perf(pt2): optimize .pt2 C++ inference path - #5407

Merged

njzjz merged 57 commits into

deepmodeling:masterfrom

wanghan-iapcm:perf-pt-expt-pt2-cpp

Apr 26, 2026

Commits on Apr 15, 2026

feat(pt_expt): multi-task training support
Han Wang
committed
512eeb6

Commits on Apr 16, 2026

fix(dpmodel): wrap fparam/aparam reshape with descriptive ValueError
Han Wang
committed
9f4d232
fix: address CodeQL findings in PR #5397
Han Wang
committed
9f1f1d8
fix(pt_expt): access unwrapped module in _compile_model for DDP compat
Han Wang
committed
f3f5474
test(pt_expt): add DDP + torch.compile training tests
Han Wang
committed
665b85a
feat(pt_expt): use inductor+dynamic for training compile
Han Wang
committed
aabb710
test(pt_expt): port silut activation + repformers accessors from #5393
Han Wang
committed
f774cd2
test(pt_expt): assert virial in compile correctness tests
Han Wang
committed
0b5468e
test(pt_expt): port silu compile and varying-natoms tests from #5393
Han Wang
committed
9bf006b
test(pt_expt): compare compiled vs uncompiled with varying natoms
Han Wang
committed
7722f52
test(pt_expt): cover DPA2/DPA3 in varying-natoms compile correctness
Han Wang
committed
be14ac2
test(pt_expt): exercise DPA2 three-body branch in compile correctness
Han Wang
committed
4c0b8ec

Commits on Apr 17, 2026

fix(dpmodel): restore nf in reshapes to fix zero-atom and add silu_backward decomp
Han Wang
committed
80c714c
fix: address CodeQL findings in PR #5397
Han Wang
committed
6158d9c
fix(pt): wrap fparam/aparam reshape with descriptive ValueError
Han Wang
committed
c2efbf1

Commits on Apr 18, 2026

feat(pt_expt): reject DPA1/se_atten_v2 with attention at compile time
Han Wang
committed
1e694a3
fix(pt_expt): remove false DPA1 attention compile guard
Han Wang
committed
6d39ddf
refactor(dpmodel): remove unused get_numb_attn_layer API
Han Wang
committed
23eb6dd
fix(test): use real path for PT water data, remove unused API
Han Wang
committed
bacd312
fix(pt_expt): rebuild FX graph after detach node removal to avoid segfaults
Han Wang
committed
f834202
fix(pt_expt): tune inductor options for compile training
Han Wang
committed
447a572
fix(pt_expt): disable DDPOptimizer to prevent compiled graph splitting
Han Wang
committed
fb25ccb
fix(test): add .cpu() before .numpy() for GPU-compatible activation tests
Han Wang
committed
479900d
fix(pt_expt): revert inductor options that cause numerical divergence
Han Wang
committed
b67a181
fix(test): make DDP tests device-adaptive instead of hardcoding CPU
Han Wang
committed
7ce7352
fix(test): correct freeze test docstrings to match dpa3 guard
Han Wang
committed
975db17
fix(pt_expt): move optimize_ddp into _compile_model, resolve test symlinks
Han Wang
committed
64dc703
fix(test): backup/restore fparam.npy in TestFparam instead of deleting
Han Wang
committed
28fbcac
fix(test): skip DDP tests when NCCL is selected with fewer than 2 GPUs
Han Wang
committed
fbb361a

Commits on Apr 20, 2026

perf(pt2): optimize .pt2 C++ inference path
Han Wang
committed
7739fad
Merge upstream/master into perf-pt-expt-pt2-cpp
Han Wang
committed
19272c2
feat(pt2): make nlist nnei dimension dynamic in .pt2 export
Han Wang
committed
b7509db
fix(pt2): pad nlist in Python eval path for dynamic nnei
Han Wang
committed
eec2528
fix(pt2): move atomic virial check before run_model and reject unsupported --atomic-virial
Han Wang
committed
217a587
fix(pt2): move nlist padding inside traced fn and strip shape assertions
Han Wang
committed
8a9fe63

Commits on Apr 21, 2026

fix(test): export test models with atomic_virial=True for .pte/.pt2
Han Wang
committed
711a1f4
perf(pt2): cache firstneigh_tensor across timesteps
Han Wang
committed
acba914
test(pt2): add regression test for oversized nlist with distance sorting
Han Wang
committed
9426572
test(pt2): add regression test for oversized nlist with distance sorting
Han Wang
committed
97870bf
refactor(pt2): remove need_sorted_nlist_for_lower monkey-patch
Han Wang
committed
755c2d3
refactor(pt2): remove redundant C++ nnei+1 nlist padding
Han Wang
committed
444b4e5
fix(test): link torch in C++ test binary, fix spin PtExpt guards
Han Wang
committed
583df2e

Commits on Apr 22, 2026

refactor(cc): unify createNlistTensor and fix oversized nlist handling
Han Wang
committed
862f560
fix(test): remove unused _DESCRIPTOR_DPA1_WITH_ATTN config
Han Wang
committed
12ad03b
fix(pt2): relax nnei dynamic shape lower bound from 2 to 1
Han Wang
committed
b0096e0

Commits on Apr 23, 2026

fix(pt2): remove eliminate_dead_code from _strip_shape_assertions
Han Wang
committed
afa732f

Commits on Apr 24, 2026

fix(pt2): neutralise shape assertions instead of erasing them
Han Wang
committed
59973be
fix(pt2): prevent CUDA NaN in attention backward by disabling kernel fusion
Han Wang
committed
39f84ac
fix(pt2): only apply realize_opcount_threshold=0 on CUDA
Han Wang
committed
57cfe3e

Commits on Apr 25, 2026

refactor(pt_expt): use extra_nlist_sort override instead of stripping shape assertions
Han Wang
committed
cdbfead
chore(pt2): note thread-safety constraint of realize_opcount_threshold
Han Wang
committed
f480456
feat(cli): warn instead of raise for --atomic-virial on non-pt_expt outputs
Han Wang
committed
d41eefd
refactor(pt2): move atomic_virial check to compute() entry; clarify error message
Han Wang
committed
c80db58
fix(cc): guard empty-vector dereferences in NeighborListData and convert_nlist
Han Wang
committed
72f95f8
test(cc): tighten oversized-nlist sanity check; verify all predictions
Han Wang
committed
43d42e6
fix(cc): apply empty-vector guard to make_inlist and DeepSpinTF; add backend-agnostic tests
Han Wang
committed
e30206d
test(cc): regression test for atomic_virial fail-fast guard
Han Wang
committed
9d2f577