LMRuntime.com / Public page

Architecture

The layered package and execution architecture from trusted local model identity through GGUF, tokenization, tensors, CPU kernels, LLaMA sessions, sampling, and the application facade.

UAIX.LmRuntime separates the model pipeline, backend-selection control plane, backend registration packages, native-asset identities, and application facade. Install the package that owns the exact layer your host must control.

Package dependency map

Foundations

AbstractionsTensorsAcceleration

Model pipeline

GgufSamplingKernels.CpuTokenizationModels.Llama

Backend branch

CpuManagedCudaDirectMLVulkanRocmMetal

Application facade

LocalEndpoint

The diagram shows responsibility lanes. It does not imply that every package depends on every item above it. Exact direct dependencies appear in the table below, every package guide, and each NuGet page.

Layer ownership

Package	Required For	Direct dependencies
`Abstractions` NuGet ↗	runtime-neutral contracts	None in the package family
`Tensors` NuGet ↗	tensor layout and storage metadata	None in the package family
`Acceleration` NuGet ↗	explicit backend registration, probing, selection, and fallback evidence	None in the package family
`Gguf` NuGet ↗	GGUF inspection and validation	Tensors
`Sampling` NuGet ↗	token selection and stop handling	Abstractions
`Kernels.Cpu` NuGet ↗	managed CPU math	Tensors
`Tokenization` NuGet ↗	tokenizer and chat-template work	Abstractions, Gguf
`Models.Llama` NuGet ↗	LLaMA graph/session internals	Abstractions, Gguf, Kernels.Cpu, Sampling, Tensors, Tokenization
`Backends.CpuManaged` NuGet ↗	the package-visible managed CPU backend and explicit CPU fallback identity	Acceleration
`Backends.Cuda` NuGet ↗	CUDA backend registration and fail-closed CUDA diagnostics	Acceleration
`Backends.DirectML` NuGet ↗	DirectML backend registration and fail-closed DirectML diagnostics	Acceleration
`Backends.Vulkan` NuGet ↗	Vulkan backend registration and fail-closed Vulkan diagnostics	Acceleration
`Backends.Rocm` NuGet ↗	ROCm backend registration and fail-closed ROCm diagnostics	Acceleration
`Backends.Metal` NuGet ↗	Metal backend registration and fail-closed Metal diagnostics	Acceleration
`Backends.Cuda.Native.win-x64` NuGet ↗	the Windows x64 modern CUDA native-asset package slot	None in the package family
`Backends.Cuda.LegacyK80.win-x64` NuGet ↗	the separate Windows x64 Tesla K80 compute-capability 3.7 native-asset package slot	None in the package family
`LocalEndpoint` NuGet ↗	application integration	Acceleration, Backends.CpuManaged, Models.Llama, Tokenization

Backend lifecycle

Register: the host explicitly adds backend instances to RuntimeBackendRegistry.
Declare: each backend exposes identity, execution kind, supported runtime identifiers, native-asset state, and capability flags.
Probe: the backend checks local evidence and returns devices, diagnostics, and availability.
Select: RuntimeBackendSelector applies an explicit require/prefer policy and records CPU fallback.
Execute: only a separately proven execution adapter may perform model work. Registration and declared compatibility are not substitutes.

Current distinctionBackends.CpuManaged reports available without native assets. The GPU registration packages expose fail-closed diagnostics until an executable adapter is proven by the host.

Managed model load path

Canonical pathSHA-256 + byte countGGUF structureTokenizer metadataLLaMA tensor bindingManaged model

LocalEndpoint performs the application-facing path. It constrains file identity before parsing, loads the bounded GGUF catalog, creates tokenizer state, validates model configuration, binds mapped weights, and exposes a disposable model scope.

Generation loop

Encode a host-prepared prompt under explicit special-token policy.
Evaluate prompt tokens into model and KV-cache state.
Produce logits for the next token position.
Select a token under explicit sampling state; LocalEndpoint uses deterministic greedy selection.
Decode safe UTF-8 output and evaluate token/text stop conditions.
Commit position and observable generation state, then repeat within the request bound.

Resource ownership

Runtime scopeOwns loaded models and the managed backend identity selected for the facade path.

ModelOwns mapped model bytes, parsed configuration, tokenizer, bound weights, and session creation.

SessionOwns position, KV-cache state, generation continuity, and immutable UAIX context evidence.

HostOwns files, hashes, licensing, prompt construction, persistence, network, tools, commands, UI, audit, cancellation, and deployment review.

Roadmap integration point

Host layer onlyThe Evidence-Gated Adaptive Runtime Control roadmap sits above this package graph. It observes public package state and may invoke only pre-authorized host actions; it does not become a hidden package dependency or alter model weights, tokenizer behavior, tensor values, or backend proof rules.

Host boundary

The packages do not download models, host an HTTP server, call a provider API, install GPU drivers, grant authority through memory, or silently change execution lanes. Hosts must retain explicit configuration, package locks, local probe results, model identities, and application-specific compatibility evidence.