Struct Qwen35Config 

pub struct Qwen35Config {

    pub vocab_size: usize,
    pub hidden_size: usize,
    pub intermediate_size: usize,
    pub num_hidden_layers: usize,
    pub nextn_predict_layers: usize,
    pub num_attention_heads: usize,
    pub num_key_value_heads: usize,
    pub key_length: usize,
    pub value_length: usize,
    pub max_position_embeddings: usize,
    pub rms_norm_eps: f64,
    pub rope_theta: f64,
    pub rope_dim_count: usize,
    pub rope_dim_sections: Vec<usize>,
    pub full_attention_interval: usize,
    pub ssm_conv_kernel: usize,
    pub ssm_group_count: usize,
    pub ssm_inner_size: usize,
    pub ssm_state_size: usize,
    pub ssm_time_step_rank: usize,
    pub tie_word_embeddings: bool,
    pub num_experts: usize,
    pub num_experts_used: usize,
    pub expert_ffn_size: usize,
    pub shared_expert_ffn_size: usize,
    pub expert_weights_scale: f32,
}

Qwen3.5 model config — fields covering both the per-layer Mamba+ Attention block and the MTP head.

Fields

vocab_size: usize

hidden_size: usize

intermediate_size: usize

num_hidden_layers: usize

Total layer count (= main layers + nextn_predict_layers MTP heads).

nextn_predict_layers: usize

Layers at index < num_hidden_layers - nextn_predict_layers use the hybrid Mamba+Attention block. The remaining nextn_predict_layers layers use standard attention for MTP.

num_attention_heads: usize

num_key_value_heads: usize

key_length: usize

Per-head Q dim. The MTP attention head uses this.

value_length: usize

Per-head V dim.

max_position_embeddings: usize

rms_norm_eps: f64

rope_theta: f64

rope_dim_count: usize

rope_dim_sections: Vec<usize>

full_attention_interval: usize

Some Qwen3.5 layers do full attention every N blocks (interspersed with the Mamba-style blocks). Read but not yet acted on.

ssm_conv_kernel: usize

ssm_group_count: usize

ssm_inner_size: usize

ssm_state_size: usize

ssm_time_step_rank: usize

tie_word_embeddings: bool

num_experts: usize

MoE (qwen35moe): routed expert count. Zero for dense models.

num_experts_used: usize

Top-k experts activated per token.

expert_ffn_size: usize

Per-expert FFN inner dim (qwen35.expert_feed_forward_length).

shared_expert_ffn_size: usize

Shared-expert FFN inner dim (qwen35.expert_shared_feed_forward_length).

expert_weights_scale: f32

Router weight multiplier applied after softmax (llama.cpp default 1.0).

Implementations

impl Qwen35Config

pub fn from_gguf(raw: &GgufFile) -> Result<Qwen35Config, Error>

Read from a GGUF file with general.architecture in {qwen35, qwen35moe, qwen36, qwen36moe}. Qwen3.6 reuses the Qwen3.5 trunk; only the metadata-key prefix differs (qwen36.* vs qwen35.*). Returns an error when any required key is missing.

pub fn is_moe(&self) -> bool

True when the GGUF arch is qwen35moe.

pub fn expert_ffn_dim(&self) -> usize

Routed-expert SwiGLU inner width.

pub fn shared_expert_ffn_dim(&self) -> usize

Shared-expert SwiGLU inner width.

pub fn mtp_layer_start(&self) -> Option<usize>

Index of the first MTP layer (= num_hidden_layers - nextn_predict_layers). Returns None when the file has no MTP heads.

pub fn from_hf_config_json(path: &Path) -> Result<Qwen35Config, Error>

Read from a HuggingFace config.json (the safetensors distribution counterpart to GGUF metadata). Maps the standard HF Qwen2.x / Qwen3.x field names; unknown fields fall back to the same defaults used by Self::from_gguf. PLAN.md M1 safetensors load path (HauhauCS catalog rows).

Trait Implementations

impl Clone for Qwen35Config

fn clone(&self) -> Qwen35Config

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for Qwen35Config

fn fmt(&self, f: &mut Formatter<'_>) -> Result<(), Error>

Formats the value using the given formatter.