[WS2][PR5][Attention] Add distributed CP drift benchmark artifacts - #285

Merged

Flink-ddd merged 26 commits into

testfrom

codex/ws2-pr5-cp-attention-drift

Aug 24, 2026

Commits on Aug 7, 2026

feat(attention): add CP backward drift validation
inaniloquentee
committed
82aa599

Commits on Aug 8, 2026

feat(attention): add CP drift benchmark artifacts
inaniloquentee
committed
561dce6

Commits on Aug 12, 2026

Merge PR3 CP reference dependency
inaniloquentee
committed
368d1bd
feat(attention): add P2P NCCL CP drift benchmark
inaniloquentee
committed
a013e1c

Commits on Aug 13, 2026

fix(attention): validate CP reference numeric inputs
inaniloquentee
committed
79a5c3a
fix(attention): bind backward gradient dtype and device
inaniloquentee
committed
578c6b2
fix(attention): enforce FP32 CP merge state
inaniloquentee
committed
504a95d
test(attention): add strict issue 235 GPU acceptance gate
inaniloquentee
committed
4d4f331
test(attention): verify P2P final-write evidence
inaniloquentee
committed
37c0b2d
test(attention): bind P2P manifest ownership
inaniloquentee
committed
2627886
style(attention): satisfy full PR lint
inaniloquentee
committed
14e60df

Commits on Aug 16, 2026

test(attention): require Q AG in GPU acceptance
inaniloquentee
committed
0ee49c3
fix(ws2): use shared BF16 logprob tolerance
Codex H100 Integration
committed
3d5e395
fix(ws2): preserve structured unavailable lanes
Codex H100 Integration
committed
0f59ace
test(ws2): compare native TE KV ring diagnostically
Codex H100 Integration
committed
20d980a
fix(ci): satisfy PR5 lint and docs checks
Codex H100 Integration
committed
099798c
fix(types): narrow split-k boundaries
Codex H100 Integration
committed
6dba1e8
fix(types): validate acceptance evidence explicitly
Codex H100 Integration
committed
5939646
fix(ws2): keep allreduce outside attention gate
Codex H100 Integration
committed
6b79059

Commits on Aug 17, 2026

test(ws2): gate p2p attention on four-rank protocol
Codex H100 Integration
committed
7c45a6c
fix(ws2): separate FlashInfer drift thresholds
Codex H100 Integration
committed
0e8e0aa
test(attention): run full deterministic communication matrix
inaniloquentee
committed
2a0a600

Commits on Aug 18, 2026

Commits on Aug 24, 2026

Merge origin/test into PR #285
inaniloquentee
committed
3b91a4b
Merge branch 'test' into codex/ws2-pr5-cp-attention-drift
inaniloquentee
authored
15f65c6