Skip to content

Pull requests: sonos/tract

New pull request New
47 Open 2,174 Closed
47 Open 2,174 Closed
Author
Filter by author
Loading
Label
Filter by label
Loading
Use alt + click/return to exclude labels
or ⇧ + click/return for logical OR
Projects
Filter by project
Loading
Milestones
Filter by milestone
Loading
Reviews
Filter by reviews
No reviews Review required Approved review Changes requested
Assignee
Filter by who’s assigned
Assigned to nobody Loading
Sort
Sort by
Newest Oldest Most commented Least commented Recently updated Least recently updated Best match
Most reactions
👍 👎 😄 🎉 😕 ❤️ 🚀 👀

Pull requests list

reduce: accumulate the generic f16 row sums in f32
#2587 opened Aug 4, 2026 by czoli1976 Contributor Loading…
softmax: accumulate the f16 row sum in f32
#2586 opened Aug 4, 2026 by czoli1976 Contributor Loading…
tanh: give aarch64 cores without FEAT_FP16 an f32-roundtrip f16 kernel
#2584 opened Aug 3, 2026 by czoli1976 Contributor Loading…
silu: serve f16 from a table instead of widening every chunk to f32
#2583 opened Aug 3, 2026 by czoli1976 Contributor Loading…
gelu: saturate the fused f32 kernel's tanh to exactly -1
#2582 opened Aug 3, 2026 by czoli1976 Contributor Loading…
1
gelu: fuse the erf-form chain into a single op
#2581 opened Aug 3, 2026 by czoli1976 Contributor Loading…
rms_norm: run f16 rows natively instead of widening the whole tensor
#2580 opened Aug 3, 2026 by czoli1976 Contributor Loading…
erf: serve f16 from a table instead of converting the whole tensor each eval
#2579 opened Aug 3, 2026 by czoli1976 Contributor Loading…
linalg/mmm: take the chunking slack only where the cache absorbs it
#2570 opened Aug 3, 2026 by czoli1976 Contributor • Draft
1
gelu: dispatch f16 through linalg and back it with a lookup table
#2568 opened Aug 3, 2026 by czoli1976 Contributor Loading…
perf(linalg): run SiLU through the fused FMA kernel on AVX-512 hosts
#2567 opened Aug 3, 2026 by cverrier Contributor • Draft
WASM PR 4: fold the repeated fused-op arms into macros
#2566 opened Aug 3, 2026 by czoli1976 Contributor Loading…
WASM PR 3: test the invariants the WASM kernels rely on, and run CI against relaxed-simd
#2565 opened Aug 3, 2026 by czoli1976 Contributor Loading…
1
WASM PR 2: move the WASM microbenches out of src and into the bench target
#2564 opened Aug 3, 2026 by czoli1976 Contributor Loading…
2
linalg: give hwbench a bandwidth probe on architectures without a hand-written one
#2562 opened Aug 3, 2026 by czoli1976 Contributor Loading…
1
3
feat(linalg): add AVX2+FMA SiLU f32 kernel
#2560 opened Aug 3, 2026 by cverrier Contributor Loading…
4
Resize PR 3: Run Resize on the Metal GPU
#2556 opened Aug 2, 2026 by czoli1976 Contributor Loading…
1
metal: max and average pooling kernels
#2553 opened Aug 2, 2026 by czoli1976 Contributor Loading…
1
metal: mlx's preferred implicit-GEMM conv kernel, plus grouped convolutions (1.2-75x)
#2551 opened Aug 1, 2026 by czoli1976 Contributor Loading…
1
metal: depthwise convolution via a ported MLX kernel (2-14x)
#2550 opened Aug 1, 2026 by czoli1976 Contributor Loading…
metal: implicit-GEMM convolution via a ported MLX kernel (9-66x)
#2549 opened Aug 1, 2026 by czoli1976 Contributor Loading…
2
metal: route skinny f16 matmuls through mlx gemv_wide (~2.9x on M4)
#2547 opened Aug 1, 2026 by czoli1976 Contributor Loading…
1
metal: drop the MFA triangular-constant probe test- fails on M1 Pro (macOS 26.5.2) and an M4 (macOS 26.6).
#2546 opened Aug 1, 2026 by czoli1976 Contributor Loading…
nnef: lazy weight loading, so a pruned model never reads what it discards
#2524 opened Jul 26, 2026 by czoli1976 Contributor Loading…
6
fix(api): keep runnable state live across run() calls
#2522 opened Jul 24, 2026 by cverrier Contributor • Draft
1
ProTip! Adding no:label will show everything without a label.