[WS2][PR7][Attention] Add FlashInfer RoPE-fused paged attention scaffold - #279

Merged

Flink-ddd merged 33 commits into

testfrom

codex/ws2-pr7-flashinfer-rope-fused-submit

Aug 24, 2026

Commits on Aug 5, 2026

feat(attention): add PR7 FlashInfer fused backend scaffold
inaniloquentee
committed
b69859e

Commits on Aug 13, 2026

fix(attention): bind decode replay to cache execution identity
inaniloquentee
committed
6e8848d
fix(attention): validate CP reference numeric inputs
inaniloquentee
committed
23b518f
fix(attention): bind backward gradient dtype and device
inaniloquentee
committed
de673d2
fix(attention): enforce FP32 CP merge state
inaniloquentee
committed
5f85361
fix(attention): fail closed on unverified FlashInfer execution
inaniloquentee
committed
b83141e
test(attention): add two-rank P2P NCCL reference check
inaniloquentee
committed
a2b2cbd
style(attention): satisfy full PR lint
inaniloquentee
committed
75365a3

Commits on Aug 16, 2026

feat(attention): execute CP query AG and deterministic fallback
inaniloquentee
committed
cee6da5
test(attention): accept CUDA unavailable fail-closed path
inaniloquentee
committed
0ed7b56
fix(ws2): record missing FlashInfer as unavailable
Codex H100 Integration
committed
3a65f4c
fix(ws2): run strict FlashInfer provenance on native FA2
Codex H100 Integration
committed
6e1b065
style(attention): satisfy PR7 pre-commit hooks
Codex H100 Integration
committed
0eecbfd
fix(types): narrow attention plan values
Codex H100 Integration
committed
d35250a

Commits on Aug 17, 2026

test(attention): cover TP2 CP2 P2P reference
Codex H100 Integration
committed
0cc82b6
feat(attention): route strict CP through shared deterministic core
inaniloquentee
committed
8309b77

Commits on Aug 18, 2026

feat(attention): reuse WS1 PR315 deterministic operators
Codex H100 Integration
committed
7bea721
fix(attention): route strict core through canonical schedule
inaniloquentee
committed
aecdf4d
feat(attention): expose canonical strict schedule
inaniloquentee
committed
8e00bad
Merge remote-tracking branch 'upstream/test' into codex/review-pr279-live
inaniloquentee
committed
cc47649
fix(attention): enforce strict schedule provenance
inaniloquentee
committed
b7f463b
fix(attention): compare strict PR7 path to shared CUDA core

Codex H100 Validation
authored and
inaniloquentee
committed
99d2eca

Commits on Aug 19, 2026

feat(attention): default strict runtime to decoupled ring schedule
Codex H100 Validation
committed
802f168
style(attention): format PR7 validation script
inaniloquentee
committed
4784c9b

Commits on Aug 20, 2026

feat(attention): use strict FA4 production core
inaniloquentee
committed
044f5b1

Commits on Aug 24, 2026

Merge origin/test into PR #279
inaniloquentee
committed
3e04a63
Merge test into PR #279
inaniloquentee
committed
e07f8b1
Merge latest test into PR #279
inaniloquentee
committed
28582d0
Merge latest test into PR #279
inaniloquentee
committed
4a60fe5