development.txt

ISSUES:

- Automatically generated unique tensor names have local scope,
  thus the corresponding tensors with automatically generated
  names will have their names differ accross different processes.
  The only way to ensure proper tensor correspondence in global
  tensor operations is to make sure all participating processes
  execute the same algorithm in a consistent fashion, like SIMD.
  That is, the order of tensor operations across all participating
  processes must be consistent such that every encountered global
  tensor operation will receive the same tensor operand irrespective
  of the difference in the locally generated tensor names. Special
  care needs to be taken in iterating over associative tensor containers,
  to ensure that the keys are consistent accross all participating
  processes. For example, automatically generated tensor names
  cannot serve as keys in an iteration procedure since they are
  inconsistent accross different processes whereas tensor Ids
  can serve as keys since they are normally consistent accross
  different processes because the container was built with the
  same structure accross all processes.

BUGS:

- Fix the bug(s) in the tensor order reduction mechanism in the TalshNodeExecutor backend.


FEATURES:

- GPU-RAM persistence:
  1. Create large tensors on GPU;
  2. Use COPY_T for smaller tensors;
  3. Cache eviction policy: Evict minimally necessary ones (prefer smaller tensors);
  4. Limit the Host buffer size to be similar to the combined devices buffer size;
  5. Introduce manual knobs to keep tensors on GPU;

- TensorNetwork: Subnetwork replacement method:
  Contract replaced tensors, then replace the contracted
  tensor with a new tensor (sub)network.

- Implement the Renormalization procedure.

- Implement SAVE/LOAD API for TensorExpansion.

- Implement TensorNetwork slice computing Generator.

- Implement bl-D procedure.

- Implement conjugate gradient optimization procedure.

- Implement DIIS convergence accelerator.

- Tensor network bond dimension adaptivity in solvers.

- Implement constrained optimization for isometric tensors.

- Fix index slicing logic (switch to two passes):
  Pass 1: Compute the throughout volume (TH) for all indices;
          TH(index) = Sum of tensor volumes for all tensors containing the index.
  Pass 2: Slice indices with maximal throughout volume;

- Implement parameterized non-linear optimization:
  dTensorFunctional/dParameter = dTensorFunctional/dTensor * dTensor/dParameter

- Introduce DAG nodes/dependencies with multiple output operands.

- Implement the guided k-way partitioning algorithm.