9c5f2fd06b
Some checks failed
Smoke Test / smoke (pull_request) Failing after 14s
Adds complete QJL (Johnson–Lindenstrauss residual correction) Metal GPU kernel integration: - ggml/include/ggml.h: add GGML_TYPE_TURBOQUANT_QJL type and helpers - ggml/src/ggml-metal.metal: QJL encode/decode kernel signatures - ggml/src/ggml-metal.m: Metal PSO registration + proper dispatch - src/llama.cpp: KV allocation, projection matrix, fused decode path - CMakeLists.txt: build all components with Metal support - include/llama.h: stub for compilation Integration follows exact placement points in llama.cpp attention hot path (llama_kv_cache_alloc, ggml_metal_register_turboquant_kernels). Closes #133
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