Repository navigation

feat: FlashInfer FA2 paged decode backend for Ada (SM89) - #162

Open

sysuTobo wants to merge 20 commits into

InferMatrixLab:mainfrom

sysuTobo:feat/flashinfer-fa2-sm89

Open

feat: FlashInfer FA2 paged decode backend for Ada (SM89)#162
sysuTobo wants to merge 20 commits into
InferMatrixLab:mainfrom
sysuTobo:feat/flashinfer-fa2-sm89

Commits on Nov 1, 2026
- kernels: add FlashInfer FA2 paged decode for Ada
  sysuTobo
  committed
  69b7b1c
- test: cover FlashInfer FA2 paged decode on SM89
  sysuTobo
  committed
  8b512b1
- bench: select the native attention backend for GSM8K
  sysuTobo
  committed
  31d4a36
- docs: document the FlashInfer FA2 Ada backend
  sysuTobo
  committed
  66e0734
- docs: record L20 GSM8K-87 results for FlashInfer FA2
  sysuTobo
  committed
  31c9f82
- docs: quantify L20 FA2 tie flips and matched throughput
  sysuTobo
  committed
  03b3089
- feat: enable CUDA Graphs and async steps in FlashInfer FA2
  sysuTobo
  committed
  8a71def
- bench: expose --cuda-graphs in the GSM8K harness
  sysuTobo
  committed
  23e28f8
- docs: record L20 FA2 CUDA Graph and async validation
  sysuTobo
  committed
  b7411f9
- docs: add Triton CUDA Graph numbers to the L20 table
  sysuTobo
  committed
  b41a756
Commits on Nov 2, 2026
- fix: keep FlashInfer FA2 CSR construction device-only
  sysuTobo
  committed
  aadc0a8
- test: cover FA2 CSR bounds, row checks and synchronization
  sysuTobo
  committed
  315a16f
- bench: validate --cuda-graphs before creating the output
  sysuTobo
  committed
  a355f9e
- docs: correct FA2 graph claims and refresh post-fix measurements
  sysuTobo
  committed
  86c09b7
- feat: make FlashInfer FA2 plans deterministic and bounded
  sysuTobo
  committed
  4afcb44
- feat: expose the resolved FA2 kernel through selection info
  sysuTobo
  committed
  55add98
- test: cover FA2 plan bounds, split invariance and wide-table graphs
  sysuTobo
  committed
  7123886
- docs: document deterministic FA2 plans and refresh measurements
  sysuTobo
  committed
  d522a0d
- perf: copy FA2 CSR metadata once per traversal
  sysuTobo
  committed
  9e170a4
- bench: fail fast for non-native --cuda-graphs
  sysuTobo
  committed
  765fe41

Commits

Commits on Nov 1, 2026

Commits on Nov 2, 2026