Repository navigation

feat(qwen4_exp): image input (vision tower + mRoPE) for Qwen3.8-Flash-Next - #386

Closed

gdevenyi wants to merge 10 commits into

FlashML-org:mainfrom

gdevenyi:feat/qwen4-exp-vision

Closed

feat(qwen4_exp): image input (vision tower + mRoPE) for Qwen3.8-Flash-Next#386
gdevenyi wants to merge 10 commits into
FlashML-org:mainfrom
gdevenyi:feat/qwen4-exp-vision

Commits on Sep 10, 2026
- feat(qwen4_exp): tensor parallelism for Qwen3.8-Flash-Next (offload backend)
  
  gdevenyi
  and
  claude
  committed
  c3d0f97
- fix(qwen4_exp): load the HF config for sharding only when TP > 1
  
  gdevenyi
  and
  claude
  committed
  96e190c
- test(qwen4_exp): TP shard test against the new expert-piece stream
  gdevenyi
  committed
  a99433c
- feat(qwen4_exp): image input (vision tower + mRoPE) for Qwen3.8-Flash-Next
  
  gdevenyi
  and
  claude
  committed
  01e10be
- feat(qwen4_exp): chunked prefill for image prompts
  
  gdevenyi
  and
  claude
  committed
  13bdd2e
- feat(qwen4_exp): key image prompts in the prefix cache by image content hash
  
  gdevenyi
  and
  claude
  committed
  b79d7b9
- fix(qwen4_exp): forward prefix through Qwen4ExpMoE, and test the CPU build
  
  gdevenyi
  and
  claude
  committed
  62a9a8c
- Merge branch 'rb/tp' into rb/vision
  gdevenyi
  committed
  0874679
- feat(quant): let the triton NVFP4 expert kernel serve a tensor-parallel shard
  
  gdevenyi
  and
  claude
  committed
  d2ad39b
- Merge branch 'rb/tp' into rb/vision
  gdevenyi
  committed
  db6571a