backport: feature-detect DSA experimental-attention spec for GLM-5/5.1/5.2#13

Open

yushengsu-thu wants to merge 11 commits into

Commits on Jun 19, 2026

backport: feature-detect DSA experimental-attention spec for GLM-5/5.1
yushengsu-thu
committed
backport: parse GLM-5.2 MLA rope dims (rotary_base + qk_pos_emb_head_dim)
yushengsu-thu
committed
backport: GLM-5.2 DSA cross-layer index sharing (CrossLayerDSAttention)
yushengsu-thu
committed

Commits on Jun 20, 2026

backport: GLM-5.2 cross-layer — assert PP stage starts on a computing layer (build time)
yushengsu-thu
committed
backport: GLM-5.2 cross-layer — reject bshd + activation recompute (holder not recompute-safe)
yushengsu-thu
committed

Commits on Jun 21, 2026

backport: GLM DSA indexer — rope-half swap on wq_b/wk/k_norm (HF<->Megatron)
yushengsu-thu
committed

Commits on Jun 22, 2026