v2.0.0

v2.0.0 — Profile v2

Profile is a physics-grounded optimizer for vLLM. It measures live GPU and scheduler telemetry against your hardware's real decode/prefill ceiling and tells you what happened, why it was slow, and exactly what to fix first.

Traffic-gated analysis. Idle servers are ignored. Under load, Profile ranks five production bottlenecks — under-batching, KV cache pressure, prefix reuse, concurrency saturation, OOM risk — and surfaces one prescriptive fix per iteration.

Closed-loop verification. Run profile diagnose, apply the fix, and Profile detects the vLLM restart automatically. It re-measures and reports throughput, latency, efficiency, and cost deltas with a definitive better / worse / plateau signal.

Physics roofline. Compute and memory-bandwidth ceilings per GPU. Efficiency % and headroom against the theoretical maximum — not a guess, a derivable number.

Economics. tok/W, $/1M tok, and joules/token tracked across every iteration.

Zero overhead. Single Rust binary. Scrape-only (NVML + /metrics). Strictly local. No external APIs.

profile diagnose --url http://localhost:8000/metrics --duration 2m

Docs

Install profile 2.0.0

Install prebuilt binaries via shell script

curl --proto '=https' --tlsv1.2 -LsSf https://github.com/jungledesh/profile/releases/download/v2.0.0/profile-installer.sh | sh

Install prebuilt binaries via powershell script

powershell -ExecutionPolicy Bypass -c "irm https://github.com/jungledesh/profile/releases/download/v2.0.0/profile-installer.ps1 | iex"

Download profile 2.0.0

File	Platform	Checksum
profile-aarch64-apple-darwin.tar.xz	Apple Silicon macOS	checksum
profile-x86_64-apple-darwin.tar.xz	Intel macOS	checksum
profile-x86_64-pc-windows-msvc.zip	x64 Windows	checksum
profile-aarch64-unknown-linux-gnu.tar.xz	ARM64 Linux	checksum
profile-x86_64-unknown-linux-gnu.tar.xz	x64 Linux	checksum
profile-x86_64-unknown-linux-musl.tar.xz	x64 MUSL Linux	checksum