Skip to content

v0.120.0

Choose a tag to compare

@ooples ooples released this 28 Jul 10:33
f616c93

0.120.0 (2026-07-28)

Features

  • add direct PTX fused-kernel blueprint (#832) (2832eb2)
  • ptx: compile and dispatch direct-PTX on net471 (all TFMs) (#881) (cc14666)
  • ptx: direct-PTX kernel foundation - 5 emitter gaps closed + convolution family (#886) (3b21f7d)
  • ptx: fuse decode QKV, RoPE, and KV-cache update (#858) (51dace6)
  • ptx: implement fused attention kernel family (#857) (8fb5e91)

Bug Fixes

  • autodiff: gather backward must scatter-ADD, not reuse the gather-shaped helper (#894) (b541350)
  • autodiff: write fused-linear grads to live buffers (#888) (4ac1684)
  • gpu: gate sparsemax on where-select availability (#866) (9b1a88a)

Performance

  • ptx: staged + double-buffered tensor-core GEMM - 0.05x to 0.57x at 4096^3 (#893) (6d41166)
  • training: element-parallel fused optimizer + reusable per-step arena (#892) (4d3c3ec)