Interfaces

TOML: the single source of truth

Every front-end consumes the same config format: [geometry] (typed lattice plus basis atoms), [grid], [path] (presets or explicit k-points), ordered [[sweeps]], [eigensolver], [solver] (including precision), and [output]. One file describes a single solve or a ten-thousand-job sweep.

CLI

blaze runs one band structure; blaze-bulk runs sweeps with fixed or adaptive threading. Output is CSV band tables (k_index, kx, ky, k_distance, band1..bandN), plus optional dielectric dumps and per-iteration convergence diagnostics as JSON.

Python: streaming without holding the GIL

The blaze2d package exposes the PyO3 module blaze._native. Its central design point: BulkDriver.run_streaming() spawns a Rust worker thread and releases the GIL while compute runs. Results flow over a crossbeam channel to a lazy BandResultIterator, so Python iterates finished band structures while the sweep is still running, and memory stays flat regardless of sweep size:

from blaze import BulkDriver
 
driver = BulkDriver("sweep.toml", threads=0)   # 0 = all cores
for result in driver.run_streaming():
    bands = result["bands"]                    # bands[k][band], already tracked

The marshaling into Python dicts happens per result at the boundary: bounded, visible data movement (board). A separate OperatorDataExtractor exposes the eigenvector-level quantities (velocity matrices, inverse-mass tensors, Born-Huang potentials) that the envelope-approximation research consumes.

WebAssembly: this site runs the real engine

The same core compiles to wasm32 (CpuBackend<f64>; rustfft works unchanged, nalgebra replaces faer for the dense step). WasmBulkDriver accepts the identical TOML string and streams one callback per solved k-point. The Examples and Workbench pages run it inside a Web Worker: every band diagram you compute in this browser is the production solver, not a demo.

The boundary rule

All three doors obey the same discipline you can trace on the board: compute stays inside Rust; boundaries move results, row by row, never intermediate state. The expensive data (eigenvector blocks, ε fields, scratch) never crosses a language boundary at all.