Add MLX backend for Apple Silicon (GPU + ANE) - #1199

Open

ChinChangYang wants to merge 59 commits into

lightvector:masterfrom

ChinChangYang:mlx-backend-squash

Commits on May 23, 2026

Add MLX backend for Apple Silicon

ChinChangYang
and
claude
committed
81b00db

Commits on May 26, 2026

MLX backend: ANE/CoreML correctness + concurrency fixes, cross-path parity smoke test (ChinChangYang/KataGo#26)
ChinChangYang
committed
628e377

Commits on Jun 2, 2026

Merge remote-tracking branch 'origin/master' into mlx-backend-squash
ChinChangYang
committed
d5647ed

Commits on Jun 3, 2026

Fix MLX createComputeContext signature for merged nninterface API

ChinChangYang
and
claude
committed
19d7617
Add transformer support to MLX backend

ChinChangYang
and
claude
committed
01d1e3a
Port CoreML/ANE transformer support to MLX backend ANE mux

ChinChangYang
and
claude
committed
be1513d
Keep MLX GPU pooling in fp16 when useFP16

ChinChangYang
and
claude
committed
8ffe8ad
Make MLX Winograd tuner robust to failing candidates and drop redundant warmup

ChinChangYang
and
claude
committed
3f45e0c
Remove dead post-project CMAKE_OSX_DEPLOYMENT_TARGET pins

ChinChangYang
and
claude
committed
735030e
Validate transformer attention projection dims in CoreML converter

ChinChangYang
and
claude
committed
829d4fb

Commits on Jun 4, 2026

Reduce CoreML conversion peak and ANE steady-state memory

ChinChangYang
and
claude
committed
b6e32af
Enforce non-owning weight-view contract at compile time

ChinChangYang
and
claude
committed
fa4feb6
RAII the gzFile handle in KataGoParser

ChinChangYang
and
claude
committed
e7f1b98
Replace WeightEntry raw ptr+count with a local FloatView

ChinChangYang
and
claude
committed
23d9822
Clarify weight-release safety comment: aneOnly is the guarantee

ChinChangYang
and
claude
committed
93289e0
Refactor weight release into per-struct releaseWeights() methods

ChinChangYang
and
claude
committed
4630a0d
Co-locate releaseWeights() defs with each struct's other methods

ChinChangYang
and
claude
committed
4cca6cc
Conform CoreML transformer derived consts to the owned-weight + FP32 contract

ChinChangYang
and
claude
committed
35ba9d8
Cover transformer descriptors in releaseWeights()

ChinChangYang
and
claude
committed
4839b37
Release in-memory weights on the MLX backend's ANE-only path

ChinChangYang
and
claude
committed
f5565a1
Validate transformer FFN matmul dimensions in CoreML parser

ChinChangYang
and
claude
committed
c741171
Use ScopedFp32 RAII for all CoreML FP32-escalation windows

ChinChangYang
and
claude
committed
95db996
Remove dead Model::apply() on the MLX backend

ChinChangYang
and
claude
committed
ce18e63

Commits on Jun 5, 2026

Normalize USE_BACKEND case before project(); clarify backend-agnostic test comment

ChinChangYang
and
claude
committed
7c71f5d
Accumulate MLX Winograd F(2,3) transforms in fp32

ChinChangYang
and
claude
committed
55b46ff

Commits on Jun 6, 2026

Make MLX Winograd auto-tune coarse; reserve the wide sweep for full tuning

ChinChangYang
and
claude
committed
3ae836d
Fail loudly on unknown MLX block kinds; fix MLX build docs

ChinChangYang
and
claude
committed
f2d89db
Wire the MLX winotuner into the ./katago tuner subcommand

ChinChangYang
and
claude
committed
3fe7526
Remove KATAGO_MLX_WINOTUNER_FULL; full sweep is command-only

ChinChangYang
and
claude
committed
da9eb7b
Remove KATAGO_MLX_WINOTUNER_FORCE; re-tune is command-only

ChinChangYang
and
claude
committed
4deeaf4

Commits on Jun 7, 2026

Add protobuf/abseil to the MLX build prerequisites in Compiling.md

ChinChangYang
and
claude
committed
ecab6c3
Key the MLX Winograd tuner cache by GPU chip, not a fixed string

ChinChangYang
and
claude
committed
51b028d
Keep MLX GPU attention in fp16 (avoid accidental fp32 promotion)

ChinChangYang
and
claude
committed
5224098

Commits on Jun 8, 2026

Resolve MLX backend review NITs: atomic tuner save, oracle test, dead param, transformer test

ChinChangYang
and
claude
committed
a6963d3

Commits on Jun 9, 2026

Resolve MLX review findings: tuner +inf guard, registerWeight rvalue delete, help text, fp16 accum test

ChinChangYang
and
claude
committed
8afe269

Commits on Jun 11, 2026

CI: add macOS MLX backend build job to build.yml

ChinChangYang
and
claude
committed
1bfc4af
MLX: serialize GPU eval across NN server threads

ChinChangYang
and
claude
committed
7455e01
MLX: faster Winograd autotuner (fast coarse path, greedy descent, cross-net memo)

ChinChangYang
and
claude
committed
c74ce51
MLX: encapsulate tune memo and drain the ComputeHandle ctor

ChinChangYang
and
claude
committed
f3ddff0

Commits on Jun 12, 2026

MLX: unify the flat-sweep diagnostic logging into one helper

ChinChangYang
and
claude
committed
1b0cbf9
Fix Windows CI: stop pinning CMake to the VS 2022 generator (#1208)

ChinChangYang
and
claude
committed
db34ce4
MLX: overlap command encoding with GPU execution

ChinChangYang
and
claude
committed
b10ff59

Commits on Jun 13, 2026

MLX: fuse BN+act and residual epilogues into Winograd untransform

ChinChangYang
and
claude
committed
63446b6
MLX: fuse the gpool block's final residual into the Winograd untransform

ChinChangYang
and
claude
committed
d1052c7
MLX: fuse gpool regularConv bias-add + BN+act into the Winograd untransform

ChinChangYang
and
claude
committed
cbb6f20
MLX: harden backend memcpy size + tuner cache parsing

ChinChangYang
and
claude
committed
90e8b9f
CoreML converter: restore parser validations + fix IO dtype mismatch

ChinChangYang
and
claude
committed
5534bfb

Commits on Jun 14, 2026

MLX: tuner/backend review hardening (memo key, isValid bounds, fail-loud)

ChinChangYang
and
claude
committed
2dc563c
CoreML parser: validate RMSNorm epsilon (in (0,1]) matching master

ChinChangYang
and
claude
committed
b40230f
CoreML parser: validate nested numBlocks and matbias numChannels

ChinChangYang
and
claude
committed
5f64ed3
Refactor MainCmds::tuner into a thin per-backend dispatcher

ChinChangYang
and
claude
committed
5b5899c

Commits on Jul 27, 2026

Merge master into mlx-backend-squash
ChinChangYang
committed
4d47bd5

Commits on Aug 6, 2026

Merge branch 'master' into mlx-backend-squash
ChinChangYang
committed
46b038d

Commits on Aug 8, 2026

Enable the large v17 transformer entries in rungpuerrortest.sh

ChinChangYang
and
claude
committed
8dc00ca

Commits on Aug 27, 2026

Merge branch 'master' (v1.18.1) into mlx-backend-squash

ChinChangYang
and
claude
committed
ca215e6
Drop USE_BACKEND_NORMALIZED, reusing upstream's pre-project normalization

ChinChangYang
and
claude
committed
507804a
MLX: fold greedy coordinate descent into mlxwinotuner.cpp

ChinChangYang
and
claude
committed
6800d66
Document the MLX backend across the user-facing docs

ChinChangYang
and
claude
committed
6fe3095

Commits on Aug 28, 2026

Print the MLX backend banner in benchmark output

ChinChangYang
and
claude
committed
6726734