Repository navigation

Flash Next TP=2: exact output-column splits, 21% faster plain decode - #1784

Open

laudney wants to merge 15 commits into

ashhart:mainfrom

laudney:pr-fn-tp2-split

Open

Flash Next TP=2: exact output-column splits, 21% faster plain decode#1784
laudney wants to merge 15 commits into
ashhart:mainfrom
laudney:pr-fn-tp2-split

Commits on Nov 2, 2026
- fix(flashnext): keep the ranked vtable static
  laudney
  committed
  ac2e81c
- feat(cuda): serve --tp to the CUDA families with ranks
  laudney
  committed
  f60cc18
- feat(flashnext): ranks agree on a computed checkpoint digest
  laudney
  committed
  0a97be7
- feat(flashnext): accept BF16 dense and shared layers in ModelOpt
  laudney
  committed
  c24e2f4
- feat(flashnext): FP8 rows in the n-gram table
  laudney
  committed
  c6e4247
- feat(flashnext): FP8 block experts in the MTP head
  laudney
  committed
  399d535
- docs(flashnext): serve NVIDIA's NVFP4 checkpoint at TP=2
  laudney
  committed
  be7eba0
- refactor(cuda): keep --tp's refusals and test out of cuda.zig
  laudney
  committed
  fd3c61d
- test(contract): gate Flash Next NVFP4's BF16 and FP8 expert kernels
  laudney
  committed
  f14b20e
- fix(flashnext): keep one GPU's expert sums at --tp 2
  laudney
  committed
  57647e4
- docs(changelog): note --tp 2 replies equal one GPU's for Flash Next
  laudney
  committed
  87dd668
- style(contract): shorten comments for the lean rule
  laudney
  committed
  6c0a0a5
- fix(flashnext): word the --tp load refusals and credit the PR
  laudney
  committed
  db7e9f5
- feat(flashnext): split dense output columns with exact gathers
  laudney
  committed
  1bfd6cc
- test(cuda): compare GDN heads against a frozen baseline
  laudney
  committed
  1077d5f