Repository navigation

CI run: Highway CPU kernels on Linux arm64 (not for merge) - #2

Draft

rcfa wants to merge 37 commits into

ci-base-p25from

t167p2-arm64

Draft

CI run: Highway CPU kernels on Linux arm64 (not for merge)#2
rcfa wants to merge 37 commits into
ci-base-p25from
t167p2-arm64

Commits on Oct 23, 2026
- Add Google Highway 1.4.0 as a submodule, not yet in any build
  rcfa
  committed
  a6e4df4
- Add a check that SwiftPM and Xcode compile the same Cmlx code on macOS
  rcfa
  committed
  fd17847
- Build the CPU backend's Highway kernels on Linux x86-64
  rcfa
  committed
  4ef6c3b
- Add CmlxCPUShim: the CPU backend's Highway, int8 and thread controls
  rcfa
  committed
  989f53e
- Add VMLXCPUKernelTests and its runner
  rcfa
  committed
  1b41593
- Show the values of failed expectations in the kernel-test runner
  rcfa
  committed
  2865fb2
- Measure MLX's polynomial accuracy on the scalar CPU code
  rcfa
  committed
  e267269
- Test the CPU quantized matmul against exact data and a float32 bound
  rcfa
  committed
  2c488ce
- Test the CPU fp32 and half-precision GEMMs around their thresholds
  rcfa
  committed
  0dac4ae
- Test CPU norms, RoPE and attention against float64 with derived bounds
  rcfa
  committed
  c80e218
- Test the CPU elementwise functions, reductions and softmax
  rcfa
  committed
  1b67ee8
- Run the CPU kernel tests in CI, and move Linux SwiftPM to Swift 6.4
  rcfa
  committed
  fc4fd9b
- Expect OpenBLAS pinned at every CPU pool size
  rcfa
  committed
  254a450
- Test that the int8 switch starts off
  rcfa
  committed
  57e28b8
- Compute the GEMM bound's reference on the host, in Double
  rcfa
  committed
  8763ba8
- Give the kernel-test runner distinct exit statuses, and a 1200 s stall
  rcfa
  committed
  a26e295
- Skip the tests that need Highway kernels, instead of passing them
  rcfa
  committed
  9686ce2
- Make a measuring run of the polynomial tests fail
  rcfa
  committed
  0b4576c
- Bound erf in half precision with the floor of its baseline
  rcfa
  committed
  55dd8ab
- Judge the matmul bounds on the host, and name the worst element
  rcfa
  committed
  ae61eff
- Check the shim's families and targets, and read back the int8 switch
  rcfa
  committed
  25fc208
- Refuse Highway kernels off x86-64, and check the highway submodule
  rcfa
  committed
  0a9ad44
- Correct the kernel tests' comments, and pin them to the CPU device
  rcfa
  committed
  7b94396
- Compare the kernel tests' exact results bit for bit
  rcfa
  committed
  717df4e
- Test the CPU gather quantized matmul, batched by expert and pooled
  rcfa
  committed
  73bfa1d
- Test untransposed and 2-bit quantized weights on the CPU
  rcfa
  committed
  09e51a4
- Test the CPU's threaded gathers and general copies
  rcfa
  committed
  ff9ace5
- Test float16 infinities and NaN through the norm kernels' loads
  rcfa
  committed
  f4bed10
- Replace private plan references with their reasons
  rcfa
  committed
  9b5125f
- Take the fork's thread-pool fix
  rcfa
  committed
  36608fb
- Take the fork's fix for Simd<bool> of a vector
  rcfa
  committed
  5514aae
- Take the fork's fixes for cores, empty arrays, float16, norms and sdpa
  rcfa
  committed
  ebd235b
- Build the Highway kernels on Linux arm64, with the SVE family disabled
  rcfa
  committed
  35a17b8
- Build without Highway on request, and name the CPU on arm64
  rcfa
  committed
  40e20b2
- Test fromFP8, dequantize and max/min against oracles of their own
  rcfa
  committed
  41a52cb
- Run arm64 CI with Highway, add a leg without it, and a TSan job
  rcfa
  committed
  17aa4c3
- Take the fork's Highway build for Linux arm64
  rcfa
  committed
  98eb14f

Commits

Commits on Oct 23, 2026