Repository navigation

[WS1][CUDA][Qwen3-Next] MoE route/combine contract with a batch-invariant grouped expert GEMM (RFC #428 stack 5/8) - #503

Open

fusheng-ji wants to merge 59 commits into

RL-Align:test-qwennextfrom

fusheng-ji:feat/ws1-qwen3-next-moe-route-combine

Commits on Oct 4, 2026
Commits on Oct 6, 2026
- Merge test-qwennext into feat/ws1-c6-gdn-recurrent-golden
  fusheng-ji
  committed
  c883134
- Merge test-qwennext into Qwen3-Next C1 norm branch
  fusheng-ji
  committed
  81eae9e
- Merge updated C1 norm base into gated RMSNorm branch
  fusheng-ji
  committed
  8ce67e0
Commits on Oct 7, 2026
- fix(norm): reject stale RMSNorm bindings and enable GPU coverage
  fusheng-ji
  committed
  bec7e7a
Commits on Oct 8, 2026
Commits on Oct 9, 2026
Commits on Oct 10, 2026
- Merge test-qwennext refactor and migrate Qwen3-Next code and tests
  fusheng-ji
  committed
  7f70b60
- style(tools): mark the Qwen3-Next TP gate imports after sys.path setup as E402-exempt
  fusheng-ji
  committed
  1513b71
Commits on Oct 11, 2026
- fix(qwen3-next): type annotations for mypy in the MoE prior-art runner
  fusheng-ji
  committed
  37bc953