From 05b5bf9db1ee4560887d34ce945f24b0446d3147 Mon Sep 17 00:00:00 2001 From: biondizzle Date: Sat, 30 May 2026 10:54:21 +0000 Subject: [PATCH] docs: mark P5 as done in NEXT_PRIORITIES.md --- NEXT_PRIORITIES.md | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-) diff --git a/NEXT_PRIORITIES.md b/NEXT_PRIORITIES.md index b4ba5bff..b9705b89 100644 --- a/NEXT_PRIORITIES.md +++ b/NEXT_PRIORITIES.md @@ -98,7 +98,11 @@ the prompt to apply doctrine rule 3, not to pick one. prefill / long-context throughput; "decode works without TMA" is true but doesn't generalize. -### **P5 — In-kernel online softmax across KV tiles (the real P2)** +### **P5 — Integrate multi-tile FMHA into production** ✅ DONE + +Shipped 2026-05-30. Wired the existing D1.5 kernel +(fmha_6warp_tma_multirow_multitile.cuh) via fmha_multitile_capi.cu + +fmha_multitile_op.py into production.py. 18 integration tests pass. After P3, the decode fast path bypasses the Python merge for `n_segments==1`. But CSA with `top_k=1024` always has `n_segments=8`, so the moment top_k > 128