Class ChatPreset

Namespace

Qavren.Edge.Chat

Assembly

Qavren.Edge.Chat.Onnx.dll

One ready-made ORT GenAI model folder: its manifest, its geometry, its defaults.

public sealed record ChatPreset : IEquatable<ChatPreset>

Inheritance

object

ChatPreset

Implements

IEquatable<ChatPreset>

Inherited Members

object.Equals(object)

object.Equals(object, object)

object.GetHashCode()

object.GetType()

object.ReferenceEquals(object, object)

object.ToString()

Properties

DefaultMaxContextTokens

Never above ContextLength.

DefaultMaxOutputTokens

The per-turn output cap when nothing nearer the call site sets one.

DefaultTemperature

The sampling temperature when nothing nearer the call site sets one.

DefaultTopK

The top-k cut-off when nothing nearer the call site sets one.

DefaultTopP

The nucleus cut-off when nothing nearer the call site sets one.

DisplayName

What an about screen or a consent sheet shows.

GenAiConfigFile

The manifest-relative name of the GenAI configuration file.

Id

The id AddOnnxChat and ById(string) use.

LicenseUri

The licence URL an about screen needs; Manifest.SpdxLicense carries the identifier.

Not every chat model's licence has an SPDX identifier, and the field must not pretend otherwise. Sub-project 2 documents SpdxLicense as "the SPDX identifier of the model's licence, for an about screen", and the consent sheet and the repo's THIRD-PARTY-NOTICES.md both surface that string verbatim. The Llama 3.2 Community Licence is not on the SPDX list, so it takes SPDX's own convention for a licence that is not: LicenseRef-LLAMA-3.2-Community, never a bare LLAMA-3.2-Community, which is not a valid SPDX expression at all. Qwen3_600MInt4 is Apache-2.0, a real identifier. The field therefore holds an SPDX expression, of which a LicenseRef- value is a legal one, and the generator script emits that spelling.

Manifest

Sub-project 2's manifest, unchanged. GraphFile names the decoder .onnx, so the content-addressed directory is keyed on the weights. genai_config.json, tokenizer.json, tokenizer_config.json and chat_template.jinja ride as Auxiliary; model.onnx.data rides as GraphExternalData.

Shape

The decoder geometry the memory budget is computed from, before anything is mapped.

StopSequences

Matched against the accumulated decoded text, never against one decoded token.