The one open piece of the performance story (still to be spiked). An aggregate-HashJoin that computes a grouped contraction Σⱼ A·B WITHOUT materializing the full join, dispatching to a matmul/BLAS kernel on the dense path — implemented as a DataFusion ExecutionPlan. Keeps the model pure-logical (one value per cell — the portable surface) while pushing BLAS-class speed into the physical plan; deliberately does NOT adopt Tang et al.'s chunked-tensor values. (design.md §4.1)
Curated from the project roadmap; see docs/design.md §8 for the milestone plan.
The one open piece of the performance story (still to be spiked). An aggregate-HashJoin that computes a grouped contraction Σⱼ A·B WITHOUT materializing the full join, dispatching to a matmul/BLAS kernel on the dense path — implemented as a DataFusion ExecutionPlan. Keeps the model pure-logical (one value per cell — the portable surface) while pushing BLAS-class speed into the physical plan; deliberately does NOT adopt Tang et al.'s chunked-tensor values. (design.md §4.1)
Curated from the project roadmap; see
docs/design.md§8 for the milestone plan.