AOTriton 0.13 Beta

Overview

To match the latest PyTorch release cycle
(released every other month), AOTriton has shifted to a maintenance-feature
release model, i.e., within four months AOTriton will make:

• A maintenance release, which ships changes without changing the compiler
  and tuning database
• A feature release, which synchronizes Triton with upstream, adds features
  to kernels, and updates the tuning database.

Specific to AOTriton 0.13: this is a maintenance release, with no Triton
compiler or database changes.

Major Changes

• AOTriton Template Instantiation (ATI)
  • Separation of the code generator from the kernel and build rules
  • Now python will be installed as the aotriton package in the venv
    for code generation
    • Conceptually the aotriton package could be considered a "compiler"
      that translates JIT kernels to the relevant AOT source code and
      kernel build rules
    • Related unit tests can be found at python/test
  • Triton kernel and build rules are migrated to modules/<family>/
    directories
• All .aks2 files are now stored in uncompressed zip containers,
  keeping their names inside the archive instead of exposing them to the
  file system
  • Previously the AKS2 file names had to avoid special characters to stay
    compatible across many layers, including but not limited to 1) the
    Win32 API and 2) Bash/Perl/Python tooling, which may not handle
    special characters well
  • Now that the names live inside the container, this restriction is
    lifted and the AKS2 naming scheme is changed to a more human-readable
    format, for example:
    Q='*fp16:16';BLOCK_DMODEL=16;PADDED_HEAD=False;ENABLE_DROPOUT=False;CAUSAL_TYPE=0;BIAS_TYPE=0
• A performance visualization tool is added to the AOTriton dashboard
  • See Performance Visualization for a screenshot
• Proper Win32 support
  • Full Unicode path support
    • Dropped dlfcn-win32, whose GetModuleFileNameA is inherently
      non-Unicode; dladdr is reimplemented with GetModuleHandleExW
      and GetModuleFileNameW
  • Replaces intptr_t with typedef fd_t for int/HANDLE
• Add debugging facility, enabled with AOTRITON_DEBUG_LEVEL
• Enable ASAN Build with TRITON_ENABLE_ASAN=1 + TheRock Clang
  • ASAN packages are released with the +asan version suffix, e.g.,
    aotriton-0.13b+asan-images-amd-gfx942.tar.gz
• Add gfx1250 Tech Preview Build
  • Announced by isArchTechPreview
  • Unreleased hardware will not be announced by
    isArchExperimentallySupported for wider enablement and testing.

Minor Changes

• Remove legacy v2python/v2src directories
• Update .claude/agents/prwriter.md to write more concise PR messages.

Performance Visualization

(Exact numbers are masked per the Performance Disclosure Policy)

What's Changed (Auto-generated)

• Remove Legacy Code and Files by @xinyazhang in #174
• test/attn_torch_function: reset persistent atomic tile counter between is_testing pre-flight and real attn_fwd by @mgehre-amd in #180
• Use SHA-256 file names by @xinyazhang in #176
• Group AKS2 files into uncompressed ZIP archive by @xinyazhang in #177
• Performance Visualization by @xinyazhang in #178
• AOTriton Template Instantiation by @xinyazhang in #186
• Rewrite Agent File prwriter.md by @xinyazhang in #189
• Support Windows wide paths for packed kernels by @jammm in #184
• test(ati): add ATI unit tests under python/test/ati/ (193 tests) by @xinyazhang in #190
• Logging with AOTRITON_DEBUG_LEVEL by @xinyazhang in #192
• Port gfx1250 and TheRock ASAN Support to Mainline by @xinyazhang in #182

New Contributors

• @mgehre-amd made their first contribution in #180

Full Changelog: 0.12b...0.13b