diff --git a/docs/02_cont_overhead.md b/docs/02_cont_overhead.md new file mode 100644 index 0000000..d46bf60 --- /dev/null +++ b/docs/02_cont_overhead.md @@ -0,0 +1,2 @@ +# Chantier 2 — CONT/copies prefill: ROI faible, non engage +416 CONT = ~17/couche GDN (reshape/permute DeltaNet, qwen3next.cpp). 4.6s mesure OPSYNC=serialise; async (defaut) les copies recouvrent le compute -> part reelle << gonfle. cont l.165 = TODO upstream (CUDA non-contig unary), pas removable safe. prefill deja 46-50, pas de bottleneck CONT prouve. Decision: low-ROI, skip. Le vrai gain prefill = GDN sur HTP (chantier 3), pas tailler des copies.