John Smith f22916bd51 Engine skeleton, reverse-engineering specs, and validated dequant kernels ...

- docs/halogen: wire protocol, checkpoint format, host logic, CLI, and
  quantized weight formats (fp8r + q4c/NVFP4 validated against base model
  weights on GPU: corr 0.9996 / 0.988; i4l layout open)
- engine: C++20 skeleton — .hgn loader (mmap + full table validation),
  wire-protocol server (PING/INFO/CSTAT/GEN/X) with the complete GEN
  grammar, deterministic stub generator; 14/14 conformance tests pass
- converter: .hgn inspector + safetensors->.hgn writer (bf16/f32/f16/i32/i64,
  fp8r absmax-448, q4c codebook quantization)
- engine/kernels: HIP gfx1151 dequant kernels (fp8r, q4c) validated on
  Strix Halo against Qwen3.8-27B base weights
- deploy/env.sh: Strix Halo build environment recipe (ROCm 7.14 wheels)

2026-09-08 20:32:12 -04:00

..

env.sh

Engine skeleton, reverse-engineering specs, and validated dequant kernels

2026-09-08 20:32:12 -04:00