cpu perf#2

Merged

richiejp merged 14 commits into

Jun 16, 2026

Commits on Jun 16, 2026

bench: PyTorch reference + ggml GEMM microbench, exact-length pf-bench

richiejp
and
claude
committed
perf: portable AVX-512 CPU build via GGML_CPU_ALL_VARIANTS

richiejp
and
claude
committed
perf: CPU ablation profiler (PF_PROF) + optional Q8_0 expert requantizer

richiejp
and
claude
committed
docs: CPU performance analysis (SSE-trap root cause, CPU_ALL_VARIANTS fix)

richiejp
and
claude
committed
perf: flash attention (default) for CPU and Vulkan

richiejp
and
claude
committed
bench: PF_WINDOW knob + memory columns in pf-bench

richiejp
and
claude
committed
perf: prototype O(n*band) banded sliding-window attention

richiejp
and
claude
committed
perf: integrate banded sliding-window attention (PF_BANDED)

richiejp
and
claude
committed
perf: chunk the MoE FFN over tokens (PF_MOE_CHUNK) to lift the large-window cap

richiejp
and
claude
committed
docs: MoE chunking + single-window result

richiejp
and
claude
committed
perf: enable banded attention + MoE chunking by default (length-gated)

richiejp
and
claude
committed
docs: refresh README Bench with the HF-vs-ours comparison

richiejp
and
claude
committed
docs: add speedup column to Bench tables, trim prose

richiejp
and
claude
committed
ci: cap build parallelism (-j4) so the portable build doesn't OOM

richiejp
and
claude
committed