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

Open

frank-2077 wants to merge 32 commits into

RL-Align:codex/ws2-rocm-strict-attentionfrom

frank-2077:feat/rocm-strict-ffn

Commits on Aug 20, 2026

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

docs(bench): add CUDA GPU and CPU comparison
frank-2077
committed
bf36f05

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

Commits on Aug 29, 2026

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

Commits on Aug 31, 2026