27ebfa3525
All checks were successful
Smoke Test / smoke (pull_request) Successful in 13s
Test matrix runner (benchmarks/run_test_matrix.py) implementing all acceptance criteria from #11: Quality Tests: - 10 practical prompts with expected-pattern matching - Perplexity proxy (WikiText-2 chunks) - Needle-in-Haystack at 8K/16K/32K contexts - Multi-turn context retention (prompt #7) Performance Tests: - tok/s at 4K/8K/16K context - TTFT proxy measurement - Peak memory (macOS/Linux) - Context ceiling binary search Outputs: - JSON: reports/test-matrix-YYYY-MM-DD.json - Markdown: reports/test-matrix-YYYY-MM-DD.md - Go/No-Go assessment with issue list Smoke test: 10/10 quality, 3/3 needle-in-haystack on qwen2.5:7b. Refs: Timmy_Foundation/turboquant#11
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)