f60604ddcc
All checks were successful
Smoke Test / smoke (pull_request) Successful in 12s
- Created comprehensive GENOME.md with full codebase analysis - Added architecture diagram (Mermaid) - Documented entry points and data flow - Identified key abstractions - Mapped API surface (C, Metal, CLI) - Identified test coverage gaps - Documented security considerations - Added basic test suite (9 tests passing) Key findings: - 73.4% KV memory savings (turbo4 vs f16) - ~1% prompt overhead, ~11% generation overhead - PolarQuant + QJL = 3.5 bits/channel - Metal shaders exist on feature branch - CPU reference incompatible with Metal dequant - QJL infrastructure present but disabled Test coverage gaps: - No unit tests for encode/decode - No integration tests - No perplexity runner (corpus exists) - No Metal vs CPU parity tests Security considerations: - Buffer overflow risk in bit packing - No constant-time implementation - No safety wrapper for C/C++ code
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)