LMRuntime.com / Package family

UAIX.LmRuntime Packages

Choose the UAIX.LmRuntime package required for each task, inspect dependency boundaries, and open version-free installation guides.

UAIX.LmRuntime is a public family of 17 NuGet packages. LocalEndpoint owns the managed CPU application path. Acceleration and backend packages make selection, compatibility declarations, and failure diagnostics explicit instead of silently changing execution.

Browse the package family on NuGet Run the managed CPU quick start

Choose The Package Required For Your Task

Required For	Package	Owned layer
runtime-neutral contracts	`Abstractions` NuGet ↗	Stable runtime contracts, request and response models, diagnostics, and governance interfaces.
tensor layout and storage metadata	`Tensors` NuGet ↗	Tensor shapes, data types, GGML storage traits, quantized-block metadata, and reference vector math.
explicit backend registration, probing, selection, and fallback evidence	`Acceleration` NuGet ↗	Explicit backend contracts, registration, capability declarations, local probing, deterministic selection, and visible CPU fallback.
GGUF inspection and validation	`Gguf` NuGet ↗	Bounded GGUF parsing, metadata and tensor catalogs, strict validation, sharding, hashing, and mapped tensor access.
token selection and stop handling	`Sampling` NuGet ↗	Greedy and probability sampling, logit transforms, deterministic random state, top-k/top-p filtering, stop matching, and generation control.
managed CPU math	`Kernels.Cpu` NuGet ↗	Reference, portable-vector, AVX2-aware, half-precision, and quantized CPU kernels with explicit dispatch and parity checks.
tokenizer and chat-template work	`Tokenization` NuGet ↗	GGUF tokenizer metadata, tokenizer factories and engines, special-token handling, chat templates, truncation, safety, and parity tools.
LLaMA graph/session internals	`Models.Llama` NuGet ↗	LLaMA-family configuration, tensor binding, mapped weight sources, reference forward execution, sessions, KV cache, generation, persistence, and parity evidence.
the package-visible managed CPU backend and explicit CPU fallback identity	`Backends.CpuManaged` NuGet ↗	The explicit managed .NET CPU backend registration, capability declaration, device identity, and no-native-asset compatibility lane.
CUDA backend registration and fail-closed CUDA diagnostics	`Backends.Cuda` NuGet ↗	CUDA backend registration, declared compatibility metadata, runtime identifiers, and fail-closed local probe diagnostics.
DirectML backend registration and fail-closed DirectML diagnostics	`Backends.DirectML` NuGet ↗	DirectML backend registration, declared compatibility metadata, runtime identifiers, and fail-closed local probe diagnostics.
Vulkan backend registration and fail-closed Vulkan diagnostics	`Backends.Vulkan` NuGet ↗	Vulkan backend registration, declared compatibility metadata, runtime identifiers, and fail-closed local probe diagnostics.
ROCm backend registration and fail-closed ROCm diagnostics	`Backends.Rocm` NuGet ↗	ROCm backend registration, declared compatibility metadata, runtime identifiers, and fail-closed local probe diagnostics.
Metal backend registration and fail-closed Metal diagnostics	`Backends.Metal` NuGet ↗	Metal backend registration, declared compatibility metadata, runtime identifiers, and fail-closed local probe diagnostics.
the Windows x64 modern CUDA native-asset package slot	`Backends.Cuda.Native.win-x64` NuGet ↗	RID-specific NuGet package slot for modern Windows x64 CUDA native assets under runtimes/win-x64/native/.
the separate Windows x64 Tesla K80 compute-capability 3.7 native-asset package slot	`Backends.Cuda.LegacyK80.win-x64` NuGet ↗	RID-specific NuGet package slot for Tesla K80 legacy CUDA native assets under runtimes/win-x64/native/.
application integration	`LocalEndpoint` NuGet ↗	The high-level local-only GGUF facade: verified files, bounded loading, isolated sessions, UAIX context evidence, and deterministic greedy generation.

Package dependency map

Foundations

AbstractionsTensorsAcceleration

Model pipeline

GgufSamplingKernels.CpuTokenizationModels.Llama

Backend registrations

CpuManagedCudaDirectMLVulkanRocmMetal

CUDA native-asset identities

Native.win-x64LegacyK80.win-x64

Managed application path

LocalEndpoint

Exact direct dependencies are listed on every guide and NuGet page. The model pipeline and backend-selection branch are separate concerns: registration and capability metadata do not prove an executable GPU adapter.

Backend packages: what installation means

Works Today

Acceleration registry, probe, policy, and selection contracts
Managed CPU backend registration and successful local probe
CUDA, DirectML, Vulkan, ROCm, and Metal capability declarations
Fail-closed diagnostics when an execution adapter is not proven
Separate CUDA native-asset package identities

Not Claimed

GPU inference merely because a registration package is referenced
A native CUDA inference binary inside the supplied native-asset slots
Automatic driver installation or model download
Silent replacement of LocalEndpoint managed execution
Hardware compatibility without a successful local probe

Package catalog

Application facade

Required For application integration

UAIX.LmRuntime.LocalEndpoint

LocalEndpoint

The high-level local-only GGUF facade: verified files, bounded loading, isolated sessions, UAIX context evidence, and deterministic greedy generation.