Commits

Match MoE reference top-k tie-break to tract on signed zeros

kali

committed

f66b50e

permute_axes: return identity permutations untouched

kali

committed

7be9993

MoveAxis: copy once with a tiled transpose

authored and

kali

committed

32224c7

build(deps): bump the rust-dependencies group with 2 updates

dependabot[bot]

authored and

kali

committed

ccabfeb

build(deps): bump dtolnay/rust-toolchain

dependabot[bot]

authored and

kali

committed

451606e

depthwise: keep channel-last zones on the per-point walk

kali

committed

b297b0c

Revert "depthwise: keep the scalar walk for more than four taps on Cortex-A53/A55"

kali

committed

213e7e7

depthwise: keep the scalar walk for more than four taps on Cortex-A53/A55

authored and

kali

committed

b77509a

depthwise: vectorise zones of more than four taps on aarch64

authored and

kali

committed

a7ed521

gpu: a copy was refused on rank before its axes were merged

kali

committed

9d6542b

Borrow MoE plans during evaluation and validate regular rows in constant time

JulienBalianSonos

authored and

kali

committed

8d086d5

Replace unowned routed-input pointers with shared tensor ownership and checked offsets, keeping reusable execution scratch free of input references.

JulienBalianSonos

authored and

kali

committed

9f4433b

Replace fixed MoE numerical examples with shared runtime property tests while retaining focused codegen regressions.

JulienBalianSonos

authored and

kali

committed

30e8773

Keep MoE lowering with its operator and separate packed Q40 execution from ordinary CPU plans.

JulienBalianSonos

authored and

kali

committed

80ac21f

Avoid redundant f32 input copies in routing and clamped SwiGLU, and keep the routed Q40 operator beside its implementations.

JulienBalianSonos

authored and

kali

committed

ce22545

Simplify MoE imports and derive the equivalent state Debug implementation.

JulienBalianSonos

authored and

kali

committed

09af7ea

Correct OLMoE coverage to reflect downstream model validation rather than only synthetic routing tests.

JulienBalianSonos

authored and

kali

committed

40aecb4

Include OLMoE-style full-softmax routing in the documented synthetic interchange coverage.

JulienBalianSonos

authored and

kali

committed

f6da297

Clarify the MoE interchange coverage for Qwen3, Granite and GPT-OSS-style FFN blocks without claiming complete model-family support.

JulienBalianSonos

authored and

kali

committed

fc19a19

NNEF activation tests enabled transformer registration; add core-only round trips to enforce ClampedSwiGlu independence.

JulienBalianSonos

authored and

kali

committed

fc554b6

Clamped SwiGLU was coupled to MoE execution; move it into core with an independent input contract, NNEF interchange and shared runtime tests.

JulienBalianSonos

authored and

kali

committed

8e15e56

Separate MoE activation and lowering from CPU execution; represent expert plans and state as explicit alternatives and add tiny interchange fixtures.

JulienBalianSonos

authored and

kali

committed

b70b1f8

Keep MoE bias equality consistent with hashing and retain runtime validation for unresolved weight shapes, while preserving routed lowering for symbolic token counts.

JulienBalianSonos

authored and

kali

committed

a61943c

Prevent incorrect MoE branch merging, validate operator contracts and stabilize routing ties; separate and document interchange, routed ops and CPU plans while reducing unchecked tensor access.

JulienBalianSonos

authored and

kali

committed

bd0e4dd

MoeFfn could be imported but not exported; serialize its portable NNEF form and report the packed-plan export boundary.

JulienBalianSonos

authored and

kali

committed

d2f09e5

MoE still called retired tensor accessors; use the plain-RAM APIs and remove redundant clones and option unwraps.

JulienBalianSonos

authored and

kali

committed

a2bf32f

fix(moe-ffn): follow upstream's EvalContext, freeze-removal and mmm-dispatch rewrites

JulienBalianSonos

authored and

kali

committed

583770b

feat(transformers): add tract_moe_ffn operator for Mixture-of-Experts FFN

JulienBalianSonos

authored and

kali

committed

fec155c

ci: run the paranoid-assertions and asan suites under opt-test

kali

committed

a3bbafa

ci: run cross.sh tests under an optimised profile with checks

kali

committed

d43f71f

Navigation Menu

Commits on Oct 5, 2026

Match MoE reference top-k tie-break to tract on signed zeros

permute_axes: return identity permutations untouched

MoveAxis: copy once with a tiled transpose

build(deps): bump the rust-dependencies group with 2 updates

build(deps): bump dtolnay/rust-toolchain

depthwise: keep channel-last zones on the per-point walk

Revert "depthwise: keep the scalar walk for more than four taps on Cortex-A53/A55"

depthwise: keep the scalar walk for more than four taps on Cortex-A53/A55

depthwise: vectorise zones of more than four taps on aarch64

gpu: a copy was refused on rank before its axes were merged

Borrow MoE plans during evaluation and validate regular rows in constant time

Replace unowned routed-input pointers with shared tensor ownership and checked offsets, keeping reusable execution scratch free of input references.

Replace fixed MoE numerical examples with shared runtime property tests while retaining focused codegen regressions.

Keep MoE lowering with its operator and separate packed Q40 execution from ordinary CPU plans.

Avoid redundant f32 input copies in routing and clamped SwiGLU, and keep the routed Q40 operator beside its implementations.

Simplify MoE imports and derive the equivalent state Debug implementation.

Correct OLMoE coverage to reflect downstream model validation rather than only synthetic routing tests.

Include OLMoE-style full-softmax routing in the documented synthetic interchange coverage.

Clarify the MoE interchange coverage for Qwen3, Granite and GPT-OSS-style FFN blocks without claiming complete model-family support.

NNEF activation tests enabled transformer registration; add core-only round trips to enforce ClampedSwiGlu independence.

Clamped SwiGLU was coupled to MoE execution; move it into core with an independent input contract, NNEF interchange and shared runtime tests.

Separate MoE activation and lowering from CPU execution; represent expert plans and state as explicit alternatives and add tiny interchange fixtures.

Keep MoE bias equality consistent with hashing and retain runtime validation for unresolved weight shapes, while preserving routed lowering for symbolic token counts.

Prevent incorrect MoE branch merging, validate operator contracts and stabilize routing ties; separate and document interchange, routed ops and CPU plans while reducing unchecked tensor access.

MoeFfn could be imported but not exported; serialize its portable NNEF form and report the packed-plan export boundary.

MoE still called retired tensor accessors; use the plain-RAM APIs and remove redundant clones and option unwraps.

fix(moe-ffn): follow upstream's EvalContext, freeze-removal and mmm-dispatch rewrites

feat(transformers): add tract_moe_ffn operator for Mixture-of-Experts FFN

ci: run the paranoid-assertions and asan suites under opt-test

ci: run cross.sh tests under an optimised profile with checks

Commits on Oct 2, 2026

Rework per review: onnx Cast becomes an Expansion wiring core Cast for everything but the String parse, which keeps a dedicated mini-op. The ElementWise wrapper was dropping uniform_tdim on the flo…

Drop an unused import in the comparison test that trips the warnings-as-errors CI step

Harden the prototype: only a strictly positive integer step yields a symbolic length (anything else falls back to a fresh symbol instead of silently coercing), and a failed symbolic gather fold lea…

Prototype (on top of #2929): keep symbolic values flowing through Cast and Gather so a Range whose limit re-derives an earlier Range's length through Shape/Gather/Cast shares one symbol instead of …

Range: i64 outputs for TDim inputs in both output_facts branches

Branch selector

User selector

Datepicker

Commit history

Commits on Oct 5, 2026

Commits on Oct 2, 2026