Struct MiniMaxH3 

pub struct MiniMaxH3 {
    pub hidden: usize,
    pub heads: usize,
    pub head_dim: usize,
    pub ffn: usize,
    pub latents_dim: usize,
    pub audio_dim: usize,
    pub text_dim: usize,
    pub shift_video: f64,
    pub shift_audio: f64,
    pub cond_aug: f64,
    pub cond_aug_audio: f64,
    /* private fields */
}

Fields

hidden: usize

heads: usize

head_dim: usize

ffn: usize

latents_dim: usize

audio_dim: usize

text_dim: usize

shift_video: f64

shift_audio: f64

cond_aug: f64

How much of a keyframe latent survives the noise blend, and therefore where its rows sit on the schedule. VISUAL_COND_ TIMESTEP is the reference’s default; 1.0 turns the blend off.

cond_aug_audio: f64

The same, for a reference soundtrack. The reference’s default is 1.0 — an audio condition is not noised at all.

Implementations

impl MiniMaxH3

pub fn from_cmf(model: &Arc<CmfModel>) -> Result<Self, String>

pub fn from_cmf_lora( model: &Arc<CmfModel>, bank: Option<&LoraBank>, ) -> Result<Self, String>

The same, with an adapter consulted for every projection.

The base weights are q4tp and stay untouched: a rank-32 update cannot be folded into a four-bit ladder without dequantizing the whole DiT, so the branch rides beside the base GEMM the way crate::ltxlora does it for LTX.

pub fn lora_bound(&self) -> usize

How many of the adapter’s branches found a projection.

pub fn lora_binds(&self, key: &str) -> bool

Whether this adapter key landed on a projection of ours.

pub fn lora_report(&self) -> Option<String>

Under CMF_LORA_PROBE=1: every branch by measured contribution, loudest first, and which ones the router switched off. This is the map that says where an adapter actually lives — for the Realism adapter it is not uniform across the fifty blocks.

pub fn refine_text(&self, states: &[f32], n: usize) -> Vec<f32>

Qwen3-VL states [n, text_dim] → refined text embeds [n, hidden]. Prompt-only, so the caller does this once per generation rather than once per step.

pub fn forward( &self, layout: &Layout, text: &[f32], video: &[f32], audio: &[f32], sigma_v: f64, cond: &[Vec<f32>], ) -> (Vec<f32>, Vec<f32>)

One denoise evaluation.

video is [latents_dim, latent_t, lat_h, lat_w] and audio is [audio_dim, 2, audio_t], both in the reference’s channel-major order. text is the refined [n, hidden] stream. Returns the two velocities, EACH ON ITS OWN SCHEDULE and unscaled.