feat(dpmodel): graph-native se_atten attention (NeighborGraph PR-D) - #5715

Merged

wanghan-iapcm merged 16 commits into

deepmodeling:masterfrom

wanghan-iapcm:feat-graph-attn-prD

Jul 6, 2026

Commits on Jul 4, 2026

feat(dpmodel): segment_max + numerically-stable mask-aware segment_softmax
Han Wang
committed
962399b
feat(dpmodel): center_edge_pairs primitive (shared by attention/angles)
Han Wang
committed
d3132d5
feat(dpmodel): graph-native se_atten transformer attention (attn_layer > 0)
Han Wang
committed
05fbd8f
test(pt_expt): graph attention make_fx (merge gate) + model force/virial parity
Han Wang
committed
61285c6
test: binding-sel audit for graph-default attention models
Han Wang
committed
298518e
test(pt_expt): pin smooth off in neighbor-list dpa1 fixture (route parity)
Han Wang
committed
ccefbbc
[pre-commit.ci] auto fixes from pre-commit.com hooks

pre-commit-ci[bot]
authored and
Han Wang
committed
1c28aa5
feat(dpmodel): make compact center_edge_pairs traceable via unbacked SymInts
Han Wang
committed
9ab2f38
feat(pt_expt): graph-form .pt2 export for dpa1 attention (attn_layer > 0)
Han Wang
committed
397c405
fix(dpmodel): cast edge_vec to descriptor precision in dpa1 call_graph
Han Wang
committed
67d140d
fix(dpmodel): shift masked values in segment_softmax to prevent segment-wide NaN
Han Wang
committed
bf09b85
docs(pt_expt): document sel-independent smooth graph attention; refresh eligibility wording
Han Wang
committed
7bf56b6
fix(pt_expt): fail fast on torch < 2.6 for graph attention tracing
Han Wang
committed
d49b9cb
docs(pt_expt): numpydoc sections for check_graph_trace_torch_version
Han Wang
committed
7c65935
test+docs: pin smooth-attention graph-vs-dense divergence; document the pt_expt graph default and legacy escape hatch
Han Wang
committed
6fc45bd

Commits on Jul 5, 2026

test(pt_expt): make CUDA graph-path tests device-robust
Han Wang
committed
84aaef5

Navigation Menu

Uh oh!

feat(dpmodel): graph-native se_atten attention (NeighborGraph PR-D) - #5715

feat(dpmodel): graph-native se_atten attention (NeighborGraph PR-D)#5715
wanghan-iapcm merged 16 commits into
deepmodeling:masterfrom
wanghan-iapcm:feat-graph-attn-prD

Commits on Jul 4, 2026

feat(dpmodel): segment_max + numerically-stable mask-aware segment_softmax

feat(dpmodel): center_edge_pairs primitive (shared by attention/angles)

feat(dpmodel): graph-native se_atten transformer attention (attn_layer > 0)

test(pt_expt): graph attention make_fx (merge gate) + model force/virial parity

test: binding-sel audit for graph-default attention models

test(pt_expt): pin smooth off in neighbor-list dpa1 fixture (route parity)

[pre-commit.ci] auto fixes from pre-commit.com hooks

feat(dpmodel): make compact center_edge_pairs traceable via unbacked SymInts

feat(pt_expt): graph-form .pt2 export for dpa1 attention (attn_layer > 0)

fix(dpmodel): cast edge_vec to descriptor precision in dpa1 call_graph

fix(dpmodel): shift masked values in segment_softmax to prevent segment-wide NaN

docs(pt_expt): document sel-independent smooth graph attention; refresh eligibility wording

fix(pt_expt): fail fast on torch < 2.6 for graph attention tracing

docs(pt_expt): numpydoc sections for check_graph_trace_torch_version

test+docs: pin smooth-attention graph-vs-dense divergence; document the pt_expt graph default and legacy escape hatch

Commits on Jul 5, 2026

test(pt_expt): make CUDA graph-path tests device-robust