perf(qwen35): batch eager decode rows under TP (split 2/4 of #946) - #1004

Open

Ma1oneZhang wants to merge 7 commits into

pegainfer-project:mainfrom

Ma1oneZhang:feat/qwen35-tp-batched-decode

Commits on Aug 30, 2026

feat(qwen35): shard linear-attention/GDR state per TP rank
Ma1oneZhang
committed
d9cd1e0
fix(qwen35): allow the f32 shard loader's pointer alignment cast
Ma1oneZhang
committed
ba12041
style(qwen35): swap the negated decode plan branch and parse fixture f32s via as_chunks
Ma1oneZhang
committed
fcdeb5a

Commits on Sep 3, 2026

docs(qwen35): record landed P2B GDR sharding in the TP implementation record
Ma1oneZhang
committed
44e6dc0

Commits on Sep 4, 2026

fix(qwen35): validate dtype and alignment in the 1D F32 weight loaders
Ma1oneZhang
committed
e6629f9

Commits on Sep 5, 2026