Parse bounded container metadata and tensor descriptors without starting inference.
Goal
Parse bounded container metadata and tensor descriptors without starting inference.
Applicability
Use this example as a host-architecture reference for the named framework. Resolve and pin dependencies through the validation solution or the consuming repository policy.
Required packages
- UAIX.LmRuntime.Gguf
Host responsibilities
- Acquire and approve the model outside request or generation paths.
- Prepare the model-specific prompt and enforce limits.
- Own cancellation, UI dispatch, error presentation, persistence, and evidence retention.
Complete code
using System;
using UAIX.LmRuntime.Gguf;
namespace GgufRuntime.Examples.MetadataInspector;
/// <summary>Prints bounded GGUF container and tensor metadata without starting inference.</summary>
public static class Program
{
/// <summary>Parses and prints one GGUF artifact.</summary>
/// <param name="arguments">The GGUF path.</param>
/// <returns>Zero when parsing succeeds; two when the path is missing.</returns>
public static Int32 Main(String[] arguments)
{
if (arguments.Length != 1)
{
Console.Error.WriteLine("Usage: MetadataInspector <model.gguf>");
return 2;
}
GgufModel model = GgufReader.Read(arguments[0], new GgufParseOptions
{
MaxTensorCount = 100_000,
MaxMetadataCount = 100_000,
MaxDimensionCount = 8,
MaxStringBytes = 16 * 1024 * 1024,
MaxArrayLength = 10_000_000,
MaxArrayDepth = 8
});
Console.WriteLine($"Artifact kind: {model.ArtifactKind}");
Console.WriteLine($"GGUF version: {model.Version}");
Console.WriteLine($"Tensor data offset: {model.TensorDataOffset}");
Console.WriteLine($"Alignment: {model.Alignment}");
if (model.TryGetString("general.name", out String modelName))
{
Console.WriteLine($"Model name: {modelName}");
}
foreach (GgufTensorDescriptor tensor in model.Tensors)
{
Console.WriteLine(
$"{tensor.Name}: {tensor.GgmlType}, " +
$"{tensor.ElementCount:N0} elements, {tensor.ByteLength:N0} bytes");
}
return 0;
}
}Expected behavior
The host either produces the narrow result described by the example or surfaces a specific identity, validation, load, generation, cancellation, or disposal failure. No hidden fallback is introduced.
Failure cases
- Invalid or untrusted input identity.
- GGUF structure, architecture, tokenizer, storage, kernel, or memory incompatibility.
- Cancellation or host shutdown.
- Framework dispatch or lifecycle misuse.
Security and trust notes
Do not accept remote model uploads, arbitrary paths, arbitrary URLs, provider keys, commands, or private-network targets. The loopback example binds to 127.0.0.1 and still requires a new threat model before broader exposure.
Disposal requirements
The host must finish or cancel generation, dispose the session, then dispose the model. Service and UI examples must connect this order to application shutdown.
Evidence produced
- Resolved package identity and version from the build.
- Model hash and byte count from intake.
- Session identifier, bounds, cancellation state, selected path, result, timing method, and limitations when the host records them.
Build-validation status
State: Experimental. Last compilation attempt UTC: 2026-06-25T00:00:00Z. The source was API-reviewed against current canonical public documentation; local compilation is blocked until a .NET SDK and package restore are available. CI is configured to produce the retained build result.
Open canonical package documentation.
Operational boundary
What this page covers: Parse bounded container metadata and tensor descriptors without starting inference.
Available now: The complete source and validation-project mapping are included in this release.
Host responsibility: Provide real inputs, run restore/build/tests, test the host lifecycle, and retain exact evidence.
Not claimed: A successful local build in this environment, finished downloadable application, universal compatibility, or production/safety certification.
Canonical sources: Canonical package guide · NuGet
Last reviewed UTC: