NVIDIA/TileGym v1.0.0

v1.0.0

Repository: NVIDIA/TileGym

Tag: v1.0.0

Published: 2026-03-11T00:21:54Z

Prerelease: yes

Release notes:

What's Changed

• [Bug fix] use padding_mode inside the kernel to process elements out of boundary for softmax by @xjmxyt in https://github.com/NVIDIA/TileGym/pull/1
• [Bug fix] use ct.gather ct.store for softmax's no-tma op by @yifeis-nv in https://github.com/NVIDIA/TileGym/pull/2
• Add PR bot to repository by @arjkesh in https://github.com/NVIDIA/TileGym/pull/3
• Update README.md by @xjmxyt in https://github.com/NVIDIA/TileGym/pull/5
• remove dead code in silu_and_mul kernel - creates output offsets (for 1D), expect n_elements param... but no need... by @lessw2020 in https://github.com/NVIDIA/TileGym/pull/6
• Initialize TileGym CI by @arjkesh in https://github.com/NVIDIA/TileGym/pull/4
• Use ruff formatter, introduce helper dev script by @arjkesh in https://github.com/NVIDIA/TileGym/pull/11
• Introduce job timeouts, speed up builds by @camille-004 in https://github.com/NVIDIA/TileGym/pull/9
• [FEA] add gelu & relu by @xjmxyt in https://github.com/NVIDIA/TileGym/pull/13
• Update dockerfile to use cuda 13.1 base image by @arjkesh in https://github.com/NVIDIA/TileGym/pull/12
• [Fix] Refactor nightly skip logic by @arjkesh in https://github.com/NVIDIA/TileGym/pull/8
• Add automatic header checks and formatting by @arjkesh in https://github.com/NVIDIA/TileGym/pull/14
• Standardize softmax.py to avoid numpy dependency by @lessw2020 in https://github.com/NVIDIA/TileGym/pull/16
• [Update] update kernels and reformat codes by @hannahli-nv in https://github.com/NVIDIA/TileGym/pull/18
• [FEA] Add dropout by @hannahli-nv in https://github.com/NVIDIA/TileGym/pull/19
• Split-K reduction kernel cleanup by @lessw2020 in https://github.com/NVIDIA/TileGym/pull/21
• Fix: moe_align_block_size() supports non-power-of-2 num_experts by @huanghua1994 in https://github.com/NVIDIA/TileGym/pull/24
• Update autotuner: use experimental autotuner in cutile-python by @xjmxyt in https://github.com/NVIDIA/TileGym/pull/25
• feat: chunked softmax implementation for large column size by @aghilann in https://github.com/NVIDIA/TileGym/pull/17
• [Update] Add benchmark and autotune for group_gemm by @xjmxyt in https://github.com/NVIDIA/TileGym/pull/26
• Fix benchmark failure cases by @arjkesh in https://github.com/NVIDIA/TileGym/pull/27
• Format benchmark files as json, add perf thresholds by @arjkesh in https://github.com/NVIDIA/TileGym/pull/15
• feat: RMSNorm backward pass kernels by @aghilann in https://github.com/NVIDIA/TileGym/pull/29
• Split-K reduction: remove un-needed scaling via INV_LOG_2 by @lessw2020 in https://github.com/NVIDIA/TileGym/pull/22
• [fix] Update benchmark sparse checkout by @arjkesh in https://github.com/NVIDIA/TileGym/pull/30
• [FEA] Add bmm by @hannahli-nv in https://github.com/NVIDIA/TileGym/pull/31
• Temporarily avoid job failures due to inconsistent benchmarks by @arjkesh in https://github.com/NVIDIA/TileGym/pull/32
• [Update] Fix bmm issue by @hannahli-nv in https://github.com/NVIDIA/TileGym/pull/34
• [FEA] Add Qwen2-7B module by @hannahli-nv in https://github.com/NVIDIA/TileGym/pull/36
• Update for ragged_bmm moe by @hannahli-nv in https://github.com/NVIDIA/TileGym/pull/37
• Add env "DISABLE_FALLBACK" & fix type hint error & other updates by @hannahli-nv in https://github.com/NVIDIA/TileGym/pull/39
• Add reusable retry workflow for runner availability timeouts by @arjkesh in https://github.com/NVIDIA/TileGym/pull/35
• Add mHC fused kernels and tests by @Edward-lyz in https://github.com/NVIDIA/TileGym/pull/38
• Update some comments by @hannahli-nv in https://github.com/NVIDIA/TileGym/pull/42
• Add tilegym wheel building by @arjkesh in https://github.com/NVIDIA/TileGym/pull/41
• fix matmul illegal address error on DGX Spark by @xjmxyt in https://github.com/NVIDIA/TileGym/pull/44
• fix qwen2 fp16 bug by @hannahli-nv in https://github.com/NVIDIA/TileGym/pull/43
• [Fix] fix num_kv_split becomes 0 by @xjmxyt in https://github.com/NVIDIA/TileGym/pull/45
• Avoid OOM for large GEMM 32k & modify layernorm cutile by @hannahli-nv in https://github.com/NVIDIA/TileGym/pull/50
• Add option to ignore specific wheel validations by @arjkesh in https://github.com/NVIDIA/TileGym/pull/51
• Add road map by @hannahli-nv in https://github.com/NVIDIA/TileGym/pull/52
• [FEA] Add SwiGLU backward pass implementation, test cases and benchmark by @Weili-0234 in https://github.com/NVIDIA/TileGym/pull/46
• Enable experimental_kernel marker by @hannahli-nv in https://github.com/NVIDIA/TileGym/pull/53
• [FEA] Add FlashAttention backward pass implementation, test cases and benchmark by @Weili-0234 in https://github.com/NVIDIA/TileGym/pull/49
• Update README.md by @xjmxyt in https://github.com/NVIDIA/TileGym/pull/54
• Add version for tilegym wheels, update reusable workflow by @arjkesh in https://github.com/NVIDIA/TileGym/pull/55
• Fix import error for experimental marker & support gemma 3 & other updates by @hannahli-nv in https://github.com/NVIDIA/TileGym/pull/57
• Add tilegym homepage to setup.py by @arjkesh in https://github.com/NVIDIA/TileGym/pull/58
• Update MoE by @hannahli-nv in https://github.com/NVIDIA/TileGym/pull/59
• fix torch dependency by @xjmxyt in https://github.com/NVIDIA/TileGym/pull/61
• feat: replace RMSNorm backward with persistent CuTile kernel by @aghilann in https://github.com/NVIDIA/TileGym/pull/60
• Scan for CVEs in wheels, fix python versions by @arjkesh in https://github.com/NVIDIA/TileGym/pull/64
• feat: add CuTile RoPE backward with tests and backward benchmark by @aghilann in https://github.com/NVIDIA/TileGym/pull/62
• A fix for silu_and_mul & Update codes & other updates by @hannahli-nv in https://github.com/NVIDIA/TileGym/pull/67
• Add workflow to prepare release tag and artifacts by @arjkesh in https://github.com/NVIDIA/TileGym/pull/66
• Update moe type hint & Update gitignore & other updates by @hannahli-nv in https://github.com/NVIDIA/TileGym/pull/68
• add cutile kernel skill and Move install_requires dependencies to requirements.txt by @hannahli-nv in https://github.com/NVIDIA/TileGym/pull/69
• Add SECURITY.md with vulnerability reporting instructions & Add SPDX license header to SECURITY.md & other updates by @hannahli-nv in https://github.com/NVIDIA/TileGym/pull/71
• feat: swiglu forward optimizations by @aghilann in https://github.com/NVIDIA/TileGym/pull/63
• feat: chunked fused linear cross-entropy kernel forward by @aghilann in…