Kazeia
/
Kazeia-engine
3
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

GDN prefill: bug=handoff multi-token (pas GQA), fix=dump bisection; defaut CPU sur

This commit is contained in:
Richard Loyer 2026-05-24 17:38:35 +02:00
parent 33e848be3c
commit 6e63110f8a
1 changed files with 3 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
docs/03_gdn_prefill_htp.md
View File

@ -1,2 +1,5 @@
# Chantier 3 — GDN prefill sur HTP # Chantier 3 — GDN prefill sur HTP
Skel gated-delta-net-ops.c:98 boucle sur n_tokens. Gate 2284 leve (test) -> pp512 50->55 MAIS sortie CHARABIA ("He thef% sur"). Kernel scalaire pas bit-correct au prefill (handoff etat S faux multi-token). +11% INVALIDE. Defaut remis CPU (GGML_HEXAGON_GDN_PREFILL=1 = opt-in test only). Vrai B-2 = corriger recurrence/etat + HVX chunkwise, bit-exact = semaines, aide partielle. Etape acquise: gate localise, kernel a fixer identifie. Pas de gain livrable. Skel gated-delta-net-ops.c:98 boucle sur n_tokens. Gate 2284 leve (test) -> pp512 50->55 MAIS sortie CHARABIA ("He thef% sur"). Kernel scalaire pas bit-correct au prefill (handoff etat S faux multi-token). +11% INVALIDE. Defaut remis CPU (GGML_HEXAGON_GDN_PREFILL=1 = opt-in test only). Vrai B-2 = corriger recurrence/etat + HVX chunkwise, bit-exact = semaines, aide partielle. Etape acquise: gate localise, kernel a fixer identifie. Pas de gain livrable.
## Diag: decode-HTP correct, prefill-HTP charabia = handoff multi-token (PAS GQA)
4B decode GDN-HTP gere num_v=32>num_k=16 -> GQA OK. Charabia uniquement prefill n_tokens>1 = layout/etat inter-token. Lead briques 4B_bug (block-repeat) = leur runtime, ici non transferable. Fix = dump HTP vs CPU par token (briques: ~5h bisection), zone scan bit-exact aide partielle. Defaut reste CPU (correct, prefill 50). B-2 = tache R&D suivie, pas patch aveugle.