Skip to main content

Dependencies

GPU profiling tools

GPU profiling reads the output of profilers the operator already runs. None of them is linked into, embedded in, or redistributed with InferCrane; the Go analysis reads their files. Licences were checked against each project’s repository on 2026-11-17. Evaluated and not integrated: Parca Agent with parcagpu (both Apache-2.0) for continuous eBPF and CUPTI profiling, because it needs a privileged node DaemonSet and an injected CUDA library; and Zymtrace, a proprietary continuous GPU profiler, which stays an integration option. InferCrane does not fork or embed vLLM, vLLM Router, SGLang, or SkyPilot. The ROCm runtime images bundle ROCm and AMD’s AITER kernels (MIT) under their own licenses; InferCrane does not build or redistribute them. Release automation must pin external tool versions and audit module/container transitive licenses and vulnerabilities. The Python SDK has no runtime dependency. The TypeScript SDK has no runtime dependency; TypeScript and Node type declarations are development-only build dependencies. No AGPL, SSPL, BSL, Commons Clause, or other source-available dependency is approved for the core. The current Terraform CLI is a separately executed, development-only qualification tool; InferCrane does not redistribute it. Operators may use an OpenTofu-compatible workflow after that integration has been independently qualified. OpenRouter and generic OpenAI-compatible APIs are remote integrations, not linked dependencies or resold services. Their terms, data handling, and charges remain between the operator and provider. LiteLLM is connected through its operator-managed OpenAI-compatible endpoint; InferCrane does not fork, embed, install, or redistribute it. Its repository licenses content outside the separately identified enterprise area under MIT; any future managed-process integration must pin and audit the exact distributed paths rather than treating the whole repository as uniformly licensed. E2B, Modal, Kubernetes sandbox implementations, MLflow, Kubeflow, SkyPilot training pipelines, and similar systems are external composition owners rather than InferCrane dependencies. Their licenses, credentials, isolation, data handling, execution, and charges remain with the operator and the chosen system. Candidate integration licenses and ownership decisions are tracked in Integration ownership and license matrix. Listing a candidate there does not register or qualify it as a supported dependency.