Repository navigation

[2/10] GDN/KDA state QAT with native forwards and INT8/FP8 state - #2519

Draft

kaix-nv wants to merge 27 commits into

mainfrom

kaix/linear-attention-decode-first

Draft

[2/10] GDN/KDA state QAT with native forwards and INT8/FP8 state#2519
kaix-nv wants to merge 27 commits into
mainfrom
kaix/linear-attention-decode-first

Commits on Oct 10, 2026
- Add GDN/KDA Torch decode QAT with INT8 recurrent state
  kaix-nv
  committed
  e38b2f9
- Use Megatron training adapters for GDN and KDA
  kaix-nv
  committed
  c396e5c
- Use Megatron Bridge workflows for linear attention QAT and QAD
  kaix-nv
  committed
  4e583c1
- Enable distributed Megatron Bridge state QAT and QAD
  kaix-nv
  committed
  82c9528
- Keep linear attention training tests minimal
  kaix-nv
  committed
  739c137
- Make linear attention state QDQ format dispatch explicit
  kaix-nv
  committed
  8142af2
- Support TensorQuantizer blockwise INT8 linear attention state QDQ
  kaix-nv
  committed
  17ccf0e
- Simplify linear attention prefill and decode plumbing
  kaix-nv
  committed
  9cf28af
- Remove redundant linear attention quantization scaffolding
  kaix-nv
  committed
  ee68955
- Separate linear attention training example from core support
  kaix-nv
  committed
  09c085a
- Align GDN and KDA state QAT with native vLLM arithmetic
  kaix-nv
  committed
  f67e37f
- Align state QAT with native Hadamard and ReplaySSM arithmetic
  kaix-nv
  committed
  4e8fdf6
- Simplify serving-aligned linear-attention state QAT
  kaix-nv
  committed
  309010a
- Simplify state QAT interfaces and fix BF16 replay
  kaix-nv
  committed
  0d03620
- Refactor linear attention policy and checkpoint lifecycle
  kaix-nv
  committed
  406343d
- Support native vLLM kernel revisions in linear attention QAT
  kaix-nv
  committed
  68759da
- Defer native vLLM inspection for documentation imports
  kaix-nv
  committed
  5cbfe54
- Fix linear-attention state QAT phase handling and serving profiles
  kaix-nv
  committed
  b3c7d67
- Fix FLA dependency setup for vLLM GPU tests
  kaix-nv
  committed
  dd01639
- Fix linear attention CI dependency and Megatron compatibility
  kaix-nv
  committed
  140735b
- Guard linear attention GPU tests on older Triton and reduce KDA compilation
  kaix-nv
  committed
  fee00a1
Commits on Oct 12, 2026
- Simplify linear attention state QAT and serving compatibility
  kaix-nv
  committed
  52b75c6
Commits on Oct 13, 2026
- Fix linear attention serving contracts and GPU tests
  kaix-nv
  committed
  3bbaf86
Commits on Oct 14, 2026
- Move ReplaySSM state QAT into a separate follow-up
  kaix-nv
  committed
  5c9797a
- Use one native state-QAT forward and reuse CI kernel shapes
  kaix-nv
  committed
  bcb2263
- Expose FP8 state block sizes through TensorQuantizer
  kaix-nv
  committed
  8cbec76
- Reuse the integration scale for prefix backward compilation
  kaix-nv
  committed
  662fd69

Commits

Commits on Oct 10, 2026

Commits on Oct 12, 2026

Commits on Oct 13, 2026

Commits on Oct 14, 2026