HuggingFaceConfig
Options & Configuration · AiDotNet.ModelLoading.Pretrained
A strongly-typed view over a Hugging Face decoder config.json (the LLaMA / Mistral /
Qwen2 text-model family). Only the fields needed to reconstruct the architecture are read;
unknown fields are ignored so newer configs still parse.
For Beginners
Every Hugging Face model ships a small config.json that
describes its shape — how wide it is, how many layers, how many attention heads, and so on.
This class turns that JSON into typed properties AiDotNet can build a matching model from.
Constructors
| Constructor | Summary |
|---|---|
HuggingFaceConfig() |
Properties
| Property | Summary |
|---|---|
IReadOnlyList<string> Architectures { get; set; } |
The declared architecture classes, e.g. |
bool TieWordEmbeddings { get; set; } |
Whether the LM head shares weights with the input embedding. |
double RmsNormEps { get; set; } |
RMSNorm epsilon. |
double RopeTheta { get; set; } |
RoPE base frequency (rope_theta). |
double? AttnLogitSoftcapping { get; set; } |
Attention-logit soft-cap magnitude (Gemma-2 attn_logit_softcapping); null when absent. |
double? FinalLogitSoftcapping { get; set; } |
Final-logit soft-cap magnitude (Gemma-2 final_logit_softcapping); null when absent. |
double? LogitScale { get; set; } |
Logit multiplier (Cohere logit_scale); null when absent. |
int HeadDim { get; set; } |
Per-head dimension. |
int HiddenSize { get; set; } |
Model (residual stream) width — the embedding / hidden dimension. |
int IntermediateSize { get; set; } |
Feed-forward inner dimension (the SwiGLU gate/up width). |
int MaxPositionEmbeddings { get; set; } |
Maximum position (context length) the RoPE table was trained for. |
int MoeIntermediateSize { get; set; } |
Routed-expert inner dimension (moe_intermediate_size); 0 when not an MoE decoder that distinguishes it from IntermediateSize. |
int NumAttentionHeads { get; set; } |
Number of query attention heads. |
int NumExpertsPerTok { get; set; } |
Experts activated per token (num_experts_per_tok); 0 for a dense decoder. |
int NumHiddenLayers { get; set; } |
Number of stacked decoder blocks. |
int NumKeyValueHeads { get; set; } |
Number of key/value heads. |
int NumLocalExperts { get; set; } |
Number of MoE experts (num_local_experts or num_experts); 0 for a dense decoder. |
int SharedExpertIntermediateSize { get; set; } |
Shared-expert inner dimension (Qwen2-MoE shared_expert_intermediate_size); 0 when absent. |
int VocabSize { get; set; } |
Vocabulary size (embedding rows and LM-head columns). |
int? BosTokenId { get; set; } |
Beginning-of-sequence token id, when declared. |
int? EosTokenId { get; set; } |
End-of-sequence token id, when declared. |
string HiddenActivation { get; set; } |
Feed-forward activation name, e.g. |
string ModelType { get; set; } |
The short model type, e.g. |
string TorchDtype { get; set; } |
On-disk weight precision, e.g. |
Methods
| Method | Summary |
|---|---|
HuggingFaceConfig FromFile(string path) |
Reads and parses a config.json file from disk. |
HuggingFaceConfig Parse(string json) |
Parses a config.json from its JSON text. |