m3ultra
5a322407f0
mlx-tune: T2 whole-DiT mx.compile (same HY3D_MLX_COMPILE flag; parity 7.3e-06, 1.01x micro — GEMM-bound, kept as free insurance)
...
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-19 12:54:44 +10:00
m3ultra
4b368eef55
mlx-tune: T1 env-gated UNet mx.compile + T3 StaticMoELayer (kills ~4800 .item() syncs/gen, bit-identical, compile-unlocked) + T4 fused DINO SDPA (1.41x micro, parity 1e-7)
...
Micro gates on m3ultra (BENCHMARKS.md): T1 parity 3.0e-05 @ 1.04x;
T3 parity 0.0 (+compiles clean); T4 parity 1.1e-07 @ 1.409x.
E2E flag-on-vs-off gen pending GPU slot.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-19 12:43:54 +10:00
modelbeast
04fd1ceb28
README: MODELBEAST build overview (provenance, our changes, quick start, hardware, license); upstream README -> README_UPSTREAM.md
2026-07-16 15:00:25 +10:00
modelbeast
e4cfa9d1e9
Clean MLX build for MODELBEAST (inference-only)
...
Fork of dgrauet/Hunyuan3D-2.1-mlx + our generate_e2e.py CLI, env-tunable
remesh (HY3D_REMESH_FACES), and HARDWARE.md. Upstream training data
(mini_trainset) and demo images stripped — inference needs none of it.
Full upstream history: github.com/dgrauet/Hunyuan3D-2.1-mlx
2026-07-16 14:38:14 +10:00
