pip: bump nvalchemi-toolkit-ops from 0.4.0 to 0.4.1

[dependabot]

Bumps nvalchemi-toolkit-ops from 0.4.0 to 0.4.1.

Release notes

Sourced from nvalchemi-toolkit-ops's releases.

v0.4.1 Release Notes

This is a hotfix release. It corrects the PyTorch Ewald/PME autograd path — energy gradients for connected charge/position/cell inputs, the reciprocal strain virial, and torch.compile support for explicit single-system batches — and hardens the neighbor-list validation surface for cluster-tile and cell-list builds. It also adds a keyword-only energy_reduction selector to the monopole electrostatics entry points and optional tile-state passthrough for PyTorch cluster-tile selective calls.

There are no API removals. Several paths that previously accepted malformed or ambiguous input now fail closed; see Behavior Changes.

---

Behavior Changes

• Compiled unified PyTorch cluster-tile dispatch rejects tensor-valued PBC (#141): it previously treated a tensor pbc as fully periodic. PBC is now validated eagerly and the direct single-system fixed-state route is compiled instead.
• Single-system Torch and JAX segmented cluster-tile COO require one exact physical interval (#141): writes are bounded by output capacity, malformed offsets fail closed, and compiled/JIT active counts are capped to writable capacity. Batched per-system physical subsegments remain supported.
• Fused Warp graph calls with an explicit max_total_cells now require an explicit neighbor_search_radius (#129).

---

Additions/New Features

• energy_reduction on monopole electrostatics (#132): Torch and JAX Ewald, PME, and slab entry points accept keyword-only energy_reduction="atom" | "system" (default "atom"). "atom" returns per-atom energies (N,); "system" returns per-system totals (B,). Direct-output fields (forces, charge gradients, virials) keep their existing shapes. Torch eager atom mode may synchronize once per participating component when a materialized uniform cotangent is proven by value inspection; system mode is structurally sync-free for arbitrary (B,) loss weights. JAX adds API/layout parity only; the underlying Warp kernels remain atom-buffer-oriented.
• Caller-owned tile state from PyTorch cluster-tile selective calls (#141): pass return_state=True to append the tile state to the return, without changing the default neighbor-list return arity.

---

Performance

• Faster JAX neighbor-list import (#130): dtype-specific direct naive and cell-list Warp wrapper registration is deferred until first use, and cluster-tile graph callbacks use bundled callback/preload registrations with lazy direct kernels

... (truncated)

Changelog

Sourced from nvalchemi-toolkit-ops's changelog.

0.4.1 - 2026-08-03

Added

• Monopole Torch and JAX Ewald, PME, and slab entry points accept keyword-only energy_reduction="atom" | "system" (default "atom"). "atom" returns per-atom energies (N,); "system" returns per-system totals (B,). Direct-output fields (forces, charge gradients, virials) keep their existing shapes. Torch eager atom mode may synchronize once per participating component when a materialized uniform cotangent is proven by value inspection; system mode is structurally sync-free for arbitrary (B,) loss weights. JAX adds API/layout parity only; underlying Warp kernels remain atom-buffer-oriented.
• PyTorch cluster-tile selective calls can append caller-owned tile state with return_state=True, without changing the default neighbor-list return arity.

Changed

• Improved JAX neighbor-list import performance by deferring dtype-specific direct naive and cell-list Warp wrapper registration until first use. Cluster-tile graph callbacks now use bundled callback/preload registrations with lazy direct kernels for naive and cell-list paths. Public behavior is unchanged.

Fixed

• Fixed Torch PME and Ewald energy gradients for connected charge, position, and cell inputs. Non-uniform or weighted energy losses and create_graph=True higher-order derivatives no longer double-count upstream chain-rule terms.
• Torch ewald_reciprocal_space now preserves graph-connected reciprocal vectors for cell/strain autograd, restoring the physical reciprocal Ewald virial when vectors are regenerated from the differentiable cell.
• Torch Ewald, PME, and slab backward paths now compile when an explicit single-system batch (batch_idx=zeros(N)) is supplied. Reciprocal PME compiled gradients are also correct when a compiled function is reused across mesh sizes.
• Torch DFT-D3 custom operators now zero caller-owned energy, forces, coordination-number, and virial buffers before empty-system or zero-edge early returns, so reused output tensors cannot retain stale values.
• JAX DFT-D3 CSR calls with atoms but no edges return zero-filled per-atom forces and coordination numbers with shapes (N, 3) and (N,), matching the neighbor-matrix contract and preserving per-system energy and virial axes.
• JAX cell-list builds now derive search radii from their realized grids, preventing missed neighbors when static capacity changes the constructed grid. Batched capacity_strategy="geometry" preserves promoted grids for all non-empty systems by reserving an equal per-system capacity; volume-based sizing remains the default. Fused Warp graph calls with explicit max_total_cells now require an explicit neighbor_search_radius.
• Unbatched JAX naive dual-cutoff PBC neighbor lists now populate both cutoff outputs when using the default wrap_positions=True. Previously this path wrapped positions but skipped the fill kernel, leaving zero counts and padded

... (truncated)

Commits

• 26dbceb fix changelog
• c4eabf9 updates from comments
• 895e6c8 small bug fixes
• 3f70e30 bump to 0.4.1
• 200452c docs: complete public API docstrings (#139)
• ad76db8 fix(dispersion): handle empty DFT-D3 neighbor graphs (#140)
• 56d9ff9 fix(electrostatics): restore reciprocal strain virial (#138)
• 5aaa8b1 fix(neighbors): harden selective cluster-tile state (#141)
• 0ed5a78 fix(electrostatics): dechain dependent autograd inputs (#131)
• 22b7e16 fix(neighbors): fill wrapped JAX dual-cutoff PBC lists (#137)
• Additional commits viewable in compare view

Note
Automatic rebases have been disabled on this pull request as it has been open for over 30 days.

[github-actions]

Original Dependabot description

Bumps nvalchemi-toolkit-ops from 0.4.0 to 0.4.1.

Release notes

Sourced from nvalchemi-toolkit-ops's releases.

v0.4.1 Release Notes

This is a hotfix release. It corrects the PyTorch Ewald/PME autograd path — energy gradients for connected charge/position/cell inputs, the reciprocal strain virial, and torch.compile support for explicit single-system batches — and hardens the neighbor-list validation surface for cluster-tile and cell-list builds. It also adds a keyword-only energy_reduction selector to the monopole electrostatics entry points and optional tile-state passthrough for PyTorch cluster-tile selective calls.

There are no API removals. Several paths that previously accepted malformed or ambiguous input now fail closed; see Behavior Changes.

---

Behavior Changes

• Compiled unified PyTorch cluster-tile dispatch rejects tensor-valued PBC (#141): it previously treated a tensor pbc as fully periodic. PBC is now validated eagerly and the direct single-system fixed-state route is compiled instead.
• Single-system Torch and JAX segmented cluster-tile COO require one exact physical interval (#141): writes are bounded by output capacity, malformed offsets fail closed, and compiled/JIT active counts are capped to writable capacity. Batched per-system physical subsegments remain supported.
• Fused Warp graph calls with an explicit max_total_cells now require an explicit neighbor_search_radius (#129).

---

Additions/New Features

• energy_reduction on monopole electrostatics (#132): Torch and JAX Ewald, PME, and slab entry points accept keyword-only energy_reduction="atom" | "system" (default "atom"). "atom" returns per-atom energies (N,); "system" returns per-system totals (B,). Direct-output fields (forces, charge gradients, virials) keep their existing shapes. Torch eager atom mode may synchronize once per participating component when a materialized uniform cotangent is proven by value inspection; system mode is structurally sync-free for arbitrary (B,) loss weights. JAX adds API/layout parity only; the underlying Warp kernels remain atom-buffer-oriented.
• Caller-owned tile state from PyTorch cluster-tile selective calls (#141): pass return_state=True to append the tile state to the return, without changing the default neighbor-list return arity.

---

Performance

• Faster JAX neighbor-list import (#130): dtype-specific direct naive and cell-list Warp wrapper registration is deferred until first use, and cluster-tile graph callbacks use bundled callback/preload registrations with lazy direct kernels

... (truncated)

Changelog

Sourced from nvalchemi-toolkit-ops's changelog.

0.4.1 - 2026-08-03

Added

• Monopole Torch and JAX Ewald, PME, and slab entry points accept keyword-only energy_reduction="atom" | "system" (default "atom"). "atom" returns per-atom energies (N,); "system" returns per-system totals (B,). Direct-output fields (forces, charge gradients, virials) keep their existing shapes. Torch eager atom mode may synchronize once per participating component when a materialized uniform cotangent is proven by value inspection; system mode is structurally sync-free for arbitrary (B,) loss weights. JAX adds API/layout parity only; underlying Warp kernels remain atom-buffer-oriented.
• PyTorch cluster-tile selective calls can append caller-owned tile state with return_state=True, without changing the default neighbor-list return arity.

Changed

• Improved JAX neighbor-list import performance by deferring dtype-specific direct naive and cell-list Warp wrapper registration until first use. Cluster-tile graph callbacks now use bundled callback/preload registrations with lazy direct kernels for naive and cell-list paths. Public behavior is unchanged.

Fixed

• Fixed Torch PME and Ewald energy gradients for connected charge, position, and cell inputs. Non-uniform or weighted energy losses and create_graph=True higher-order derivatives no longer double-count upstream chain-rule terms.
• Torch ewald_reciprocal_space now preserves graph-connected reciprocal vectors for cell/strain autograd, restoring the physical reciprocal Ewald virial when vectors are regenerated from the differentiable cell.
• Torch Ewald, PME, and slab backward paths now compile when an explicit single-system batch (batch_idx=zeros(N)) is supplied. Reciprocal PME compiled gradients are also correct when a compiled function is reused across mesh sizes.
• Torch DFT-D3 custom operators now zero caller-owned energy, forces, coordination-number, and virial buffers before empty-system or zero-edge early returns, so reused output tensors cannot retain stale values.
• JAX DFT-D3 CSR calls with atoms but no edges return zero-filled per-atom forces and coordination numbers with shapes (N, 3) and (N,), matching the neighbor-matrix contract and preserving per-system energy and virial axes.
• JAX cell-list builds now derive search radii from their realized grids, preventing missed neighbors when static capacity changes the constructed grid. Batched capacity_strategy="geometry" preserves promoted grids for all non-empty systems by reserving an equal per-system capacity; volume-based sizing remains the default. Fused Warp graph calls with explicit max_total_cells now require an explicit neighbor_search_radius.
• Unbatched JAX naive dual-cutoff PBC neighbor lists now populate both cutoff outputs when using the default wrap_positions=True. Previously this path wrapped positions but skipped the fill kernel, leaving zero counts and padded

... (truncated)

Commits

• 26dbceb fix changelog
• c4eabf9 updates from comments
• 895e6c8 small bug fixes
• 3f70e30 bump to 0.4.1
• 200452c docs: complete public API docstrings (#139)
• ad76db8 fix(dispersion): handle empty DFT-D3 neighbor graphs (#140)
• 56d9ff9 fix(electrostatics): restore reciprocal strain virial (#138)
• 5aaa8b1 fix(neighbors): harden selective cluster-tile state (#141)
• 0ed5a78 fix(electrostatics): dechain dependent autograd inputs (#131)
• 22b7e16 fix(neighbors): fill wrapped JAX dual-cutoff PBC lists (#137)
• Additional commits viewable in compare view

Dependabot will resolve any conflicts with this PR as long as you don't alter it yourself. You can also trigger a rebase manually by commenting @dependabot rebase.

---

Dependabot commands and options

You can trigger Dependabot actions by commenting on this PR:

• @dependabot rebase will rebase this PR
• @dependabot recreate will recreate this PR, overwriting any edits that have been made to it
• @dependabot show <dependency name> ignore conditions will show all of the ignore conditions of the specified dependency
• @dependabot ignore this major version will close this PR and stop Dependabot creating any more for this major version (unless you reopen the PR or upgrade to it yourself)
• @dependabot ignore this minor version will close this PR and stop Dependabot creating any more for this minor version (unless you reopen the PR or upgrade to it yourself)
• @dependabot ignore this dependency will close this PR and stop Dependabot creating any more for this dependency (unless you reopen the PR or upgrade to it yourself)

[frostedoyster]

This one is risky. nvalchemi updates break things sometimes. I would add more graph construction tests to this one before merging