Crate zsfm_burn 

Burn-backend inference.

Shared Burn tensor ops (linear, norm, attention, RoPE), direct GGUF weight loading via GgufWeights, and op microbenchmarks. No model weights needed for the op checks: they use fixed synthetic inputs so they run offline in CI.

Every measurement lands in benchmark/burn_migration.md.

Re-exports

pub use bench::bench_op;

pub use bench::max_abs_err;

pub use bench::OpMeasurement;

pub use gguf_weights::open_gguf_weights;

pub use gguf_weights::GgufWeights;

Modules

attn

bench

gguf_weights

Direct GGUF weight loading for Burn inference — no candle.

linear

norm

rope