14tok/s
2026-05-25 11:18:38 +02:00
23f2c9fcfc
#277 : dist=static ngl0 OK; crash dev=stale shared so + ancienne signature 1-arg
2026-05-25 10:44:53 +02:00
9b62420089
v4 fix #277 : ngl0 CPU decode 14tok/s, 4s/tour, usable
2026-05-25 10:18:31 +02:00
7a5780ca1d
Bridge v3 utilisable: thinking-off ChatML, sys+usr, cap; static intact
2026-05-25 09:41:04 +02:00
fe6826a3be
Integration #277 : hang=thinking non bride, fix budget0+cap; 4B>9B CPU
2026-05-25 09:38:25 +02:00
76f21a21cb
Bridge STATIC autonome (42MB): zero NEEDED partage, fix collision TTS libllama, valide device
2026-05-24 22:43:45 +02:00
f877902903
STATUS.md: validé/limites pour integrateur
2026-05-24 22:17:42 +02:00
9675280760
GDN chunkwise: HMX fp16 bloque bit-exact, HVX qf32 3-5sem; reste 55/15 CPU exact
2026-05-24 22:12:47 +02:00
1a6637bd67
TTS: Talker GGUF chargeable, audio e2e=JNI app non valide shell; GDN agent en cours
2026-05-24 22:12:23 +02:00
c5c2738281
Valide device: generate() bout-en-bout OK (test_native), bridge pret integration
2026-05-24 22:05:44 +02:00
7dbf38db97
Doc complete: README+MODELS+PERF+PITFALLS, build nettoye
2026-05-24 21:59:56 +02:00
15b9baf6b2
JNI bridge complet: generate() prefill+decode, build arm64 OK (libkazeia_engine.so 209K)
2026-05-24 21:58:00 +02:00
0f92fa723f
Dist Kazeia-Engine: libs arm64+headers+JNI bridge+INTEGRATION/TTS doc; STT reste ORT
2026-05-24 21:55:01 +02:00
1024167bdc
Empreinte: engine GGUF 2.4/5.0 vs pte 3.3/5.9, -15-30%
2026-05-24 21:18:16 +02:00
9f7bc1e6c3
Comparatif FR: 9B>4B (psy juste vs derive role), engine OK budget0
2026-05-24 21:12:10 +02:00
431865f07f
Fix cascade: reasoning-budget 0 (engine OK, FR coherent); A/B = JNI
2026-05-24 20:15:41 +02:00
25b1344c19
Cmp 4B/9B: 9B+riche, mais eval FR shell impossible, requiert JNI
2026-05-24 19:50:24 +02:00
db6c063705
Cascade 9B+Guard4B/8B: speed+RAM mesures, reco Guard4B; eval texte = JNI
2026-05-24 19:30:56 +02:00
7f9dda4a75
Cascade 9B+Guard4B tient (7.5GB), Guard8B ok aussi; verdict gros modeles
2026-05-24 19:21:08 +02:00
f82fa61281
9B: 41/8.4 5GB tient; 30B/35B OOM. Plafond=9B+4B parallele, MoE hors-jeu
2026-05-24 19:13:20 +02:00
47e75d74a4
Prefill cap = recurrence serielle, HVX inutile; gain=chunkwise GDN matriciel (weeks)
2026-05-24 18:55:49 +02:00
e1a6d4e764
Prefill: HMX/batch satures, GDN scalaire ancre, plafond ~103 via HVX chunkwise
2026-05-24 18:29:46 +02:00
ce5d91946e
Comparatif: Q3-4B 103/17, Q3.5-4B 55/15, ExecuTorch 14-21. Engine egale decode sans pte
2026-05-24 18:19:27 +02:00
eb72df4767
Mesure split ref: prefill HTP 55 / decode CPU 15 Qwen3.5-4B, end-to-end ~13s
2026-05-24 18:08:39 +02:00
1b64bbb7eb
Clarif: gen lente = decode-NPU cli, pas GDN; split prefill-NPU/decode-CPU = role JNI
2026-05-24 18:06:40 +02:00
8c793a22b5
FIX GDN prefill HTP: GQA broadcast h%heads, charabia resolu, pp512 58 t/s (+16%)
2026-05-24 17:57:20 +02:00
31afd64eec
GDN dump: FARF hors logcat (sink dedie req), lib nettoyee, B-2 tracke. Produit 50/13 pret
2026-05-24 17:42:34 +02:00
6e63110f8a
GDN prefill: bug=handoff multi-token (pas GQA), fix=dump bisection; defaut CPU sur
2026-05-24 17:38:35 +02:00
33e848be3c
Chantier 3: gate GDN-prefill leve = charabia, +11% INVALIDE, defaut CPU. Honnete: kernel a fixer
2026-05-24 17:22:52 +02:00
b82420fd46
Chantier 3 etape 1: GDN prefill HTP, gate leve, pp512 +11% (50->55)
2026-05-24 17:20:09 +02:00
ae842df5ae
Chantier 2 clos: CONT low-ROI (opsync gonfle, reshape structurel) -> 3 prioritaire
2026-05-24 17:17:24 +02:00
104c6b4ad5
CLAUDE: chantier1 clos
2026-05-24 17:16:44 +02:00
024f3b8ed6
Chantier 1 clos: prefill pp2048 sain (46t/s), hang=cli only hors JNI
2026-05-24 17:16:44 +02:00
17a78f1c6a
Profil: GDN absent HTP (100% CPU confirme), CONT+SSM_CONV+MULMAT dominent; lib restauree
2026-05-24 16:53:08 +02:00
Richard Loyer