v0.3.2

What's Changed

• docs: Update configs for documentation to host on Read the Docs by @peterjunpark in #981
• [Kernel]Remove recast iter from sliding window attention kernel. by @amd-nprotaso in #997
• docs: Update docs style to fit style guide by @peterjunpark in #982
• [MoE] Port moe_gemm_2stage (stage1+stage2) to the new fx.* pipeline — fp8 (gfx942 + gfx950) by @coderfeli in #947
• Refactor gfx950 A16W16 GEMM to use the universal kernel by @xytpai in #992
• CI: upgrade PyTorch images to ROCm 7.14 by @coderfeli in #989
• Use stable torch reference API for a16w16 GFX950 GEMM testing by @xytpai in #1000
• docs: add rocm-docs-core package to requirements by @nhaehnle in #1005
• Fix GitHub Pages docs deployment by @gyohuangxin in #1006
• [compiler] isolate mutable state across dynamic if branches by @kefan203 in #1003
• [Expr] Add type-aware extrema and integer ceiling division by @Phil-amd in #995
• [Feat] Add fx.num_warp_threads() as constant accessor by @sjfeng1999 in #996
• [fix] Resolve normalization-related CI regressions by @cschenjunlin in #991
• [Docs] Elaboration for composite type and storage by @sjfeng1999 in #1004
• DSL-ify raw arith float ops in kernels (flash/mla/pa/moe), fastmath v… by @xudoyuan in #930
• [fp4_gemm_4wave] MXFP4 GEMM perf optimize by @benenzhu in #990
• [CI] Skip amdgpu repo during manylinux image dnf install by @coderfeli in #1014
• [Kernel][MI350] Add bias, alibi bias and sink to flash attention by @amd-nprotaso in #960
• [ROCDL] Add Global/BufferLoadAsyncLDS atoms by @sjfeng1999 in #1023
• [ROCDL] Fix wave size for gfx1250 (wave32, not wave64) by @sjfeng1999 in #1024
• [Test] Run every RUN line in the MLIR FileCheck harness by @sjfeng1999 in #1021
• [Kernel] Add row-wise Softmax backward by @jhinpan in #1019
• [Kernel] fp8 conv3d: 8-wave GEMM pipeline + BIG_IN fix by @jiacao-amd in #860
• [FlyToROCDL] Honor fly pointer alignment in ptr.load/ptr.store lowering by @xudoyuan in #1030
• [ROCDL] Accept IntegerAttr in the wait-counter for backward compatibility by @sjfeng1999 in #1029
• [CI] Track Softmax backward in the benchmark dashboard by @jhinpan in #1026
• [Docs] Remove Disclaimer section from README by @coderfeli in #1039
• [CI] Bypass runner git-cache in promote checkout by @coderfeli in #1040
• [GEMM] Refine GFX950 A16W16 async load scheduling by @xytpai in #1041
• [Fix] Let fx.maxnumf inherit the ambient fastmath by @JohnQinAMD in #1035
• [Ext][Coop] Add cooperative warp- and block-scope collectives by @sjfeng1999 in #1031
• [Bugfix] Fix two A-tile loading bugs in the preshuffle GEMM by @JohnQinAMD in #1007
• [Kernel][Perf] Enable the gmem->LDS async copy for f16/bf16 preshuffle GEMM by @JohnQinAMD in #1008
• Release: bump version to 0.3.2 by @coderfeli in #1044
• [Fix] Restore the gfx950 XCD mapping for bf16 dense attention by @JohnQinAMD in #1009
• fix(ci): isolate self-hosted runner artifacts by @jhinpan in #1046
• [Fix] Fix alignment reported by ptr.load/ptr.store lowering by @sjfeng1999 in #1043
• [Kernel][Perf] Speed up the RDNA3 GEMM up to 1.85x by choosing the tile from the shape by @vlluvia in #1028
• [Kernel] Replace raw scf ops with Python control flow in kernels by @xudoyuan in #1025
• [Kernel][Perf] Optimize paged-attention metadata decode by @fsx950223 in #910
• [Fix] crd2idx uses wrong dynamic basis leaf by @sjfeng1999 in #1052
• [ROCDL] Lower approximate exp2 to native intrinsic by @jhinpan in #1037
• [Skills] Fix silently-ignored frontmatter, de-drift kernel skills and CLAUDE.md by @jhinpan in #1027
• [Bugfix][Kernel] Fix NaN from the lazy rescale on a wide score range by @JohnQinAMD in #1033
• [gfx1250][GEMM] Add const-init to the test and support mxscale tile_m=16 by @aoli26 in #1060
• [GEMM] Track gfx950 async LDS loads with asyncmark by @xytpai in #1063

New Contributors

• @peterjunpark made their first contribution in #981
• @nhaehnle made their first contribution in #1005

Full Changelog: v0.3.1...v0.3.2