[Apple Silicon] Add MPS and Metal inference support - #175

Open

Jourloy wants to merge 24 commits into

microsoft:mainfrom

Jourloy:main

Commits on Jul 17, 2026

add mlx backend for apple silicon with metal gpu acceleration

pedronaugusto
authored and
Jourloy
committed
inline env defaults in api_server, remove start script, drop intermediate mx.eval syncs

pedronaugusto
authored and
Jourloy
committed
sparse/config: probe flex_gemm on MPS, soft-fall-back to pytorch backend

authored and
Jourloy
committed
sparse/config: probe flex_gemm masked path too, not just dense

authored and
Jourloy
committed
sparse/attention: add 'flex_gemm_sparse_attn' backend option

authored and
Jourloy
committed
sparse/norm: CPU-staged zeros_like workaround for MPS

authored and
Jourloy
committed
sparse/attention: drop FUSED_ATTN_MAX_SEQLEN threshold to match CUDA-path precedent

pedronaugusto
authored and
Jourloy
committed
benchmarks/e2e_decoder: shell-conv + attention block micro-bench

pedronaugusto
authored and
Jourloy
committed
sparse/config: default ATTN to flex_gemm_sparse_attn on Darwin when flex_gemm probe passes

pedronaugusto
authored and
Jourloy
committed
deps: pin torch>=2.11.0 in requirements_macos.txt

pedronaugusto
authored and
Jourloy
committed
feat: add native Apple Silicon asset pipeline
Jourloy
committed
fix: cache only pinned runtime weights
Jourloy
committed
fix: retry Metal build with ABI fallback
Jourloy
committed
test: probe production split-k Metal kernel
Jourloy
committed
fix: distinguish requested and safety decimation
Jourloy
committed
fix: use SDPA for dense attention on macOS
Jourloy
committed
test: isolate standalone Metal integration smoke
Jourloy
committed
feat: measured mesh-integrity metrics module (WP1)

Jourloy
and
claude
committed
feat: bench_remesh harness for Metal narrow-band DC (WP2)

Jourloy
and
claude
committed
feat: remesh is the production candidate path (WP3)

Jourloy
and
claude
committed
test: postprocess parity + stability suite (WP4)

Jourloy
and
claude
committed
feat: default remesh_project 0.7 (detail recovery on clean raws)

Jourloy
and
claude
committed
feat: opt-in fp32 decode thresholds (WP9)

Jourloy
and
claude
committed
feat: guarded project_back (distance + normal-agreement + flip-revert)

Jourloy
and
claude
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[Apple Silicon] Add MPS and Metal inference support - #175

[Apple Silicon] Add MPS and Metal inference support#175
Jourloy wants to merge 24 commits into
microsoft:mainfrom
Jourloy:main

Commits on Jul 17, 2026

add mlx backend for apple silicon with metal gpu acceleration

inline env defaults in api_server, remove start script, drop intermediate mx.eval syncs

sparse/config: probe flex_gemm on MPS, soft-fall-back to pytorch backend

sparse/config: probe flex_gemm masked path too, not just dense

sparse/attention: add 'flex_gemm_sparse_attn' backend option

sparse/norm: CPU-staged zeros_like workaround for MPS

sparse/attention: drop FUSED_ATTN_MAX_SEQLEN threshold to match CUDA-path precedent

benchmarks/e2e_decoder: shell-conv + attention block micro-bench

sparse/config: default ATTN to flex_gemm_sparse_attn on Darwin when flex_gemm probe passes

deps: pin torch>=2.11.0 in requirements_macos.txt

feat: add native Apple Silicon asset pipeline

fix: cache only pinned runtime weights

fix: retry Metal build with ABI fallback

test: probe production split-k Metal kernel

fix: distinguish requested and safety decimation

fix: use SDPA for dense attention on macOS

test: isolate standalone Metal integration smoke

feat: measured mesh-integrity metrics module (WP1)

feat: bench_remesh harness for Metal narrow-band DC (WP2)

feat: remesh is the production candidate path (WP3)

test: postprocess parity + stability suite (WP4)

feat: default remesh_project 0.7 (detail recovery on clean raws)

feat: opt-in fp32 decode thresholds (WP9)

feat: guarded project_back (distance + normal-agreement + flip-revert)