b76312b024
All checks were successful
Smoke Test / smoke (pull_request) Successful in 34s
Monitoring tool for tracking when TurboQuant lands in upstream llama.cpp and Ollama. Checks GitHub PRs/issues for TurboQuant, PolarQuant, QJL mentions, checks Ollama releases, and compares fork freshness against upstream. scripts/upstream_watch.py — Automated monitoring: - Search llama.cpp/ggml/ollama for TurboQuant keywords - Check Ollama releases for KV cache mentions - Compare fork commit age vs upstream - Generate report or JSON output - Run: python3 scripts/upstream_watch.py --since 30d docs/upstream-watch-report.md — Current status: - TurboQuant has NOT landed upstream yet - Fork is CURRENT with upstream llama.cpp - Continue using TheTom/llama-cpp-turboquant fork
TurboQuant
KV cache compression for local inference on M4 Max MacBook Pro.
What
TurboQuant (Google, ICLR 2026) is a three-stage KV cache compression method:
- PolarQuant — WHT rotation + polar coordinates + Lloyd-Max codebook (~4.2x compression)
- QJL — 1-bit quantized Johnson-Lindenstrauss residual correction
- TurboQuant — PolarQuant + QJL = ~3.5 bits/channel, zero accuracy loss
Why
Unlock 64K-128K context on qwen3.5:27b within 32GB unified memory. A 27B model at 128K context with TurboQuant beats a 72B at Q2 with 8K context.
Status
See issues for current progress.
Roles
- Strago: Build spec author
- Cid: Implementation, benchmarks, deployment
- Locke: Research support, upstream watch
- John: Quality review
- Frankie: Coordination
Source Repos
- TheTom/llama-cpp-turboquant — llama.cpp fork with Metal
- TheTom/turboquant_plus — Reference impl, 511+ tests
- amirzandieh/QJL — Author QJL code (CUDA)
- rachittshah/mlx-turboquant — MLX fallback
Docs
- BUILD-SPEC.md — Full build specification (Strago, v2.2)