Module common 

Common model components shared across text model architectures.

Re-exports

pub use chatml_history::*;

Modules

chatml_history

text_model

Structs

Cache

Abstraction over cosine and sine tables, kv-caching and attention masking.

CausalSelfAttention

Config

Generalized LLM configuration shared by all decoder-only text models.

LinearAttnConfig

Configuration for linear (recurrent) attention layers (e.g. Gated DeltaNet in Qwen3.5).

MLP

Multi-perceptron implementation with fused gate+up projection.

RopeScaling

RoPE scaling configuration for models with extended context (e.g. LLaMA 3.1+).

Transformer

Transformer block with causal self attention and several caching strategies.

Enums

EosTokenId

EOS token ID(s) — deserializes from either a single u32 or an array of u32.

Functions

detect_text_model_arch

Auto-detect text model architecture from config.json’s “architectures” field.

load_rms_norm

Load an RMS norm, optionally applying the residual weight pattern (1 + weight). When residual is true (Qwen3.5), the stored weight is treated as a residual and 1.0 is added.