[V100/SM70] gemma-4 (31B + MTP) support: fully-FA hybrid attention (sliding-window + head_dim-512)#59

Closed

rivetphilbot wants to merge 25 commits into

1CatAI:mainfrom

rivetphilbot:feat/gemma4-mtp

Commits on May 19, 2026

[SM70] Add V100 dense WNA16 TurboMind linear kernel
rivetphilbot
committed
[SM70] Admit V100 (CC 7.0) in CompressedTensorsWNA16
rivetphilbot
committed
[SM70] Wire compressed-tensors MoE decode buffers for V100
rivetphilbot
committed
[Qwen3] Keep router gate and split GDN projections unquantized under compressed-tensors
rivetphilbot
committed
[Qwen3.5] Skip tuple-shard split for non-output-dim CT params
rivetphilbot
committed
[SM70] Sync sm70_884_4.cu kernel registry to lmdeploy main (gs32)
rivetphilbot
committed

Commits on Jun 5, 2026

Commits on Jun 7, 2026