Reading the repo — stats and the shipped feed land in about ten seconds…
Quantization, kernels, runtime and inference engine for mobiles, wearables, smart home and robots.
6.1k 515 112 openLast shipped 10d ago
Quantization, kernels, runtime and inference engine for mobiles, wearables, smart home and robots.
A deterministic 0–100 hygiene score — README, license, CI, tests, docs, and freshness.
Who ships this repo — author concentration and the bus factor across the last 300 mainline commits.
How welcoming this repo is to contributors — issue throughput, close time, responsiveness, and good-first-issue count.
What this project is built on — dependency count by ecosystem, the license mix, and anything worth a legal look before you adopt it.
Whether this project's CI can be trusted — pass rate, run times, flaky runs, and which workflow is the weak link.
Grounded in cactus's README, structure, and recent commits — answers won't invent code they haven't seen.
A Monday email with what shipped, in plain English — no account needed.
Showing raw commit titles for the newest commits. Sign in to generate AI summaries.
Released version 2.2.0
Repo Wrapped is now at version 2.2.0, a minor update from the previous 2.1.0 release.
Fixed Parakeet audio encoder padding leak
Corrected a critical bug where padded audio frames were leaking into the transcription output, causing short clips to lose their last word and generate hallucinated words. The encoder now properly masks padded frames, uses the correct mel-scale normalization settings (Slaney instead of HTK), and respects the configured audio time limit instead of a hardcoded example length.
Deduplicate shared memory-mapped weights
Fixed weight file deduplication across component graphs by implementing a registry system that ensures the same memory-mapped weight file is loaded only once and shared across multiple graph instances, reducing memory overhead.
Enhanced transpiler with cache attention
Improved the custom transpiler with key-value cache attention support for generic prefills and expanded component splitting capabilities. This includes updates to fusion handling, lowering cache management, and model profiling with new test coverage for transpiler reporting and component split generation.
A floor, not a guess: counts only commits whose author, co-author trailer, or message explicitly credits an AI tool (Claude, Copilot, Cursor, aider, Codex…). Based on 30 mainline commits. Unattributed AI code isn't counted here — the full audit estimates that separately.
