Aperture inspects your hardware with permission, reads the model link or local files you choose, and explains a memory-aware configuration. For compatible GGUF models, it can start a local chat or bounded experiment using an existing inference runtime, including CPU/GPU split execution when the checkpoint exceeds available VRAM. Aperture can also compare the machine it actually observes with source-pinned inference recipes without downloading or running the referenced models.
Website · Get started · Verified support · Release 0.4.7 · MIT license
Install Node.js LTS first. Requires Node.js 20.11 or newer; Node 22 and 24 have release control coverage. The package command does not require Git, an npm account, or a cloud-model subscription.
Windows PowerShell
npx.cmd --yes --package=https://github.com/BigBirdReturns/aperture/releases/download/v0.4.7/bigbirdreturns-aperture-0.4.7.tgz apertureLinux / macOS terminal
npx --yes --package=https://github.com/BigBirdReturns/aperture/releases/download/v0.4.7/bigbirdreturns-aperture-0.4.7.tgz apertureWindows and Linux have recorded native runs; native Mac inference remains unverified. With Git installed, use npx --yes github:BigBirdReturns/aperture#v0.4.7 instead. Distribution is through GitHub, not the npm registry namespace.
Approve the read-only scan, paste your model link or drag in a local file, and choose the exact representation when prompted. Review the configuration, then start a local chat, save it for later, generate one answer, or opt into an experiment. Metadata access, weight downloads, native-runtime installation, and execution have separate permissions.
Use /new for a fresh chat and /exit to release the model. Saved configurations are listed with aperture list and reopened with aperture chat ANSWER.json; retain the npx --package=… aperture prefix when the executable is not installed globally.
When a saved remote GGUF already exists as a complete exact artifact in Aperture's managed cache, the 0.4.6 path binds that artifact locally before asking for model-host or weight-transfer approval. Aperture still hashes every selected component and repeats native fit before loading. Missing, changed, symbolic, differently identified, or unsupported cache entries remain on the existing permissioned path. Safetensors are not admitted through this shortcut.
Aperture 0.4.7 adds a source-pinned recipe resolver:
aperture recipes --allow-scanSave the complete local report with:
aperture recipes --allow-scan --out aperture-recipes.jsonThe first catalog is derived from factual observations in local-inference-lab/rtx6kpro, pinned to an exact upstream commit. Aperture does not vendor or execute that project's prose, scripts, Dockerfiles, or other executable material. It converts selected observations into structured compatibility records for Qwen3.8, Qwen3.5, MiniMax, GLM-5, and Kimi configurations.
Every result is classified as QUALIFIED, CANDIDATE, UNKNOWN, or BLOCKED. A reference machine is never silently promoted into a minimum. If an experiment ran on a 96 GB-class GPU but did not establish a lower memory boundary, a 32 GiB card receives UNKNOWN, with the exact fit/load canary required to resolve it. Hard rejection is reserved for evidence such as too few GPUs or a known numerical requirement that the observed hardware cannot satisfy. Approximate working-set observations remain reference evidence rather than fabricated floors.
The recipe command performs only the permissioned hardware scan. It does not read model files, contact model hosts, download weights, install a runtime, execute inference, stop workloads, or run a benchmark. See Recipe Lab for the evidence and claim boundary.
Aperture retains the opt-in hardware receipt introduced in 0.4.5 for reproducing fit and backend failures without posting the raw scan:
aperture support --allow-scan --out aperture-support.jsonThe receipt retains CPU topology, RAM and allocation headroom, integrated and discrete graphics, NPU inventory, drivers, storage classes and capacities, available link observations, runtime presence, scan status, and explicit unmeasured fields. It excludes host and user names, paths and mount labels, stable device and partition identifiers, GPU UUIDs, drive product names and serials, network adapter names and addresses, model locations, prompts, generated text, credentials, and environment variables. It is never uploaded automatically.
The file is reduced, not anonymous. Hardware model names, exact capacities, driver versions, and its timestamp can still fingerprint a machine. Inspect the JSON before sharing it.
Before asking to acquire remote GGUF weights, Aperture obtains permission for the pinned runtime and assesses the complete selected tensor tables. Bounded prefixes are capped at 8 MiB per shard and 64 MiB total. A failed or unavailable assessment stops acquisition with the model, quantization, context, sequence count, and explicit device/layer requirements unchanged. Automatic placement checks layer counts for the highest estimated fitting count, not optimal speed.
After download and complete integrity hashing, generation and chat repeat the assessment against refreshed hardware immediately before loading. Each experiment trial rehashes and reassesses independently. A later refusal preserves earlier trial results. Native memory checks and actual GPU-layer/context readback remain enabled. Shared RAM and independent GPUs are not pooled.
These are pinned-runtime estimates, not architecture or throughput guarantees. Prefixes can include initial tensor bytes but do not acquire the full checkpoint. --answer-only remains a lightweight provisional answer without native installation or assessment. See VERIFICATION.md for native observations and retained limits.
Local GGUF files and complete shard sets, extensionless GGUF blobs, Hugging Face repository/file links, and direct HTTPS GGUF files are supported for inspection. Safetensors inspection and a Python/Accelerate compatibility route are also present, with separate runtime prerequisites. Model architecture support depends on the pinned runtime.
The managed GGUF path selects one compatible CUDA or Vulkan device, Apple Metal, or CPU. It preserves your selected artifact and explicit context. Independent devices and shared RAM are not pooled into fictitious capacity. NPU inventory is not NPU inference. The guided runner supports one sequence; automatic external-harness connections and distributed execution are not implemented. Recipe matching can reason about multi-GPU source configurations, but it does not itself implement distributed serving.
VERIFICATION.md records native runs separately from source/control tests. Release 0.4.4 established Qwen2.5-14B Q4_K_M generating 42 through a CPU/RTX 4060 split at 2,048-token context while the 8,982,142,976-byte tensor payload exceeded the device's reported physical memory. Qualified automatic runs selected 27 or 28 of 49 layers as admission-time headroom changed, and the system-headroom watchdog completed the exact plan that the former RSS proxy aborted.
Release 0.4.5 added the permissioned support-receipt path. Release 0.4.6 added exact managed-GGUF return use and separately verified a Windows x64 cache-resume passage through CUDA. Release 0.4.7 adds capability-gated recipe matching and additional NVIDIA compute-capability and current PCIe-link observations. The recipe result remains compatibility evidence only; it does not claim that an upstream throughput or model-quality result has been reproduced on the scanned machine.
npm test
node scripts/package-smoke.mjsNo telemetry or automatic result upload is enabled. Raw local run records can contain prompts, output, paths, and hardware identifiers. Use the reduced support command for an initial external report, then inspect the JSON before sharing it. See CONTRIBUTING.md and SECURITY.md for reporting and development.
Aperture is an independent MIT-licensed controller. Magnitude inspired the scan/select/setup experience. node-llama-cpp, llama.cpp, and Hugging Face Accelerate provide distinct runtime capabilities under their own licenses. The first Recipe Lab catalog cites factual observations from local-inference-lab/rtx6kpro without redistributing its prose or executable material. Model checkpoints keep their publishers' license and access conditions. No model weights are included here.