78352b0a91
Some checks failed
Smoke Test / smoke (pull_request) Failing after 4s
50-turn multi-phase conversation test that detects quality degradation under sustained context pressure. Supports Ollama and llama-server backends with KV cache type configuration. Phases: code_gen -> debug -> refactor -> test -> iterate Metrics: quality score, coherence drift, hallucinated references, repetition ratio, prompt relevance. Includes --compare mode for side-by-side KV type comparison. Acceptance: run on both TurboQuant and FP16, compare results.
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)