Repository navigation

feat(models): add nvidia/Qwen3.8-Flash-Next-FP4-Local checkpoint support - #708

Closed

chrisqianz wants to merge 14 commits into

FlashML-org:mainfrom

chrisqianz:qwen4_exp-fp4-local

Closed

feat(models): add nvidia/Qwen3.8-Flash-Next-FP4-Local checkpoint support#708
chrisqianz wants to merge 14 commits into
FlashML-org:mainfrom
chrisqianz:qwen4_exp-fp4-local

Commits on Oct 16, 2026
- feat(models): decode CSF slab scales for the FP4-Local NVFP4 experts
  chrisqian
  committed
  2fc2020
- feat(qwen4_exp): load the FP4-Local PLE table as packed e2m1
  chrisqian
  committed
  2db65bd
- test(qwen4_exp): import save_file locally in the NVFP4 disk rejection test
  chrisqian
  committed
  c624c02
- fix(models): int32 CSF exception math and accept plain kinds in pair mode
  chrisqian
  committed
  b708c45
- test(qwen4_exp): build CSF test records as Python ints and compare slab shapes
  chrisqian
  committed
  c5e0120
- test(qwen4_exp): store CSF U32 exception records as signed int32 bit patterns
  chrisqian
  committed
  2d19499
- test(qwen4_exp): index the e2m1 LUT with int64 nibbles
  chrisqian
  committed
  54b1dfb
- fix(qwen4_exp): widen PLE nibbles to int64 before the e2m1 LUT lookup
  chrisqian
  committed
  b3135b5
- fix(models): find index-only checkpoints that keep shards in a subdirectory
  chrisqian
  committed
  3ff0cf0
- fix(qwen4_exp): accept the transformers 5.16 full-attention layer alias
  chrisqian
  committed
  545a7fe
- feat(qwen4_exp): serve the FP4-Local dense projections as MXFP8 / W4A16
  chrisqian
  committed
  2ca3fce
- fix(qwen4_exp): fall back to the pinned PLE backend for packed nvfp4 tables
  chrisqian
  committed
  59953e3
- fix(models): read the U32 CSF exceptions stream in the parallel expert loader
  chrisqian
  committed
  f4a9df2
- fix(kernels): decode the e2m1 sign without a non-boolean tl.where
  chrisqian
  committed
  9945046

Commits

Commits on Oct 16, 2026