Scope
Move the torch-only deepmd/kernels/ package to deepmd/pt_expt/kernels/.
The package is implemented through torch.library.triton_op, wrap_triton, and torch.ops.deepmd.*; it is not backend-agnostic and cannot be consumed by the dpmodel, TensorFlow, JAX, or Paddle backends. pt_expt is the long-term owner and the design target for these exportable kernels.
Dependency direction
During the legacy pt backend's deprecation window, pt should import kernels from pt_expt. The dependency must not point from pt_expt back to pt, so removing the legacy backend also removes the temporary imports cleanly.
Invariant
deepmd/dpmodel/** must not import the torch-specific kernels package.
This issue tracks the relocation separately from DPA1 CUDA performance PR #5758 and should include the existing DPA4 kernels in the same package-level move.
Scope
Move the torch-only
deepmd/kernels/package todeepmd/pt_expt/kernels/.The package is implemented through
torch.library.triton_op,wrap_triton, andtorch.ops.deepmd.*; it is not backend-agnostic and cannot be consumed by the dpmodel, TensorFlow, JAX, or Paddle backends.pt_exptis the long-term owner and the design target for these exportable kernels.Dependency direction
During the legacy
ptbackend's deprecation window,ptshould import kernels frompt_expt. The dependency must not point frompt_exptback topt, so removing the legacy backend also removes the temporary imports cleanly.Invariant
deepmd/dpmodel/**must not import the torch-specific kernels package.This issue tracks the relocation separately from DPA1 CUDA performance PR #5758 and should include the existing DPA4 kernels in the same package-level move.