expXa

Research

Q6_K

Studies involving Q6_K weights: the two-token expert kernel, Q6_K CPU tiles and the Qwen3.6-35B-A3B Q6_K profile.

  1. expXa vs PXA 4 on Qwen3.6-35B-A3B Q6_K

    Matched dual-V100 comparison on Qwen3.6-35B-A3B Q6_K: both engines with and without MTP, 512–65,536 input tokens, every run and correctness checks.

    No change

  2. expXa vs PXA 4 on Qwen3-30B-A3B Q4_K_M

    Matched V100 comparison of expXa and PXA 4.0.1 on Qwen3-30B-A3B Q4_K_M with 40 CPU expert layers: per-run results, method, and how expXa went from 49 s to 24 s.

    Adopted

  3. Warp sharing in expert and attention kernels

    Testing warp-level sharing of expert IDs, weight fragments and attention scalars in expXa's Volta kernels on two V100s. No change earned adoption.

    Not adopted

  4. VNNI CPU tiles in serving

    Before/after serving measurement of hoisted and AVX-512 VNNI Q4_K/Q6_K CPU tiles in expXa on two MoE models: outputs unchanged, no measurable serving speedup.

    Adopted

  5. Reordering expert work for weight reuse

    Grouping MoE expert jobs by expert and traversing row bands on V100: about 4% less operator time at 255 active jobs, slower for small batches and the paired kernel.

    Not adopted

Other topics: Attention AVX-512 VNNI CPU kernels CUDA kernels MoE CPU offload PXA comparison Q4_K_M NVIDIA V100