Parameters

Every preset in Aspose.LLM for .NET exposes nine independent parameter bags — small POCO classes that group related settings. Tune any bag before calling AsposeLLMApi.Create(preset) to change how the engine loads and runs the model.

This section covers each bag in detail: the public fields, their types and defaults, and the typical reasons to change them.

At a glance

Bag	What it controls	Frequently changed for
Model source	Where the model file comes from	Local files, Hugging Face repos, custom caches
Model inference	How the model loads into memory	GPU offload, tensor split, memory mapping
Context	Runtime context and batch sizes	Context length, flash attention, KV cache quantization
Chat	Session-level conversation settings	System prompt, max tokens, cache cleanup
Sampler	Token sampling behavior	Temperature, penalties, DRY, Mirostat
Engine	Engine-wide defaults	Cache path, debug logging, default threads
Binary manager	Native binary deployment	Release tag, preferred acceleration backend
Multimodal context	mtmd (vision) context	GPU offload for projector, verbosity

PresetCoreBase exposes nine properties, but only eight parameter class types — the MmprojSourceParameters property for the vision projector uses the same ModelSourceParameters type as BaseModelSourceParameters. Both are covered on the Model source page.

General guidance

• Start with a built-in preset. Presets ship with carefully chosen defaults for each model family. Override only fields you have a concrete reason to change.
• Every field is a public get/set property. Mutations on the preset before Create take effect; after Create, the engine has already read the preset and further changes are ignored.
• Every bag is lazy-initialized on PresetCoreBase. Accessing any property gives you an instance even if you never set it — safe to write preset.ContextParameters.ContextSize = 8192 without a null check.
• Null means “use native default” on nullable fields. Set an explicit value only when you want to override.

Sections

• Model source — model file resolution order and cache layout.
• Model inference — model-load knobs (GPU, memory, tensor split).
• Context — llama.cpp context parameters.
• Chat — session-level conversation settings.
• Sampler — sampler knobs.
• Engine — engine-wide defaults.
• Binary manager — native binary deployment.
• Multimodal context — mtmd context for vision.

What’s next

• Presets — how PresetCoreBase composes the eight bags.
• Custom preset — override patterns for the most common scenarios.
• Supported presets — built-in presets with their default parameter values.