Skip to content

Add current MLX arg-reduce native validation - #1936

Open
NripeshN wants to merge 11 commits into
mainfrom
mlx-current-corpus-native-validation
Open

NripeshN wants to merge 11 commits into
mainfrom
mlx-current-corpus-native-validation

Conversation

@NripeshN

@NripeshN NripeshN commented Sep 14, 2026

Copy link
Copy Markdown
Member

Summary

  • pin the current MLX Metal kernel tree at d9add9d11f3154111a4c85f267ec2fd307ecd18e and record its exact 42-unit / 17,478-entry zero-diagnostic census
  • translate and package all 24 discovered arg_reduce.metal entries deterministically for Metal, OpenGL, and DirectX, pinning all 72 target artifacts
  • require native macOS Metal, Linux glslang/SPIR-V, and Windows DXC compiler validation for all 24 entries while retaining numerical runtime/upstream-metallib parity for argmin_float32 and argmax_float32
  • add fail-closed backend support for exact nested software-subgroup helper identities and two-word shuffle lowering in OpenGL, internal-only bfloat groupshared register representation in DirectX, and native-width narrow local arrays in Metal
  • retain the initial current-tree aggregate-constructor, integral-reference pointer-offset, constrained free-function SFINAE, strict-JSON IEEE-754 input-token, and source-sensitive DXC native-16 repairs

Scope

This increment proves 24 / 17,478 entries for deterministic translation and native compiler validation, producing 72 pinned Metal/OpenGL/DirectX artifacts. Numerical runtime parity remains deliberately bounded to 2 / 17,478 entries: argmin_float32 and argmax_float32.

The compact contract explicitly does not claim complete-corpus translation, upstream MLX test-suite execution, or MLX host-runtime redirection yet.

Validation

  • complete target harnesses: Metal 25 passed in 211.37s; OpenGL 25 passed in 219.68s; DirectX command-contract proof 25 passed in 218.83s
  • Metal harness compiled and linked all 24 entries natively and executed the two float32 runtime/upstream-metallib parity entries
  • OpenGL harness validated all 24 entries with glslangValidator and spirv-val; two-entry Mesa EGL runtime remains required in Linux CI
  • DirectX harness reproduced all 24 artifacts through a strict exact-nine-argument shim requiring -enable-16bit-types, -WX, cs_6_6, and CSMain; real DXC/WARP remains required in Windows CI
  • broad affected suites: 8,545 passed, 30 expected skips
  • canonical Python 3.11 full suite: 18,436 passed, 32,528 expected skips, one known Hypothesis warning
  • all-file pre-commit and warning-fatal Sphinx 8.2.3 passed
  • support matrix is fresh; all 957 evidence rows are present with zero missing; CI coverage passed; 36-issue sync dry-run was non-mutating
  • exact current-tree census: 42 units, 17,478 entries, zero diagnostics

Support issue traceability: no issue closed

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant