feat(tree-native): add tree-attention metadata + position_ids surface for tree training#3

Open

racky-scitix wants to merge 8 commits into

sirl-dev-tree-native

Commits on Apr 22, 2026

feat(tree-native): add Megatron PackedSeqParams.tree_metadata + TreeMetadata dataclass

MaoChouHJM
authored and
racky-scitix
committed
feat(tree-native): add Megatron tree-attention extensions

MaoChouHJM
authored and
racky-scitix
committed
refactor(tree-native): slim TETreeDotProductAttention to TE delegation

MaoChouHJM
authored and
racky-scitix
committed
refactor(tree-native): data-driven tree dispatch, eliminate spec-level tree classes

MaoChouHJM
authored and
racky-scitix
committed
refactor(tree-native): remove tree-specific code from TEDotProductAttention

MaoChouHJM
authored and
racky-scitix
committed
refactor(tree-native): generic position_ids on PackedSeqParams for RoPE

MaoChouHJM
authored and
racky-scitix
committed
refactor(tree-native): position_ids flows through model.forward, not PackedSeqParams

MaoChouHJM
authored and
racky-scitix
committed

Commits on Apr 23, 2026

refactor(tree-native): drop stray blank line in transformer_engine.py
racky-scitix
committed