LMRuntime.com / Public page
Overview
A concise explanation of what LMRuntime is, where it fits, and what it does not claim.
UAIX.LmRuntime is a public .NET package family for bounded GGUF intake, tokenizer behavior, tensor semantics, managed CPU kernels, LLaMA-family execution, sampling, backend selection contracts, and application-facing local sessions.
Inspectable local inference
Local inference is easier to reason about when container parsing, tokenizer selection, tensor storage, model binding, mathematical execution, token selection, backend discovery, and host orchestration remain separate. UAIX.LmRuntime exposes those layers as focused packages instead of hiding them behind a remote service or silent network fallback.
Seventeen packages, explicit responsibilities
Abstractions, Tensors, and Acceleration define contracts, storage semantics, and backend control-plane types.Gguf, Tokenization, Sampling, Kernels.Cpu, and Models.Llama own container-to-token execution responsibilities.CpuManaged, Cuda, DirectML, Vulkan, Rocm, and Metal register backend identities and capability surfaces with explicit probes.LocalEndpoint owns the verified managed CPU application path.Open the full package selection guide →
The application-facing path
The host supplies a prepared prompt and trusted model identity. LocalEndpoint verifies and maps the artifact, validates supported metadata, creates model and session state, produces logits, applies explicit token selection, and returns output and observable generation state.
Backend registration is not execution proof
Acceleration separates registration, declared capabilities, runtime-identifier compatibility, probing, selection policy, and fallback evidence. Backends.CpuManaged reports available without native assets. GPU registration packages fail closed until a host-provided adapter proves the local runtime, assets, driver, and device.
Where it fits
- .NET applications that require in-process local inference with no provider API or model downloader in the runtime path.
- Tools that inspect GGUF metadata, tensors, tokenizer configuration, and backend diagnostics.
- Runtime engineering and test suites that need a legible reference path and parity evidence.
- Desktop or worker integrations that keep prompt assembly, persistence, network, tools, commands, and UI in the host application.
What publication does not imply
A published package is a consumable artifact, not a blanket guarantee for every model, quantization, operating system, processor, GPU, driver, context length, or performance target. Package pages document exact responsibilities and known boundaries; the capability matrix separates API presence from executed evidence.