• Fixed a memory bug that triggered anytime the kernel needed to reallocate memory for a growing output tensor.

• Fixed bug in LLVM code generation where loops with too many iterations overflowed the stack. All local variables are now hoisted when using the LLVM as the backend compiler.

• Added Windows support. This only covers the base package and the tensora[numpy] extra. The cffi and taco extras are still only available on Linux and Mac.
• Fixed sum nodes not fusing iterations loops of the same variable when the sum node also had nodes of other iteration variables or terminal expressions. This is an optimization improvement for the LLVM backend and a invalid code fix for the C code generation.

Tensora now uses llvmlite as the backend compiler by default. The old method, using cffi to shell out to the system C compiler, is still available.

• Add llvmlite as a backend compiler
  • Put the cffi compiler behind the tensora[cffi] extra
    • Make setuptools a dependency only when this extra is installed
  • Add backend_compiler option to tensor_method with BackendCompiler.llvm and BackendCompiler.cffi as legal values
  • Add evaluate_cffi, which uses Tensora as the tensor compiler and cffi as the backend compiler
  • Change evaluate to use LLVM as the backend compiler, while still using Tensora as the tensor compiler
  • Add CLI option --language with "c" and "llvm" as legal values
• Remove HashOutput, which never actually worked
• Add support for Python 3.13
• Add support for Returns 0.23, 0.24, and 0.25

This release splits out the TACO binary into a separate package tensora-taco. This package is optional, making the TACO code generator optional. The native Tensora code generator is still included and used by default. TACO can be still be used as an alternative code generator if the tensora[taco] extra is installed.

Also, these other changes:

• The signature of tensor_method was changed so that the output_format and input_formats were combined into one formats dictionary
• Renamed PureTensorMethod to TensorMethod to reflect the decision that mutable assignments will never be supported in Tensora
• Most of the attributes on TensorMethod were made private
• Cache on a parsed Problem, instead of on the string versions of assignments and formats
• Raise a nice exception on TensorMethod when trying to broadcast to target indexes because there is no way for the dimensions of those indexes to be known. Generating such kernels via the tensora CLI is still permitted.
• Raise a nice exception when a tensor and and index have the same name in an assignment
• Remove Tensor.from_scalar; you could always use Tensora.from_lol
• Do not cast the result of a vector inner product a @ b to a Python scalar; leave it as a scalar Tensor

This release removes a lot of partially implemented features, making the remaining behavior more reliable.

• Scalars are no longer supported. Users must use x() instead of x in all expressions.
• Added option to tensora CLI to allow choosing the external TACO binary as the code generator
• Better exceptions are raised for many kinds of invalid input
• Appended an underscore in front of many Python files to make is explicit that they are not part of the external API

This is a quick release to fix some issues in the previous release.

• Raise a nice exception when no solution is found by the tensor algebra compiler
• Rename assemble to assemble everywhere so that the tense is consistent with compute and evaluate
• Fix bug where compute code was being emitted when assemble code was requested

This release includes two major changes:

• A new tensor algebra compiler written in Python specifically for Tensora
  • Added tensora.TensorCompiler enum with values TensorCompiler.tensora and TensorCompiler.taco
  • tensor_method accepts an optional argument compiler: TensorCompiler that selects the tensor algebra compiler; tensora is the default
  • Split evaluate into evaluate_tensora and evaluate_taco to select the tensor algebra compiler; evaluate is an alias for evaluate_tensora
• A tensora CLI that will emit the code for given tensor algebra problems

And one minor change:

• Dropped support for Python 3.8 and 3.9 (allows for the match statement, which was helpful in writing the new tensora algebra compiler)