Product
One optimization layer, from detection to remediation.
Not three products — one pipeline. Detect what's wrong, decide and fix it with a human in the loop, at the scale of your entire fleet.
→
Telemetry
vLLM, Triton, KServe, SGLang, Prometheus — read continuously
01
Detect
Your cluster's X-ray.
02
Decide & Fix
Detect. Recommend. Fix — with a human in the loop.
03
Fleet Scale
The same detect-and-fix loop, across every cluster you operate.
Paralleliq is one optimization layer that runs this loop continuously. It starts with piqc, our free open-source scanner, or a lightweight always-on agent — either way, it maps every model in your fleet to the hardware it's running on. Every finding becomes a recommendation with its dollar impact shown up front; nothing executes until an operator approves it, and every approved change is reversible and logged under a named operator. And because it was built to run this way from the start, the same loop covers one cluster or every cluster you operate — including your customers' — in a single view.
- 1Auto-discovers vLLM, Triton, KServe, SGLang, Ollama, TGI
- 1Understands model-hardware fit — not just GPU utilization
- 1Continuous safety signals: KV cache pressure, OOM risk, queue depth — every 15 seconds
- 1Performance and structural signals collected at longer intervals — zero added load on workloads
- 1Reads from your existing Prometheus — no duplicate scraping
- 2Dollar-impact framing per finding — margin recovery for providers, cost reduction for teams
- 2Blast radius preview before any change executes
- 2Operator approval queue with role-based controls
- 2Immutable audit trail — every decision logged under a named operator
- 2One-click rollback for any approved action
- 3Multi-cluster, multi-cloud, multi-region in one view
- 3Covers your clusters and your customers' clusters equally
- 3Findings attributed to workload, customer, and cluster automatically
- 3Cost and capacity intelligence across the full fleet
paralleliq.yaml — one config, one loop
discover: runtimes: [vllm, triton, kserve, sglang] mode: read-only policy: kv-cache-pressure: when: kv_hit_rate < 0.4 recommend: [rebalance, evict_idle_replicas] requires: operator_approval fleet: clusters: [us-east-prod, eu-west-prod, on-prem-dc1] view: unified