Tracking Issue: Tensor Extension Types #6865

This issue tracks tensor extension types in vortex-tensor. These types support ML, AI, and scientific data, including images, video, sensor data, time-series data, embeddings, and matrices. Parent Epic: Extension Types #7683.

Design

Tensor support uses Vortex extension types with canonical storage dtypes. Tensor metadata and scalar functions live in vortex-tensor.

The tracked tensor types are:

• FixedShapeTensor: A tensor with a fixed physical shape and dimension count.
• VariableShapeTensor: A tensor whose shape can change while its dimension count stays fixed.
• Matrix: A two-dimensional tensor.
• Vector: A one-dimensional tensor.

Normalization is an explicit compute operation, not a logical refinement or physical encoding:

• L2Norm(v) returns the magnitude of each tensor or vector.
• L2Normalize(v) returns a struct with the normalized value and the same norm.

The scalar functions compute from decoded coordinates. They do not use encoding-specific arithmetic shortcuts.

Design references:

• Extension Types RFC
• Fixed-shape Tensor RFC
• Apache Arrow canonical extensions

Steps

• Establish the extension-type support that tensor work requires.
• Add the initial fixed-shape tensor extension type in vortex-tensor.
• Reorganize vortex-tensor for multiple tensor-related types.
• Add the initial Vector extension type and L2Norm expression.
• Define explicit L2 normalization semantics without a refinement dtype or physical encoding.
• Add logical shape casts for fixed-shape tensors.
• Add Arrow export support.
• Add NumPy export support.
• Add PyTorch export support.
• Add variable-shape tensor support.
• Documentation.
• Public API stabilization.

Unresolved questions

• Define the intended stable public API for tensor extension types.
• Decide the relationship between FixedShapeTensor, VariableShapeTensor, Matrix, and Vector.
• Decide what validation belongs in tensor extension metadata versus storage-array validation.
• Define export semantics for Arrow, NumPy, and PyTorch.
• Decide whether tensor slicing and indexing produce lazy tensor views, materialized arrays, or both.

Implementation history

• Extension Types RFC established the extension-type design used by tensor work.
• Fixed-shape Tensor RFC described the first tensor extension design.
• #6812 added the experimental fixed-shape tensor extension and CosineSimilarity.
• #6857 reorganized vortex-tensor for more tensor-related types.
• #6964 added Vector, L2Norm, and AnyTensor.
• #9767 replaces the Normalized encoding with explicit L2Normalize semantics.
• #9768 moves L2Norm to RowFn.
• #9769 moves inner product and cosine similarity to RowFn.