monster/modelbeast
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

benchmarks: Hunyuan mlx-tune final — 243s vs 260s (1.07x), M1 rebench suggested

Browse Source 
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This commit is contained in:
m3ultra 2026-07-19 13:54:46 +10:00
parent d488d87490
commit a0e1ef1f18
1 changed files with 7 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7
BENCHMARKS.md
View File

@ -236,3 +236,10 @@ anatomy bench image, defaults (50 steps, octree 256): shape 146s + tex 101s
+ T1 UNet compile); shape delta noise-level. Caveat: baseline row used a
different input image; same settings though. Next lever: Target 2
whole-DiT compile (unblocked by T3).
Addendum — with T2 (whole-DiT compile) included: shape 144s + tex 99s =
**243s total**. Full mlx-tune vs 260s baseline ≈ **1.07× e2e**. Verdict:
on M3 Ultra these 2B GEMM-bound models get ~1% from mx.compile fusion;
the tune's real wins were T4 fused attention (texture stage) + T3
compile-safety. Branch mlx-tune @ 5a32240; flag HY3D_MLX_COMPILE=1.
Worth re-benching on M1 Ultra where launch overhead is proportionally larger.