Presets

A preset bundles every parameter needed to configure a language model for inference: where to find the model file, how to load it, how to generate text, and — for vision presets — how to process images. Pass a preset to AsposeLLMApi.Create(preset) and the engine applies the whole configuration at once.

Base type

All presets derive from PresetCoreBase in namespace Aspose.LLM.Abstractions.Parameters.Presets. The base type exposes nine parameter bags, each an independently overridable POCO:

Property Type Purpose
BaseModelSourceParameters ModelSourceParameters Where to load the main model (local path, Aspose model ID, or Hugging Face repo and file name).
MmprojSourceParameters ModelSourceParameters Optional vision projector (mmproj). Set by vision presets; empty on text presets.
BinaryManagerParameters BinaryManagerParameters Native llama.cpp release tag, binary cache path, and preferred acceleration backend.
EngineParameters EngineParameters Engine-wide settings: model cache path, debug logging, log directory, default threads.
ChatParameters ChatParameters System prompt, optional history, max tokens, cache cleanup strategy.
ContextParameters ContextParameters llama.cpp context knobs: context size, batch sizes, rope scaling, flash attention, KV cache dtype.
SamplerParameters SamplerParameters Sampler knobs: temperature, top-p, top-k, min-p, penalties, DRY, XTC, mirostat, seed, logit bias.
BaseModelInferenceParameters ModelInferenceParameters Model-load knobs: GPU layers, main GPU, split mode, tensor split, memory mapping, KV overrides.
MtmdContextParameters MultimodalContextParameters mtmd (vision) context: GPU use, timings, thread count, verbosity, media marker. Set by vision presets.

Every bag is lazy-initialized — accessing any property is safe even if you did not set it explicitly.

Create the API with a preset

using Aspose.LLM;
using Aspose.LLM.Abstractions.Parameters.Presets;

var preset = new Qwen25Preset();
using var api = AsposeLLMApi.Create(preset);

The engine reads every bag on the preset during construction. Mutations to the preset after Create returns do not affect the alive engine — set everything you need first.

Override preset parameters before Create

You can tweak any field on any bag before passing the preset to Create.

var preset = new Qwen25Preset();

preset.ContextParameters.ContextSize = 8192;            // smaller context = less memory
preset.SamplerParameters.Temperature = 0.2f;            // more deterministic output
preset.BaseModelInferenceParameters.GpuLayers = 0;      // force CPU-only inference

using var api = AsposeLLMApi.Create(preset);

Common overrides:

  • ContextParameters.ContextSize — reduce to save memory or raise to fit long conversations.
  • SamplerParameters.Temperature — lower for deterministic tasks, higher for creative output.
  • BaseModelInferenceParameters.GpuLayers — control GPU offload (0 = CPU only; 999 = full offload).
  • BinaryManagerParameters.PreferredAcceleration — force CUDA, Metal, Vulkan, HIP, or CPU.
  • ChatParameters.SystemPrompt — set a default system prompt for new sessions.
  • ChatParameters.MaxTokens — cap the length of a single response.

See the custom preset use case for longer examples.

Extend a built-in preset

Create a subclass that sets your customizations in the constructor. This is the cleanest approach when you use the same tuning in several places.

using Aspose.LLM.Abstractions.Parameters.Presets;

public class LowTemperatureQwen : Qwen25Preset
{
    public LowTemperatureQwen()
    {
        ContextParameters.ContextSize = 16384;
        SamplerParameters.Temperature = 0.1f;
        ChatParameters.SystemPrompt = "You are a precise assistant. Answer briefly.";
    }
}

// Use it
using var api = AsposeLLMApi.Create(new LowTemperatureQwen());

Create a preset from scratch

For a fully custom model, extend PresetCoreBase directly and populate the bags you need. At minimum, you must set BaseModelSourceParameters so the engine can find the model.

using Aspose.LLM.Abstractions.Parameters.Presets;

public class MyCustomPreset : PresetCoreBase
{
    public MyCustomPreset()
    {
        BaseModelSourceParameters.HuggingFaceRepoId = "your-org/your-model-GGUF";
        BaseModelSourceParameters.HuggingFaceFileName = "your-model.Q4_K_M.gguf";

        ContextParameters.ContextSize = 4096;
        ContextParameters.NBatch = 1024;

        SamplerParameters.Temperature = 0.7f;
        SamplerParameters.TopP = 0.9f;

        BaseModelInferenceParameters.GpuLayers = 999;
    }
}

See the custom preset use case for a full walkthrough including chat template selection for non-standard models.

Default preset

AsposeLLMApi.GetDefaultPreset() returns a fresh Qwen25Preset instance. It is a convenience helper for code that needs a sensible preset when none has been chosen yet — not a snapshot of the engine’s internal defaults.

PresetCoreBase defaultPreset = api.GetDefaultPreset();
// Equivalent to: new Qwen25Preset()

For raw parameter values without a preset, call:

(ModelInferenceParameters inference,
 ContextParameters context,
 ChatParameters chat,
 SamplerParameters sampler) = await api.GetDefaultParametersAsync();

Supported presets

See Supported presets for the full list of built-in text and vision presets with their Hugging Face model sources, context sizes, and quantization levels.

What’s next

AsposeLLMApi facade Parameters