solve#

Explicit triangular inverse policies for differentiable prefill emulation.

Functions

neumann_inverse_reference

Independently sum powers 0..degree, including the actual polynomial gradient.

triangular_inverse

Materialize exact or polynomial inverse of I plus the strict lower triangle.
neumann_inverse_reference(lower, degree)#

Independently sum powers 0..degree, including the actual polynomial gradient.

The strictly lower triangle is used. There is no residual-based fallback. Use FP64 to distinguish polynomial error from working-arithmetic error.

Parameters:

  • lower (Tensor)

  • degree (int)

Return type:

Tensor

triangular_inverse(lower, policy)#

Materialize exact or polynomial inverse of I plus the strict lower triangle.

Materializing the inverse retains the WY operand-QDQ sites. The Triton option requires CUDA FP32 chunk-64 matrices and supports first-order gradients only.

Parameters:
Return type:

Tensor