FlashSlice

Fused Triton kernels for the slice/deslice bottleneck of physics-attention (Transolver-style layers): the membership tensor that couples every mesh point to every slice token is never written to memory, and the layer's memory stops growing with the token count.

This is a placeholder release. The kernels, the instrumented Transolver layer, the parity gate and the benchmarks are released together with the paper Does Transolver Need a Transformer? (under review); until then the repository is private and this package holds only the eager reference of the two operators, so that the interface is on record:

from flashslice.reference import slice_eager, deslice_eager

slice_eager(x_mid, fx_mid, weight, bias, tau) pools point features onto slots through a softmax membership; deslice_eager(x_mid, weight, bias, tau, tokens) broadcasts slot values back to the points through the same membership. The fused kernels implement exactly these two maps and their gradients.

Apache-2.0. Copyright 2026 Shizheng Wen.