Skip to content

v0.1.0 — Initial public release

Latest
Latest

Choose a tag to compare

View all tags
@danielfleischer danielfleischer released this 19 Apr 14:26
· 4 commits to main since this release
v0.1.0
0e7662e
This commit was signed with the committer’s verified signature.
danielfleischer Daniel Fleischer
SSH Key Fingerprint: 1m5EWa4RlXDHND+CUPKDLGSNcLK6V5sMfMHaiFJkeAQ
Verified
Learn about vigilant mode.

Triton8 v0.1.0

LLM-driven optimization loop that transforms PyTorch code into fast Triton kernels for Intel XPU. Claude Code explores kernel variants in a branching trial tree, guided by a curated knowledge base and hardware profiler feedback.

Highlights

  • Automated trial loop — configurable multi-trial optimization with branching from best results
  • Tool suite — static analysis, validation, benchmarking (via AI-bench), and VTune GPU profiling
  • Knowledge base — correctness constraints, XPU-specific patterns (tensor descriptors, GRF tuning, tile swizzling), fusion
    heuristics, and annotated examples
  • Kernel templates — GEMM, GEMM+epilogue, and reduction starting points
  • KernelBench Level 2 results on Intel Battlemage G21 / Arc Pro B50

Requirements

  • Intel XPU hardware, PyTorch with XPU support, Intel XPU Backend for Triton
  • Claude Code
  • VTune Profiler (optional)

⚠️ Pre-1.0: APIs, tool interfaces, and KB schema may change between releases.

Assets 2
Loading