Skip to content

Composed MoE Engine v0.3.2 - GPU bug fix & real dense baseline

Latest

Choose a tag to compare

@Mattral Mattral released this 17 Jun 09:06

Composed MoE Engine v0.3.2 - GPU bug fix & real dense baseline

This patch release fixes two runtime bugs exposed by real GPU runs, aligns dependencies, and integrates the second Colab benchmark data.

Fixes

  • Triton router kernels: K now declared as tl.constexpr to resolve static_range compilation error. Two regression tests added (147 total).
  • Dependency drift: pytest-repeat added to requirements.txt so --count works with the documented install path.

Benchmarks

  • Integrated second Colab run data (cpu_results_colab_run2.json), including dense baseline ratios (8.3×–12.3× routing overhead at ep_size=1).
  • Updated BENCHMARKS.md and RESULTS.md with v0.3.2 summary and populated dense baseline table.

Documentation

  • Updated README test badge and version note.
  • Revised docs/testing.md to reflect both pip install -e ".[dev]" and requirements.txt paths.