Prefill cap = recurrence serielle, HVX inutile; gain=chunkwise GDN matriciel (weeks)
This commit is contained in:
parent
e1a6d4e764
commit
47e75d74a4
|
|
@ -75,3 +75,6 @@ Qwen3-4B Engine: prefill HTP 103 / decode CPU 17. Qwen3.5-4B: 55/15. ExecuTorch
|
||||||
|
|
||||||
## Prefill 3.5: leviers gratuits satures
|
## Prefill 3.5: leviers gratuits satures
|
||||||
HMX on/off=54.7 identique, b1024=54.7 -> GDN scalaire = ancre (pas matmul). Plafond=~103 (dense). Levier unique = GDN HVX chunkwise (semaines, bit-exact frontiere). state f16 modere. decode 15 fige (BW). B-2 phase2 = HVX.
|
HMX on/off=54.7 identique, b1024=54.7 -> GDN scalaire = ancre (pas matmul). Plafond=~103 (dense). Levier unique = GDN HVX chunkwise (semaines, bit-exact frontiere). state f16 modere. decode 15 fige (BW). B-2 phase2 = HVX.
|
||||||
|
|
||||||
|
## Plafond prefill: HVX deja actif, inutile (55.5=55.9 scalar=hvx)
|
||||||
|
Mur = dependance serielle 512tok, pas calcul. HVX inner-dot ~0 gain. Cap 55 = recurrence. Seul levier = chunkwise GDN (delta rule matricielle->HMX->~100), changement algo, semaines, bit-exact frontiere. Aucun gain rapide. decode 15 fige.
|
||||||
|
|
|
||||||
Loading…
Reference in New Issue