v0.119.0

0.119.0 (2026-07-21)

Features

• simd: block-Q8_0 GEMM matching ggml_vec_dot_q8_0_q8_0 (#830) (3a4d27e)

Bug Fixes

• cuda: power_scalar returned NaN for negative bases under --use_fast_math (#827) (4ae2377)
• gpu: add tape-gradient parity and correct backward kernels (#829) (f1b6dff)

Performance

• fft: blittable buffer-reusing NativeComplexFFT — kill per-call Complex<T> allocation (#855) (b61cd73)
• fft: wire NativeComplexFFT butterfly to SIMD radix-2 (test-gated) (#831) (7938eb2)