Repository navigation

[9/10] GDN/KDA prefill operand and arithmetic QAT - #2503

Draft

kaix-nv wants to merge 8 commits into

kaix/linear-attention-replayssm-servingfrom

kaix/linear-attention-qat-m2

Draft

[9/10] GDN/KDA prefill operand and arithmetic QAT#2503
kaix-nv wants to merge 8 commits into
kaix/linear-attention-replayssm-servingfrom
kaix/linear-attention-qat-m2

Commits on Sep 22, 2026
- Add GDN state and WY activation QAT emulation
  kaix-nv
  committed
  07d39f3
Commits on Sep 28, 2026
- Qualify GDN QAT on Hopper and Megatron training
  kaix-nv
  committed
  bfbb13e
Commits on Sep 29, 2026
- Move GDN test compilation into setup fixtures
  kaix-nv
  committed
  95766de
- Clarify dynamic per-tile FP8 state quantization
  kaix-nv
  committed
  3896181
- Add GDN/KDA Torch decode QAT with INT8 recurrent state
  kaix-nv
  committed
  d03c732
- Add fused Triton GDN/KDA decode QAT
  kaix-nv
  committed
  756c330
- Add vLLM GDN/KDA state-only fake quantization
  kaix-nv
  committed
  bae8fa2
- Add GDN and KDA prefill QAT after decode infrastructure
  kaix-nv
  committed
  bd2be2b