feat(gguf,dsv4): GGUF block-dequant + DSv4 kernel surface (WIP) - #9

Closed

TheTom wants to merge 18 commits into

thewafflehaus:devfrom

TheTom:tom/wip/gguf-dequant

Commits on May 30, 2026

feat(gguf): WIP block-dequant kernels — Q8_0 + Q2_K + IQ2_XXS (scaffold)
TheTom
committed
fix(gguf): switch to bare #[kernel] — legacy bench(...) registration rejects mixed-dtype param sets
TheTom
committed
fix(iq2_xxs): use canonical ksigns_iq2xs[128] table, not bit-parity reconstruction
TheTom
committed
feat(dsv4): sqrt(softplus) router scoring kernel
TheTom
committed
feat(dsv4): MXFP4 (OCP FP4 e2m1) block dequant kernel
TheTom
committed
feat(dsv4): block-FP8 e4m3 dequant (attention + router weight format)
TheTom
committed
docs(dsv4): forward-path kernel arc — verified shapes, gotchas, remaining work
TheTom
committed
fix(router): mark sqrtsoftplus outputs with mut for codegen Op::Store validation
TheTom
committed
feat(dsv4): CSA/HCA compressor pool — softmax-gated weighted sum + APE
TheTom
committed
feat(dsv4): HCA d512+sink + mHC pre/post + sqrtsoftplus codegen fix
TheTom
committed
feat(dsv4): Lightning Indexer per-position aggregate score kernel
TheTom
committed
feat(dsv4): Lightning Indexer top-k index selector (single-block bitonic)
TheTom
committed
feat(dsv4): CSA sparse-gather SDPA decode (head_dim=512)
TheTom
committed
chore: copyright headers to team form across ffai/
TheTom
committed
feat(dsv4): mHC dynamic-mix sinkhorn-split kernel + comment scrub
TheTom
committed
feat(dsv4): mHC collapse + expand replace static pre/post placeholder
TheTom
committed
feat(dsv4): partial RoPE — rotate the tail n_rot dims per head
TheTom
committed
feat: ffai_vector_add — generic elementwise out[i] = a[i] + b[i]
TheTom
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

feat(gguf,dsv4): GGUF block-dequant + DSv4 kernel surface (WIP) - #9

feat(gguf,dsv4): GGUF block-dequant + DSv4 kernel surface (WIP)#9
TheTom wants to merge 18 commits into
thewafflehaus:devfrom
TheTom:tom/wip/gguf-dequant

Commits on May 30, 2026

feat(gguf): WIP block-dequant kernels — Q8_0 + Q2_K + IQ2_XXS (scaffold)

fix(gguf): switch to bare #[kernel] — legacy bench(...) registration rejects mixed-dtype param sets

fix(iq2_xxs): use canonical ksigns_iq2xs[128] table, not bit-parity reconstruction

feat(dsv4): sqrt(softplus) router scoring kernel

feat(dsv4): MXFP4 (OCP FP4 e2m1) block dequant kernel

feat(dsv4): block-FP8 e4m3 dequant (attention + router weight format)

docs(dsv4): forward-path kernel arc — verified shapes, gotchas, remaining work

fix(router): mark sqrtsoftplus outputs with mut for codegen Op::Store validation

feat(dsv4): CSA/HCA compressor pool — softmax-gated weighted sum + APE

feat(dsv4): HCA d512+sink + mHC pre/post + sqrtsoftplus codegen fix

feat(dsv4): Lightning Indexer per-position aggregate score kernel

feat(dsv4): Lightning Indexer top-k index selector (single-block bitonic)

feat(dsv4): CSA sparse-gather SDPA decode (head_dim=512)

chore: copyright headers to team form across ffai/

feat(dsv4): mHC dynamic-mix sinkhorn-split kernel + comment scrub

feat(dsv4): mHC collapse + expand replace static pre/post placeholder

feat(dsv4): partial RoPE — rotate the tail n_rot dims per head

feat: ffai_vector_add — generic elementwise out[i] = a[i] + b[i]