v1.4.0

What's Changed

• cutile recurrent_gated_delta_rule: autotune + persistent variant & other updates by @hannahli-nv in #129
• [recurrent_gated_delta_rule] Add stream-v kernel and tuning huristics & other updates by @hannahli-nv in #131
• ci: add request-nvskills-ci.yml workflow by @hannahli-nv in #133
• Add tilecpp backend by @hannahli-nv in #134
• feat(softmax): add explicit multi-wave kernel option by @liqiangxl in #128
• Initial NVSkills-CI onboarding for TileGym skills by @hannahli-nv in #135
• benchmark_fn_cupti: surface kernel_name and kernel_times in result dict & other updates by @hannahli-nv in #138
• Skill Evals: tilegym-cutile-python by @hannahli-nv in #144
• Add pre-commit CI checks and skills-signature validation by @hannahli-nv in #137
• Add FlashInfer-style transformer kernel inventory by @hannahli-nv in #139
• Fix pre-commit format failure and update format.sh by @hannahli-nv in #149
• Skill Evals: cutile-to-triton + improve-cutile-perf + cutile-to-julia by @hannahli-nv in #152
• add some cuTile kernels for Liger suites & Update mla autotune config & other updates by @hannahli-nv in #151
• perf(liger/cutile): tune cross_entropy BLOCK_SIZE by GPU arch & Modularize transformer HF benchmark & other updates by @hannahli-nv in #153
• [perf][cutile]autotune RoPE occupancy & Fix flashinfer cutile kernels & other updates by @hannahli-nv in #156
• Update CI image to CUDA 13.3 / Ubuntu 24.04 and shard the GPU test & benchmark jobs by @hannahli-nv in #158
• ci: refresh pytest-split .test_durations by @github-actions[bot] in #159
• Update recurrent_gated_delta_rule and matmul & Add FlashInfer-backed transformer kernel inventory workflow by @hannahli-nv in #160
• ci: add transformer inventory runtime test to GitHub CI by @lirundong in #162
• [cutile] matmul: expose ct.load cost (latency) as an autotune knob & other updates by @hannahli-nv in #161
• fix result mismatch for Test_GroupGemm & Add Triton backend support for ops kernels & other updates by @hannahli-nv in #163
• ci: refresh pytest-split .test_durations by @github-actions[bot] in #165
• [tests]Shrink matmul 16384^3 config to 8192^3 to avoid SM89 OOM & Update tilecpp tests & other updates by @hannahli-nv in #164
• ci: refresh pytest-split .test_durations by @github-actions[bot] in #167
• Update matmul op args & Update test_mla_decoding_split_kv parameter & other updates by @hannahli-nv in #166
• Raise cuda.tile's autotune launch-timeout budget when requested via env & other updates by @hannahli-nv in #168
• ci: refresh pytest-split .test_durations by @github-actions[bot] in #170
• Bump version from 1.3.0 to 1.4.0 by @hannahli-nv in #172

New Contributors

• @github-actions[bot] made their first contribution in #159
• @lirundong made their first contribution in #162

Full Changelog: v1.3.0...v1.4.0