v0.4.0

What's Changed

• [llvm] Drop four LLVM tuning knobs that never reached codegen by @Phil-amd in #1153
• [CI] Fix cross-process cache IR assertion by @coderfeli in #1171
• [Verion] Bump version to 0.4.0 by @coderfeli in #1174
• [Misc] Fix 5 Dependabot vulnerabilities by @Phil-amd in #1164
• [Ext][Coop] Add more warp-level primitives by @sjfeng1999 in #1167
• [CI] Remove ancestor-of-main gate from PyPI publish workflow by @jli-melchior in #1175
• smem cleanup: move capacity helpers off legacy allocator, migrate tes… by @xudoyuan in #1062
• [Kernel][FA] Fix the DAZ flag never reaching the hardware by @Phil-amd in #1177
• [llvm][Skill] Add an LLVM kernel-tuning and verification skill by @Phil-amd in #1154
• refactor communication ops by @yanboshao in #1116
• [API] Add extension as stable, exclude experimental module paths by @sjfeng1999 in #1181
• Reject global->LDS direct loads on gfx11 by @mgehre-amd in #1122
• [a16w4] Sync moe_2stage_a16wmix from aiter (gfx942 MXFP4 + 7 earlier … by @MHYangAMD in #1182
• [Fix][Coop] Align warp array collective semantics by @sjfeng1999 in #1178
• [Kernel][GEMM] Add gfx950 2-wave preshuffle tiles by @XiaobingSuper in #1183
• [Kernel] Refactor preshuffle GEMM indexing with layout algebra by @coderfeli in #1131
• Support variadic operands in fully expanded tiled GEMM by @coderfeli in #1128
• Revert "Refactor preshuffle GEMM indexing with layout algebra (#1131)" by @coderfeli in #1189
• [CI] Fix dumpir.sh silently succeeding with no command by @Phil-amd in #1187
• [AOT] Export linkable C objects and support CPU-only builds by @coderfeli in #1202
• [Chore][gfx1250] Drop leftovers of the GEMM modules removed in #885 by @jhinpan in #1197
• [Fix][gfx1250] Reject a K that tile_k does not divide in the A8W4 MoE GEMM by @jhinpan in #1199
• [Fix][gfx1250] Time the TDM multicast-add benchmark on the device by @jhinpan in #1200
• [CI] Reduce cooperative test matrix by @coderfeli in #1207
• Issue causal flash q-blocks longest-first by @RichardChamberlain1 in #1186
• [Fix][Examples] Remove duplicated k-stride computation in preshuffle GEMM example by @yicyang in #1208
• [Kernel][Perf] Add GLM-5 and Kimi-K3 MonoKernels by @coderfeli in #1204
• [JIT] Attach exactly one #rocdl.target per gpu.module by @mustafayildirim in #1195
• [Docs] Audit and refresh repository documentation by @coderfeli in #1203
• [AOT] Embed backend runtime into exported objects by @coderfeli in #1209
• [fix] mono kernel glm precision fixed by @james-huang09 in #1213
• [Build] Find ROCm through rocm-sdk before falling back to /opt/rocm* by @xinyazhang in #1142
• [Bugfix][Test] Fix atomic_fmin/fmax Out left at +inf on side-stream launch by @dimitri91209 in #1220
• [Perf] Optimize rmsnorm/layernorm by @cschenjunlin in #848
• [Perf] Tile GLM MXFP4 scales and pair gate/up rows by @james-huang09 in #1241
• [Kernel]add mega moe tp kernel for gfx950 by @binding7012 in #1255
• [DSL] Add a category independent cute algebra framework by @sjfeng1999 in #1260
• [Enh] Add CompiledFunction.max_blocks_per_cu() by @akii96 in #1259
• [DSL] Expose scalar buffer load API by @coderfeli in #1261
• [fix] Build the zipped_divide guide from the tiler shape by @xudoyuan in #1258
• [llvm] Add support for custom LLVM extensions by @Phil-amd in #1190
• [DSL][Runtime] Add ktrace capture with target providers by @coderfeli in #1268
• support scale nonshuffle by @solinzby1 in #1275
• [Bugfix][JIT] Restore compile launches and defer C export preparation by @coderfeli in #1278

New Contributors

• @XiaobingSuper made their first contribution in #1183
• @RichardChamberlain1 made their first contribution in #1186
• @yicyang made their first contribution in #1208
• @mustafayildirim made their first contribution in #1195
• @james-huang09 made their first contribution in #1213
• @dimitri91209 made their first contribution in #1220
• @akii96 made their first contribution in #1259

Full Changelog: v0.3.4.1...v0.4.0