rlx-bench 0.2.1

PLAN L5 — uniform benchmark harness for RLX backends + patterns

Coverage
62.86%
22 out of 35 items documented0 out of 16 items with examples
Size
Source code size: 75.2 kB This is the summed size of all the files inside the crates.io package for this release.
Documentation size: 859.1 kB This is the summed size of all files generated by rustdoc for all configured targets
Ø build duration
this release: 20s Average build duration of successful builds.
all releases: 24s Average build duration of successful builds in releases after 2024-10-23.
Links
Homepage
MIT-RLX/rlx
32 7 1
crates.io
Dependencies
Versions
Owners

rlx-bench

Uniform benchmark harness for RLX backends + workload patterns. The canonical answer for "how fast does my graph run on Device X with PrecisionPolicy Y?"

What's here

BenchmarkPattern — common workload shapes (matmul-only, matmul + GELU, full FFN block, attention-only, end-to-end BERT layer).
Per-backend timing harness — measures each pattern across Device::Cpu / Metal / Mlx / Cuda / Rocm / Wgpu / Tpu / Fpga (whichever are enabled), reports p50 / p95 / GFLOP/s.
Examples:
- bench_all — sweep every (pattern × device × policy) cell.
- bench_autodiff — measure reverse-mode AD overhead per op.

Install

[dependencies]
rlx-bench = "0.1"

Run

cargo run -p rlx-bench --release --example bench_all

License

GPL-3.0-only.