Skip to content

v2.2.0 - MTP, wall power, and community flow

Choose a tag to compare

View all tags
@hogeheer499-commits hogeheer499-commits released this 17 May 13:31
· 160 commits to main since this release
v2.2.0
de24847

May 17 organization and evidence update for the Strix Halo local LLM guide.

Highlights:

  • Added the Qwen3.6 MTP IQ4_XS-Q8nextn llama-server route: 90.80 t/s average over six prompts, 110.61 t/s best prompt. Scoped as speculative server evidence, not a broad 100 t/s claim.
  • Added Qwen3-Coder Q4_K_S llama-server ngram follow-up: best average 95.21 t/s, below the direct 98.51 t/s speed-first headline.
  • Imported community wall-power evidence from issue #6 into data/community_power.csv, COMMUNITY_RESULTS.md, and POWER_BASELINE.md.
  • Added a dedicated power / efficiency issue template for wall-power, board-power, UPS, and smart-plug reports.
  • Reworked CONTRIBUTING.md and the benchmark template so future community reports include the metadata needed for reproducibility.
  • Updated SHARE.md and README changelog so public sharing text matches the current evidence.
  • Clarified the historical v1.0.0 release note now that v2.2.0 is the current latest release.
  • Closed already-integrated community report issues with links to the docs/data they now support.

Validation:

  • GitHub Actions Validate passed for commit 932ee31.
  • python3 scripts/validate_repo.py passed locally.

The live main branch remains the source of truth for ongoing benchmark and community-data updates.

Assets 2
Loading