v0.4.1

What's Changed

• [Skills] Add review-pr: advisory PR review adapted from aiter and ATOM by @jhinpan in #1279
• [Kernel][gfx1250] Support K tails in the BF16 GEMM by @jhinpan in #1230
• [fix] Make make_tile treat every argument list as modes by @xudoyuan in #1284
• [fix] Support waves_per_eu min,max range for full VGPR utilization by @jli-melchior in #1184
• [Tool][Debug] Inject hand-edited assembly into a JIT cache entry by @Phil-amd in #1188
• [fix][CI] Run aiter's install_triton.sh to match aiter's Triton version by @jli-melchior in #1300
• [ROCDL] Gate the TDM tile_dim0 span check on padding by @xudoyuan in #1302
• [Skills] Ensure flydsl-kernel-authoring follows Anthropic best practices by @danielholanda in #1305
• [Skills] Add evals and skill cards for kernel authoring and trace analysis by @danielholanda in #1304
• [llvm] bump llvm to 1df311a6 by @jli-melchior in #1243
• Align bf16 implicit Conv3d with aiter by @huizzhan in #1217
• [Fix] Bitcast small floats across converted ops in convert-fly-to-rocdl by @Phil-amd in #1290
• [gfx1250] Add A8W8/A8W4/A4W4 compute-bound GEMM by @aoli26 in #971
• [Core] Add four concrete layout categories by @sjfeng1999 in #1292
• [CI] Fix the SGLang nightly setup after sgl-kernel moved into sglang.kernels.aot by @jhinpan in #1287
• [Kernel] Reuse typed APIs in GLM and Kimi-K3 monokernels by @coderfeli in #1310
• [compiler] [llvm] xnack perf by @jli-melchior in #1308
• [Release] Bump version to 0.4.1 by @jli-melchior in #1312
• 1,Balance the prologue's extra wave_m == 1 barrier so both groups' ep… by @solinzby1 in #1306
• [Dialect][ROCDL] Add gfx120 integer WMMA and an INT4 GEMM by @qiangpan2 in #1266
• fix(coop): preserve declared operator element types by @sjfeng1999 in #1185
• [Kernel] Remove redundant scalar, vector, and IR wrappers by @coderfeli in #1262
• [Kernel] Replace raw memory glue and manual reductions with fx APIs by @coderfeli in #1315
• [Skills] review-pr: review a PR when a maintainer adds the ai-review label by @jhinpan in #1311
• [Skills] review-pr: strip the API key and show why model calls failed by @jhinpan in #1317
• kernel-trace-analysis: fix SKILL.md nits so the evals pass by @danielholanda in #1322
• [Ext][Coop] Add more block-level primitives by @sjfeng1999 in #1314
• [LLVM] Roll back to 941a04e and cherry-pick unclaused VMEM patches by @jli-melchior in #1327

New Contributors

• @danielholanda made their first contribution in #1305
• @huizzhan made their first contribution in #1217

Full Changelog: v0.4.0...v0.4.1