feat(kernel): deepseek4 hc glue ops, cb caps, faster iq gather decode - #31

Merged

asher merged 15 commits into

mainfrom

moe-gather-variants

Aug 8, 2026

Commits on Aug 8, 2026

experiment(gather): x-staged ext GLU gather, opt-in only (slower on M5)
asher
committed
78a8a1f
perf(gather): fold iq block scale once per chunk dot, byte-index iq2_xxs grid
asher
committed
b11cb25
perf(gather): hoisted-scale chunk dequant for the remaining iq codecs
asher
committed
728e8e6
perf(matmul): hoisted-scale iq chunk dequant in the ext mat-vec
asher
committed
eecd258
feat(kernel): deepseek4 hyper-connection M=1 glue ops
asher
committed
ec4dfa7
test(hc-glue): numpy references and fused self-consistency
asher
committed
24c138e
feat(runtime): command-buffer cap get/set for per-phase serving
asher
committed
1ad38e0
docs: changelog for iq deq family, hc glue ops, cb caps
asher
committed
7f0fe0e
feat(moe): slot-parallel mix_ns and row-paired GLU gather variants
asher
committed
49c579e
refactor(iq): vectorize grid unpack in ext deq_chunk16s
asher
committed
0709c40
docs: changelog for gather variant kernels
asher
committed
233fca2
fix(moe): default row-paired GLU gather off; isolation wins not serving-real
asher
committed
38b3364
feat(moe): silu_limit epilogue on the shared-expert GLU ride-along
asher
committed
6053bca
revert(moe): drop the row-paired GLU gather and the shexp silu_limit act
asher
committed
0931b4c
strip: remove x-staged MoE GLU gather variant (_xs)
asher
committed
05d3ab6