rlx-bench 0.2.2

PLAN L5 — uniform benchmark harness for RLX backends + patterns
Documentation

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.
    • bench_fft — batch × N sweep across backends; set RLX_BENCH_DISPATCH_ONLY=1 on wgpu to skip readback and isolate dispatch time.

Install

[dependencies]
rlx-bench = "0.2"

FFT benchmark

just throttle
cargo run -p rlx-bench --release --example bench_fft --features metal,gpu
RLX_BENCH_DISPATCH_ONLY=1 cargo run -p rlx-bench --release --example bench_fft --features gpu
./scripts/bench_fft_rig.sh   # remote CUDA rig (Windows + WSL)

Run

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

License

GPL-3.0-only.