Thriving Thriving Python PY Shell SH
31
strix-halo-llamacpp Nathanw1014/strix-halo-llamacpp

Performance-tuned llama.cpp for AMD Strix Halo (gfx1151): FA + MoE-prefill fixes with a bundled current Mesa driver. Vulkan and HIP; portable dir, Docker, and distrobox.

HEALTH 83 / Healthy
// solo builder// release machine amdflash-attentiongfx1151llama-cppllm-inferencelocal-llmradvrocmryzen-ai-maxspeculative-decoding

// readme

Strix Halo llama.cpp toolbox (FA + MoE-prefill fixes)

A ready-to-run llama.cpp for AMD Strix Halo (Ryzen AI Max+ 395 / Radeon 8060S / gfx1151), tuned for long-context, quantized-KV workloads. It bundles a set of Flash-Attention and MoE-prefill fixes plus a current GPU driver, so quantized KV cache is fast instead of a penalty.

The measurements behind these fixes (matrices, methodology, raw data) live in the companion evidence pack.

Speedups at a glance

v0.6.4 payload vs stock master, same box, same session (amd_iommu=off, pp512 / tg32 t/s, -b 512 -ub 512, -r 3, captured 2026-08-18 between 05:28 and 06:04 UTC). Baseline is stock upstream 9f0d017 at f16 KV, the commit this release merged; fixed is the released v0.6.4 payload (baf6360b on strix-halo-vulkan). Both arms were built with the same pinned glslc and run against the same pinned Mesa, so only llama.cpp differs. Every cell traces to a raw run under benchmarks/results/glance-20260818/.

| Model (arch) | KV |…

The Undervalued Score +

How much a project earns versus how much attention it actually gets. Above 50 means the work is outrunning its audience. Recomputed nightly from commit velocity, contributor effort, issue resolution, fork utility, release cadence, and project maturity — divided by a logarithmic reach factor.

score  = signal / reach

signal = 0.25·commit_velocity   // commits in last 90 days (cap 30)
       + 0.20·contributor_work  // unique authors × velocity (cap 100)
       + 0.20·issue_resolution  // closed ÷ total issues
       + 0.20·fork_ratio        // forks ÷ stars (proxy for real usage)
       + 0.10·release_cadence   // releases in 90 days (cap 3)
       + age_bonus              // +0 to +0.30 after 6 months
       + homepage_bonus         // +0.05 if homepage is set

reach  = log₁₀(stars + watchers + 10)
The Health Score +

Is the project alive and maintained right now? A 0–100 pulse recomputed nightly from commit recency, rhythm, how fast issues close, and how quickly PRs get merged.

health = 0.35·recency       // days since last commit (90d decay)
       + 0.25·cadence       // commit rhythm consistency
       + 0.20·issue_health  // closed ÷ total issues
       + 0.20·pr_health     // merged ÷ total PRs
Health bands +

The colour and label on every card come straight from the health score.

Healthy   80 – 100   active, responsive, regular releases
Stable    60 – 79    maintained, steady, no alarms
Quiet     40 – 59    slowing down — watch this one
At Risk    0 – 39    going dark · candidate for rescue
// Tags — what each label means +

Tags are independent behavioral signals computed nightly. A project can hold multiple at once. They drive the home page sections.

solo_builder      one person holds > 80% of commits (last 180d)
needs_contributors has open "help wanted" or "good first issue" labels
hidden_gem        < 100 stars · active in last 3 months · documented
legacy_hero       repo > 5 years old · committed this year
fork_magnet       forks/stars > 0.5 · used as template or dependency
release_machine   5+ releases in the last 90 days
under_pressure    > 10 open issues · ≤ 2 contributors · health ≥ 60
community_watch   watchers > stars · devs tracking before the public
community_hub     GitHub Discussions enabled · > 20 discussions
funded            maintainer has active funding channel
Why rank against stars at all? +

Stars are an outcome, not effort. A project with 8 stars and daily commits is doing more interesting work than one coasting on 8k. We measure the building, then divide by the attention already received — so the genuinely undervalued rise to the top.

// stars   = lagging indicator
// commits = leading indicator
// we rank by the leading one