4b2b8fc081
Some checks failed
Smoke Test / smoke (pull_request) Failing after 8s
Issue #55 — security hardening for PolarQuant Turbo4 codec. - Add llama-turbo-safety.h/.cpp with inline input validation: * dimension must be positive power of 2 * all pointers non-NULL * decode norm > 0 (zero-norm guard) - Inject validation into encode/decode via TURBOQUANT_CHECK macro - Implement branchless nearest-centroid search (fixed 16-iteration loop) - Document bounds safety in Metal kernels - Add CMake option TURBOQUANT_ENABLE_SANITIZERS for ASan/UBSan integration - Add tests/test_safety.py (smoke test wrapper) Validation: standalone roundtrip tests pass; ASan build passes; constant-time properties verified (fixed loop counts + branchless selection). Closes #55
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
- Project Status — Full project status and build specification