Skip to content

v1.6.0

Choose a tag to compare

@hannahli-nv hannahli-nv released this 09 Nov 03:41

What's Changed

  • cuTile prefill ragged: guard-if body, persistent/latency autotune variants, sm12x/sm107/sm90 configs & other updates by @hannahli-nv in #205
  • ci: refresh .bench_durations by @github-actions[bot] in #206
  • Unify cold-cache CUDA Graph measurement and autotuning & style(tilecpp): reflow the nvcc invocation & other updates by @hannahli-nv in #208
  • [liger] fused_linear_cross_entropy: chunked grad_weight GEMM on tensor cores (2.4x faster above 4 GB of logits) by @Aryan95614 in #212
  • feat(cutile): add dropout backward with mask recomputation by @frankliee in #213
  • feat(qwen3.8): integrate Qwen/Qwen3.8-27B with TileGym cuTile kernels + skinny GEMV/GEMM groups — 1.73x speedup & other updates by @hannahli-nv in #215
  • Fix agent skills: normalize unofficial frontmatter fields per the Agent Skills spec by @hannahli-nv in #217
  • [liger] Add experimental cuTile fused linear DPO loss (sigmoid) by @Aryan95614 in #211
  • chore(hf-bench): upgrade transformers 5.10.2 -> 5.19.0 (+ tokenizers 0.23.3) & other updates by @hannahli-nv in #219
  • fix(ci): use run_all_json result file names in the selective baseline merge by @MaxFreedomPollard in #220
  • ci: refresh .bench_durations and fix benchmark result parsing by @github-actions[bot] in #218
  • fix(cutile-rs): build kernels offline first to avoid crates.io flakes & other updates by @hannahli-nv in #224

New Contributors

Full Changelog: v1.5.1...v1.6.0