zeekayandHanzo Dev 501e62a4f8 perf(metal): cache compiled MPSGraph conv per shape -> fixes UNet regression ...

Per-conv MPSGraph::new()+compile was a net loss on the UNets many small convs
(64->88ms). Caching the compiled graph+placeholders per (shape,dtype,params):
UNet back to 62ms (baseline parity), FULL-FAST 5.38->10.42 fps, full-VAE
1.04->2.24 fps. cosine still 1.000000.

Co-authored-by: Hanzo Dev <[email protected]>

2026-06-16 11:29:14 -07:00

patches

perf(metal): cache compiled MPSGraph conv per shape -> fixes UNet regression

2026-06-16 11:29:14 -07:00

src

feat(bench): convbench-metal A/B (im2col vs MPSGraph) + cosine gate

2026-06-16 11:24:59 -07:00

.gitignore

baseline: spark authoritative musetalk-bench (cuda src) before metal gating

2026-06-15 19:48:32 -07:00

Cargo.lock

wip(metal): env-gated MPSGraph conv2d path (HANZO_METAL_MPS=1)

2026-06-16 11:21:29 -07:00

Cargo.toml

feat(metal): gate cuda-only code, add metal device path

2026-06-15 19:58:00 -07:00

S

Description

Local-only checkout; no git remote existed

114 KiB

Languages

Rust 96.7%

Vim Snippet 3.3%