Major Changes from Release 0.11 Beta

API Changes

• BREAKING Varlen LSE tensor shape changes to (H, Total_seqlen) (#149).
• Two new varlen layouts: PaddedVarlen and StridedVarlen for
  TransformerEngine compatibility — V3 API only (#150).
• BREAKING LazyTensor::acquire now receives self instead of cookie
  (#164).
• Support hdim_qk != hdim_vo; dispatcher infers both from Q and V (#135).
• attn_options::deterministic = true forces the deterministic split-kernel
  backward path (#134).
• All V2 launch functions marked [[deprecated]]; removal planned for next
  feature release (#164). Exception: check_gpu and
  debug_simulate_encoded_softmax are un-deprecated in aotriton/flash.h
  (#173).

GPU Targets

• Experimental: gfx1103, gfx1152, gfx1153 iGPU support added (#138, #142).
• gfx1100 and gfx1151 promoted out of experimental (#173).
• gfx11xx split into gfx110x (RDNA 3) and gfx115x (RDNA 3.5) build
  packs for independent release tarballs (#173).

Performance

• gfx950: pipelining + XCD remapping; forward improves from ~753 to
  ~904 TFLOPS on MI355X with Triton mainline (hdim=128, non-causal) (#162).
• Triton compiler bumped to 9c446b40 (ROCm/triton, Apr 2026).
• AITER ASM kernels updated to v0.1.11 for gfx942 and gfx950 (#158).
• Updated tuning databases for gfx942, gfx950, gfx1100, and gfx1201 (#172).

Bug Fixes

• GQA + attention bias backward produced wrong gradients; fixed by moving
  bias pointer init inside the Q-head loop (#170).
• Unsupported AOTRITON_TARGET_ARCH now fails loudly with a diagnostic
  instead of a cryptic downstream argparse error (#171).

Build and Dependency Changes

• Tuning database sharded into per-arch files under
  v3python/database/<vendor>/<arch>/ (#133).
• pybind11 and incbin submodules removed; pybind11 now comes from pip
  (#152).
• Alternative Triton wheel YAML config mechanism added; pyyaml is a new
  build dependency (#132, #153).
• __signature__ now includes AOTRITON_GIT_TREESHA1 (root tree SHA1,
  injectable via env var) (#173).

Minor Changes from Release 0.11 Beta

• Tests default to V3 API; FWD_IMPL env var selects attn_fwd backend
  (#158).
• pkg-config added as a build dependency (#159).
• Windows build fixed for gfx942 affine kernels (wide pstring_view) (#156).

Known Problems

• gfx1100: a small number of unit tests fail due to compiler accuracy issues.
• gfx1201: a small number of unit tests fail due to a hipblasLt GPU segfault.
• gfx950: hdim=48/80 backward kernels disabled pending a compiler fix.
• gfx950: attn_fwd with hdim=16 silently rounds up to hdim=32 at
  runtime until the upstream compiler bug is resolved.
• AITER ASM kernels for dropout, SWA/GQA, and MQA/GQA fall back to Triton.

Notes Generated by GitHub

What's Changed

• Use MOLD linker in CI scripts by @xinyazhang in #125
• Update README.md to include PyTorch 2.9 by @xinyazhang in #126
• Add CI Script to Test Triton Upstream by @xinyazhang in #127
• Fix warning 'missing .note.GNU-stack section implies executable stack' by @xinyazhang in #129
• Tuner V3 Part 1: Celery Execution Framework by @xinyazhang in #131
• Alternative Wheels by @xinyazhang in #132
• Sharding the Database by @xinyazhang in #133
• Misc. build changes by @xinyazhang in #136
• Add deterministic algorithm support to v3::flash::attn_options by @xinyazhang in #134
• add gfx1103 iGPU support by @lamikr in #138
• Add support of hdim_qk != hdim_vo by @xinyazhang in #135
• Add gfx1152 and gfx1153 iGPU support by @roberteg16 in #142
• Fix window size in v3 by @alugorey in #144
• Port Changes and Fixes from release/0.11 to main by @xinyazhang in #147
• API Change: Use Compact LSE Tensor for Varlen Inputs by @xinyazhang in #149
• API Change: Support Additional Varlen Memory Layouts by @xinyazhang in #150
• [ISSUE-98]: Alternative Wheel Mechanism by @aoguntayo in #153
• Fix Windows build for gfx942 affine kernels: use wide string for pstring_view by @jammm in #156
• Sunset Git Submodules Other Than Triton by @xinyazhang in #152
• Update README to include pkg-config dependency by @xinyazhang in #159
• Remove stray line in CMakeLists.txt by @sstamenk in #160
• Bump to AITER asm v0.1.11 by @xinyazhang in #158
• Enable Pipelining and XCD Remapping on gfx950 GPUs by @xinyazhang in #162
• Migrate the Tuner to V3 API by @xinyazhang in #164
• fix(gpu_targets): fail loud when arch filter drops every requested target (#169) by @1fanwang in #171
• Tuner V3.5 and Tuning Database. by @xinyazhang in #170
• Remove unnecessary Windows ccache guard by @astrelsky in #167
• 0.12b Operator Tuning Database and related Tuner Changes. by @xinyazhang in #172
• Misc. Changes for 0.12b Release by @xinyazhang in #173

New Contributors

• @lamikr made their first contribution in #138
• @alugorey made their first contribution in #144
• @aoguntayo made their first contribution in #153
• @1fanwang made their first contribution in #171
• @astrelsky made their first contribution in #167

Full Changelog: 0.11b...0.12b