v0.1.0a5
Highlights
- Added native
vllm bench serveworkloads alongside GuideLLM. - Added opt-in parallel trials on explicit, non-overlapping GPU assignments.
- Kept lifecycle, scoring, artifacts, and reports consistent across backends.
Release artifacts
| Platform | Install |
|---|---|
| PyPI | pip install "vtune==0.1.0a5" |
| Linux/WSL runtime | pip install "vtune[runtime]==0.1.0a5" |
The universal wheel and source distribution are attached to the GitHub release.
Benchmark engine
- Added
benchmark.engine: vllm, forward-compatible benchmark arguments, canonical throughput aliases, request accounting, raw JSON, logs, commands, versions, repeats, retries, and timeout evidence. - Documented random, ShareGPT, Hugging Face, custom, and prefix-repetition data.
Parallel execution
- Added
execution.mode: local_parallelwith named workers, GPU sets, and deterministic ports. - Rejected overlapping devices, GPU sharing, manual ports, and incompatible tensor parallelism before execution.
- Isolated failures and cleaned owned processes on cancellation.
Validation and limitations
- All 147 private tests and 23 subtests passed.
- Python 3.11/3.12 package, documentation, metadata, and controlled two-process integration checks passed.
- One OPT-125M vLLM 0.28.0 WSL2 run passed on an RTX 3080. Real multi-GPU execution was not tested; parallel measurements remained experimental.
Full changelog: v0.1.0a4...v0.1.0a5.