feat(ffn): add deterministic distributed Triton FFN for ROCm - #325

Merged

Flink-ddd merged 78 commits into

RL-Align:testfrom

frank-2077:feat/rocm-strict-ffn

Sep 3, 2026

Commits on Aug 19, 2026

Merge main into ROCm strict Attention
root
committed
a9b2746
feat(attention): add strict ROCm Attention and RCCL path
root
committed
41bac5d
test(attention): make ROCm acceptance imports explicit
root
committed
21486d9
style-rocm-attention-lint
root
committed
674d7ba
test-unify-cuda-rocm-attention
root
committed
0934ca4
refactor(attention): consolidate ROCm implementation
root
committed
4c744b3
feat(attention): default strict runtime to decoupled ring schedule
Codex H100 Validation
committed
22d3780

Commits on Aug 20, 2026

fix(attention): gate ROCm strict provenance
inaniloquentee
committed
aac0fb0
feat(attention): use strict vendor cores by platform
inaniloquentee
committed
d146d43
test(attention): accept strict ROCm RoPE provenance
inaniloquentee
committed
4418d18
test(attention): validate reported platform provenance
inaniloquentee
committed
fe39824
feat(gemm): align BF16 K reduction with TP tree
frank-2077
committed
6424b88
feat(distributed): add ROCm transport-only deterministic collectives
frank-2077
committed
6c02774
feat(ffn): add deterministic Qwen3 FFN assembly
frank-2077
committed
6ad03ca
test(ffn): cover staged BF16 FFN contract
frank-2077
committed
f7dc7a9
test(ffn): cover TP CP and SP topologies
frank-2077
committed
49fa1c5
test(attention): run strict references by logical row
inaniloquentee
committed
f6ff094
test(attention): report executed vendor backend
inaniloquentee
committed
9ff0e6d
fix(attention): narrow injected AITER callables
inaniloquentee
committed
1bc4177

Commits on Aug 21, 2026

test(ffn): allow ROCm RCCL teardown to complete
frank-2077
committed
7d04efe
perf: benchmark deterministic FFN on ROCm
frank-2077
committed
2f3f61d

Commits on Aug 23, 2026

feat: add strict distributed Triton FFN for ROCm
frank-2077
committed
9dc9a68
docs: report MI300X Triton FFN performance
frank-2077
committed
809789c
refactor(ffn): make ROCm path Triton-native
frank-2077
committed
5222b68
docs(ffn): publish Triton ROCm performance report
frank-2077
committed
5fe86f4
docs(bench): label shapes axes and accuracy references
frank-2077
committed
43a31ae
docs(bench): refresh shape and topology figures
frank-2077
committed
6e61ef0
bench(rocm): compare Triton FFN with official TP1
frank-2077
committed
08f47d9
docs(bench): report TP1 FFN speed and exactness
frank-2077
committed
981c7ab
docs(bench): simplify ROCm result figures
frank-2077
committed
e64abab

Commits on Aug 24, 2026

Merge branch 'test' into codex/ws2-rocm-strict-attention

zhangj1an
and
claude
committed
13897a7
docs(bench): add CUDA GPU and CPU comparison
frank-2077
committed
bf36f05
feat(attention): route strict ROCm attention to Vime via contract dispatch

zhangj1an
and
claude
committed
3247e3e
feat(attention): bind TP/CP degree and add multi-rank strict validation

zhangj1an
and
claude
committed
3446042
Merge branch 'test' into codex/ws2-rocm-strict-attention

zhangj1an
and
claude
committed
97d97ee

Commits on Aug 26, 2026

perf(ffn): eliminate deterministic layout copies
frank-2077
committed
7380bc8
perf(ffn): optimize deterministic gfx942 leaf
frank-2077
committed
6e0c220
Merge branch 'test' into codex/ws2-rocm-strict-attention

zhangj1an
and
claude
committed
2162c19

Commits on Aug 28, 2026

Commits on Aug 29, 2026

Merge remote-tracking branch 'origin/test' into codex/ws2-rocm-strict-attention
zhangj1an
committed
0e7b442
Merge remote-tracking branch 'origin/feat/rocm-deterministic-collectives' into codex/ws2-rocm-strict-attention
zhangj1an
committed
d5041a2
fix(build): repair ops.cpp preprocessor guards mangled by the merge

zhangj1an
and
claude
committed
e8e79e7
feat(kernels): add bitwise Triton deterministic Attention core

zhangj1an
and
claude
committed
a8deab9
bench(attention): add WS2 strict ROCm Attention report on MI300X

zhangj1an
and
claude
committed
cc76c48
bench(attention): measure the per-KV-group schedule and correct the ~3x estimate

zhangj1an
and
claude
committed
0f8a4f9
bench(attention): drop the stale pre-per-KV-group MI300X performance artifacts

zhangj1an
and
claude
committed
e24b594
bench(attention): support CPU and CUDA hosts, add the exactness heatmap

zhangj1an
and
claude
committed
f4602c0
bench(attention): migrate result keys to the reference-native rename

zhangj1an
and
claude
committed
fd2b329
bench(attention): add the host run and fold the schedule cost into the script

zhangj1an
and
claude
committed
15e123f
bench(attention): skip a cell instead of waiting on it forever

zhangj1an
and
claude
committed
279ee20
chore: trigger dco app
Flink-ddd
committed
f0513d2
bench(attention): add the CPU host sweep and a hard per-cell time budget

zhangj1an
and
claude
committed
1a2e4fc
perf(distributed): optimize deterministic ROCm collectives
maxiaosong1124
committed
a55233f
perf(distributed): adopt optimized ROCm collectives
frank-2077
committed
caef501
bench(rocm): update distributed FFN results
frank-2077
committed
19e0ee6

Commits on Aug 30, 2026

Merge pull request #361 from RL-Align/test
Flink-ddd
authored
0831aa8
Merge branch 'main' into codex/ws2-rocm-strict-attention
zhangj1an
committed
2440a1b
feat(attention): add the strict ROCm runtime and open CP on the dispatch path

zhangj1an
and
claude
committed
b5c948a
Optimize deterministic rollout tensor-parallel all-reduce
inaniloquentee
committed
c57a8ae
Optimize deterministic rollout tensor-parallel all-reduce (#365)
inaniloquentee
authored
9d5732b
refactor(attention): resolve the ROCm CP transport through the shared collective

zhangj1an
and
claude
committed
99eebae
merge: resolve test conflicts for ROCm collectives
Flink-ddd
committed
553e9a5
style(distributed): satisfy collective formatting
Flink-ddd
committed
285dda1
Merge latest PR #356 into PR #357
maxiaosong1124
committed
66c6276
Merge PR #357 (perf/rocm-deterministic-collectives-ipc) into codex/ws2-rocm-strict-attention

zhangj1an
and
claude
committed
52a41fb
feat(attention): validate the AITER schema the strict ROCm core calls into

zhangj1an
and
claude
committed
ca5313f
feat(attention): add the strict ROCm decode path over a paged KV cache

zhangj1an
and
claude
committed
4f8704b
Enable exact-batch CUDA graphs for strict rollout
inaniloquentee
committed
8e32313

Commits on Aug 31, 2026