w8a8

W8A8/W4A8 inference + optimized SDPA on Apple Silicon — unlocking unused INT8 TensorOps in M5 for 1.2–1.9× faster LLM prefill, plus FlashInfer-inspired GQA decode attention for up to 1.6× SDPA speedup, built as MLX custom primitives.

Open, bit-reproducible low-precision number format (b-posit / AI-Posit) for AI inference — 30–50% less memory, with results identical to the bit on any GPU, CPU, or RISC-V. Reference + conformance suite + CORE-ET RTL

Reproducible single-GPU Qwen3.6-35B-A3B W8A8 inference tuning for Hygon K100AI (gfx928) on vLLM 0.18.1.

Bounded Posits at IEEE Tensor-Core Throughput on Commodity NVIDIA Blackwell, with Bit-Exact Reproducibility