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
- Rust toolchain (see repo root
README.md)
Installation
Usage
# Run using an example config, without real hardware (GPU="sim")
# Run against detected hardware, writing results to a custom directory
# Override warmup/iteration counts from the CLI
# After a run, write a candidate tuning/<gpu>.json profile for contribution
# Aggregate results/<gpu>-<ts>.json files into the BENCHMARKS.md 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.