NVIDIA/TileGym v1.3.0

v1.3.0

Repository: NVIDIA/TileGym

Tag: v1.3.0

Published: 2026-05-19T06:26:33Z

Prerelease: no

Release notes:

What's Changed

• perf(rmsnorm): skip bounds checking in gather kernel when tile evenly divides N by @liqiangxl in https://github.com/NVIDIA/TileGym/pull/111
• experimental/swa_attention_cutile by @DevTechJr in https://github.com/NVIDIA/TileGym/pull/107
• feat: migrate cuda.tile_experimental.autotune_launch → cuda.tile.tune.exhaustive_search & other updates by @hannahli-nv in https://github.com/NVIDIA/TileGym/pull/114
• Add cutile-python programming skill by @hannahli-nv in https://github.com/NVIDIA/TileGym/pull/115
• Update cutile kernels & other updates by @hannahli-nv in https://github.com/NVIDIA/TileGym/pull/116
• add multi_wave_cached rms norm by @liqiangxl in https://github.com/NVIDIA/TileGym/pull/113
• Add cuTile autotune disable policy & refactor(attention): update _fmha_autotune_configs & other updates by @hannahli-nv in https://github.com/NVIDIA/TileGym/pull/117
• skills: update .agents/skills/ license by @hannahli-nv in https://github.com/NVIDIA/TileGym/pull/119
• Update improve-cutile-kernel-perf skill & Add liger kernels & other updates by @hannahli-nv in https://github.com/NVIDIA/TileGym/pull/120
• skills: Skip non-SKILL.md skill docs in SPDX header check by @hannahli-nv in https://github.com/NVIDIA/TileGym/pull/122
• Refine tilegym ops kernel style by @hannahli-nv in https://github.com/NVIDIA/TileGym/pull/121
• Refactor and release Transformers patching and auto-kernelize loop & other updates by @hannahli-nv in https://github.com/NVIDIA/TileGym/pull/123
• Bump setup.py version from 1.2.0 to 1.3.0 by @hannahli-nv in https://github.com/NVIDIA/TileGym/pull/124
• Update attention and matmul tests & other updates by @hannahli-nv in https://github.com/NVIDIA/TileGym/pull/125
• [benchmark] add torch.profiler based bench utils with L2 cache flush by @liqiangxl in https://github.com/NVIDIA/TileGym/pull/118
• bench(softmax): use profile_with_l2flush for more realistic bandwidth measurements by @liqiangxl in https://github.com/NVIDIA/TileGym/pull/126
• [tests] test_gemma_attention: guard set_backend in test_op with try/skip & other updates by @hannahli-nv in https://github.com/NVIDIA/TileGym/pull/127

New Contributors

• @liqiangxl made their first contribution in https://github.com/NVIDIA/TileGym/pull/111
• @DevTechJr made their first contribution in https://github.com/NVIDIA/TileGym/pull/107

Full Changelog: https://github.com/NVIDIA/TileGym/compare/v1.2.0...v1.3.0