tpt-gpu-bench 0.1.0

User-runnable dynamic benchmark suite for TPT GPU kernels
tpt-gpu-bench-0.1.0 is not a library.

TPT-GenBench

User-runnable dynamic GPU benchmark harness. Runs a set of GEMM/Attention/Conv2D workloads defined in a bench.toml config against a detected (or simulated) GPU, checks correctness against a tolerance, and compares timing against a baseline profile in tuning/.

Prerequisites

Installation

cd crates/tpt-gpu-bench
cargo build --release

Usage

# Run using an example config, without real hardware (GPU="sim")
cargo run -- --config examples/quick_smoke.toml --gpu sim

# Run against detected hardware, writing results to a custom directory
cargo run -- --config bench.toml --output results/

# Override warmup/iteration counts from the CLI
cargo run -- --config bench.toml --warmup 5 --iterations 20

# After a run, write a candidate tuning/<gpu>.json profile for contribution
cargo run -- --config bench.toml --contribute

# Aggregate results/<gpu>-<ts>.json files into the BENCHMARKS.md scoreboard
cargo run -- --scoreboard

Example configs live in examples/ — quick_smoke.toml (fast CI sanity check), llama3_shapes.toml, and resnet50_shapes.toml (workload shapes matching those models). Write your own bench.toml following the same [target] / [run] / [[workload]] structure.

Results feed the community scoreboard in BENCHMARKS.md.