Richard Loyer
|
7f9dda4a75
|
Cascade 9B+Guard4B tient (7.5GB), Guard8B ok aussi; verdict gros modeles
|
2026-05-24 19:21:08 +02:00 |
|
Richard Loyer
|
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 |
|
Richard Loyer
|
47e75d74a4
|
Prefill cap = recurrence serielle, HVX inutile; gain=chunkwise GDN matriciel (weeks)
|
2026-05-24 18:55:49 +02:00 |
|
Richard Loyer
|
e1a6d4e764
|
Prefill: HMX/batch satures, GDN scalaire ancre, plafond ~103 via HVX chunkwise
|
2026-05-24 18:29:46 +02:00 |
|
Richard Loyer
|
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 |
|
Richard Loyer
|
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 |
|
Richard Loyer
|
1b64bbb7eb
|
Clarif: gen lente = decode-NPU cli, pas GDN; split prefill-NPU/decode-CPU = role JNI
|
2026-05-24 18:06:40 +02:00 |
|
Richard Loyer
|
104c6b4ad5
|
CLAUDE: chantier1 clos
|
2026-05-24 17:16:44 +02:00 |
|
Richard Loyer
|
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
|
3a5f735ab9
|
Optim prefill: ngl99/t8/b512=49.9 t/s (+11%), full offload, decode CPU
|
2026-05-24 16:35:52 +02:00 |
|
Richard Loyer
|
627f9a80e2
|
Correctif: OPFILTER inexistant, isolation GDN_cpu = bench DEBUG; bugs+qualite requalifies
|
2026-05-24 15:58:18 +02:00 |
|
Richard Loyer
|
d05483b13a
|
Bug cli = overflow buffer HTP gros prefill (2^44), pas init; qualite 4B OK
|
2026-05-24 15:52:33 +02:00 |
|
Richard Loyer
|
753d8be2b0
|
pp1024 plat (saturation batchable), bornes GDN: cout B-2 non mesure, profiler bloque par cli
|
2026-05-24 15:44:45 +02:00 |
|
Richard Loyer
|
4dd863ecb5
|
Verdict: GDN HTP decode-only (2284), prefill GDN fallback CPU = ancre. B chiffre.
|
2026-05-24 14:49:39 +02:00 |
|
Richard Loyer
|
eceacaba9a
|
Mesures 4B: HTP prefill 44 / CPU decode 13.4, qwen35_compile HTP fonctionne
|
2026-05-24 14:33:17 +02:00 |
|
Richard Loyer
|
2779babcab
|
Rebuild upstream+hexagon: configure OK (NDK r27d, SDK 6.5, v79), build lance
|
2026-05-24 13:01:16 +02:00 |
|
Richard Loyer
|
6aa4d06c95
|
Ambition haute: cible = Qwen3.5-4B near-ExecuTorch, prefill GDN sur HTP, decode CPU
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|
2026-05-24 12:57:37 +02:00 |
|
Richard Loyer
|
12217d3bbc
|
Decision A: integration produit, base upstream master, pas de kernel GDN custom
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|
2026-05-24 12:54:28 +02:00 |
|
Richard Loyer
|
376b4a4a70
|
Angle revise: base upstream master (pas chraac), trou unique = prefill GDN Hexagon
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|
2026-05-24 12:52:08 +02:00 |
|
Richard Loyer
|
cc5b033762
|
Engine cadrage: CLAUDE.md analyse R&D, verdict NPU-prefill/CPU-decode
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|
2026-05-24 12:34:14 +02:00 |