From 8c793a22b5d8b158b70c4c668c121fe45d05f2b2 Mon Sep 17 00:00:00 2001
From: Richard Loyer <richard.loyer@c42.lu>
Date: Sun, 24 May 2026 17:57:20 +0200
Subject: [PATCH] FIX GDN prefill HTP: GQA broadcast h%heads, charabia resolu,
 pp512 58 t/s (+16%)

---
 docs/03_gdn_prefill_htp.md | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/docs/03_gdn_prefill_htp.md b/docs/03_gdn_prefill_htp.md
index eeee224..149ae42 100644
--- a/docs/03_gdn_prefill_htp.md
+++ b/docs/03_gdn_prefill_htp.md
@@ -6,3 +6,6 @@ Skel gated-delta-net-ops.c:98 boucle sur n_tokens. Gate 2284 leve (test) -> pp51
 
 ## Harness dump: FARF non capte par logcat -> sink DSP dedie requis (briques ~5h)
 Instrumente FARF GDNDUMP, mais logs DSP hors logcat sans ADSP_FARF config. dump bit-exact = vrai dump dans dst buffer + ref ggml = chantier multi-h. Instrument retiree, lib propre. Default CPU. B-2 = R&D suivie, pas a finir a chaud. Produit pret: prefill 50 CPU correct / decode 13.
+
+## FIX bit-correct: GQA broadcast (resolu)
+Cause: kernel HTP lisait h jusqu'a num_v_heads(32) sur q/k=16 tetes -> hors-tete = charabia. CPU ref ops.cpp:10466 fait iv1%neq1 / iv1%nek1 (interleaved). Fix gated-delta-net-ops.c: hq=h%q_t->ne[1], hk=h%k_t->ne[1] sur q/k/g/beta, v=h. Resultat: sortie coherente (avant 'He thef% sur'), prefill pp512 50->58 t/s (+16%) GDN sur HTP. Decode inchange CPU 13. B-2 etape 1 DONE bit-correct scalaire. Suite: HVX chunkwise pour >58.