From 043097286a45e771fe42894dc4771f306d264ca1 Mon Sep 17 00:00:00 2001 From: bong-water-water-bong Date: Sun, 15 Nov 2026 05:15:27 -0400 Subject: [PATCH] config: gpu-kernels pin 2022afb74 -> baabfc3d6 (hip.fa_decode_d256 at n_kv >= 128 for multi-row attention, 4-row step -4.1 %) One commit, gpu-kernels#88: the split-K HIP decode for head-256 gated attention (Qwen3.5 / 3.6 / 3.8) engages from n_kv 128 when the attention has two or more query rows (llama-server -np N -kvu, speculative verify); one-token decode keeps 1024. Qwen3.8-27B Q4_K_M, 4 streams, dispatch profile: the Loom gate kernel's 3.30 ms/step (one workgroup per query head) becomes 0.32 ms/step on hip.fa_decode_d256, step GPU span 99.51 -> 95.39 ms. Decode-path KLD at 4 rows and at one sequence x 4 rows is within ~1 sigma of the floor. Co-Authored-By: Claude Opus 5.5 --- config/gpu-kernels.config | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/config/gpu-kernels.config b/config/gpu-kernels.config index 33365c6c6..1e9bb29f3 100644 --- a/config/gpu-kernels.config +++ b/config/gpu-kernels.config @@ -22,6 +22,6 @@ # # Move the pin by hand, after the HRX checks on Strix Halo (docs/hrx.md). repo=1bit-MONSTER/gpu-kernels -sha=2022afb745cced3bcee04f22ed4917cc937e3813 +sha=baabfc3d6eae1dd52292cedc6703ff2960e5d103 path=addons/hip-hrx require=addon.cmake kernels