feat(dpmodel): backend-agnostic DPA4 descriptor and fitting (core path) - #5515

Merged

wanghan-iapcm merged 31 commits into

deepmodeling:masterfrom

wanghan-iapcm:feat-dpmodel-dpa4

Jun 12, 2026

Commits on Jun 11, 2026

feat(dpmodel): relocate Lebedev rules with numpy loader
Han Wang
committed
03cc9ee
test(dpmodel): cover Lebedev loader error paths; review fixes
Han Wang
committed
eccbe4d
feat(dpmodel): general-lmax real spherical harmonics (e3nn conventions)
Han Wang
committed
ea0c471
fix(dpmodel): match e3nn zero-vector handling in real SH; e3nn-free convention tests
Han Wang
committed
f8798e1
feat(dpmodel): dpa4_nn indexing + utils
Han Wang
committed
18dde4b
test(dpmodel): accurate parity-file docstring, import order, literal index anchor
Han Wang
committed
8a12493
feat(dpmodel): dpa4_nn radial basis, envelope, radial MLP
Han Wang
committed
f51d986
feat(dpmodel): dpa4_nn Wigner-D calculator + edge quaternions
Han Wang
committed
80fc4a8
feat(dpmodel): dpa4_nn norms, SO3 linear, gated activation
Han Wang
committed
3916c21
test(dpmodel): tighten fp32 parity tolerances to measured ulp
Han Wang
committed
dee01f5
feat(dpmodel): dpa4_nn S2 grid projection + grid nets (Lebedev)
Han Wang
committed
23107df
test(dpmodel): keep fp32 parity tests tolerance-based for truncation robustness
Han Wang
committed
3115285
feat(dpmodel): dpa4_nn SO2 linear + convolution with padded-edge aggregation
Han Wang
committed
66be7c3
docs(dpmodel): clarify EdgeCache padded-layout and cache-reuse contract
Han Wang
committed
a89ce39
feat(dpmodel): dpa4_nn type, geometric, environment embeddings
Han Wang
committed
2d6fa3a
fix(dpmodel): uniform precision-key lookup; honest GIE zonal test scope note
Han Wang
committed
fab9bf3
feat(dpmodel): dpa4_nn padded edge-cache builder
Han Wang
committed
abc0574
feat(dpmodel): dpa4_nn equivariant FFN + interaction block
Han Wang
committed
0c7c822
feat(dpmodel): DPA4 descriptor (core path)
Han Wang
committed
c8ebada
feat(dpmodel): dpa4_ener GLU energy fitting
Han Wang
committed
ce2e2a9
test(dpmodel): DPA4 cross-backend consistency tests and model-dict contract
Han Wang
committed
b2bbfff
[pre-commit.ci] auto fixes from pre-commit.com hooks
pre-commit-ci[bot]
committed
80b4ba1
fix(dpmodel): address review findings (xp-namespace constants, validation, lint)
Han Wang
committed
93e4d0b
test(dpmodel): ulp tolerance for fp32/fp16 safe_norm parity (CI BLAS variance)
Han Wang
committed
a625528
style(dpmodel): NumPy capitalization; explicit dtypes for pylint
Han Wang
committed
225c8c5
fix(dpmodel): xp-namespace conversion for remaining numpy buffers; KeyError in collection lookup
Han Wang
committed
70702ce
test(dpmodel): pin pt reference modules to CPU in parity tests (CUDA runners)
Han Wang
committed
177c057
ci: retrigger checks
Han Wang
committed
cb395c8

Commits on Jun 12, 2026

test(dpmodel): also patch import-time device snapshots for CPU parity pinning
Han Wang
committed
0b37974
test(dpmodel): run pt parity references on native device with device-conditional tolerances
Han Wang
committed
33180fa
test(dpmodel): move edge-cache fixture tensors to the pt device too
Han Wang
committed
ed0e303

Navigation Menu

Uh oh!

feat(dpmodel): backend-agnostic DPA4 descriptor and fitting (core path) - #5515

feat(dpmodel): backend-agnostic DPA4 descriptor and fitting (core path)#5515
wanghan-iapcm merged 31 commits into
deepmodeling:masterfrom
wanghan-iapcm:feat-dpmodel-dpa4

Commits on Jun 11, 2026

feat(dpmodel): relocate Lebedev rules with numpy loader

test(dpmodel): cover Lebedev loader error paths; review fixes

feat(dpmodel): general-lmax real spherical harmonics (e3nn conventions)

fix(dpmodel): match e3nn zero-vector handling in real SH; e3nn-free convention tests

feat(dpmodel): dpa4_nn indexing + utils

test(dpmodel): accurate parity-file docstring, import order, literal index anchor

feat(dpmodel): dpa4_nn radial basis, envelope, radial MLP

feat(dpmodel): dpa4_nn Wigner-D calculator + edge quaternions

feat(dpmodel): dpa4_nn norms, SO3 linear, gated activation

test(dpmodel): tighten fp32 parity tolerances to measured ulp

feat(dpmodel): dpa4_nn S2 grid projection + grid nets (Lebedev)

test(dpmodel): keep fp32 parity tests tolerance-based for truncation robustness

feat(dpmodel): dpa4_nn SO2 linear + convolution with padded-edge aggregation

docs(dpmodel): clarify EdgeCache padded-layout and cache-reuse contract

feat(dpmodel): dpa4_nn type, geometric, environment embeddings

fix(dpmodel): uniform precision-key lookup; honest GIE zonal test scope note

feat(dpmodel): dpa4_nn padded edge-cache builder

feat(dpmodel): dpa4_nn equivariant FFN + interaction block

feat(dpmodel): DPA4 descriptor (core path)

feat(dpmodel): dpa4_ener GLU energy fitting

test(dpmodel): DPA4 cross-backend consistency tests and model-dict contract

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

fix(dpmodel): address review findings (xp-namespace constants, validation, lint)

test(dpmodel): ulp tolerance for fp32/fp16 safe_norm parity (CI BLAS variance)

style(dpmodel): NumPy capitalization; explicit dtypes for pylint

fix(dpmodel): xp-namespace conversion for remaining numpy buffers; KeyError in collection lookup

test(dpmodel): pin pt reference modules to CPU in parity tests (CUDA runners)

ci: retrigger checks

Commits on Jun 12, 2026

test(dpmodel): also patch import-time device snapshots for CPU parity pinning

test(dpmodel): run pt parity references on native device with device-conditional tolerances

test(dpmodel): move edge-cache fixture tensors to the pt device too