Struct ModelPaths 

pub struct ModelPaths {

    pub transformer: PathBuf,
    pub transformer_shards: Vec<PathBuf>,
    pub low_noise_transformer: Option<PathBuf>,
    pub vae: PathBuf,
    pub spatial_upscaler: Option<PathBuf>,
    pub temporal_upscaler: Option<PathBuf>,
    pub distilled_lora: Option<PathBuf>,
    pub low_noise_distilled_lora: Option<PathBuf>,
    pub t5_encoder: Option<PathBuf>,
    pub clip_encoder: Option<PathBuf>,
    pub t5_tokenizer: Option<PathBuf>,
    pub clip_tokenizer: Option<PathBuf>,
    pub clip_encoder_2: Option<PathBuf>,
    pub clip_tokenizer_2: Option<PathBuf>,
    pub text_encoder_files: Vec<PathBuf>,
    pub text_tokenizer: Option<PathBuf>,
    pub decoder: Option<PathBuf>,
}

Resolved model file paths. For diffusion models, transformer and vae are always required. For upscaler models, only transformer (weights) is required; vae is empty. For utility models, only transformer is required; vae may be empty. Other paths are optional — each engine validates what it needs at load time.

Fields

transformer: PathBuf

transformer_shards: Vec<PathBuf>

Multi-shard transformer paths (Z-Image BF16); empty means use single transformer

low_noise_transformer: Option<PathBuf>

The low-noise expert of a two-expert checkpoint pair (Wan 2.2 A14B).

Some is the two-expert predicate. The high-noise expert lives in transformer because it runs first — the sampler starts at high sigma — so every consumer that only understands one transformer keeps working against the one that opens the generation.

vae: PathBuf

spatial_upscaler: Option<PathBuf>

temporal_upscaler: Option<PathBuf>

distilled_lora: Option<PathBuf>

low_noise_distilled_lora: Option<PathBuf>

The distilled adapter for low_noise_transformer. Each expert of an A14B pair is distilled separately; swapping them is not a degradation, it is the wrong model.

t5_encoder: Option<PathBuf>

clip_encoder: Option<PathBuf>

t5_tokenizer: Option<PathBuf>

clip_tokenizer: Option<PathBuf>

clip_encoder_2: Option<PathBuf>

CLIP-G / OpenCLIP encoder (SDXL only)

clip_tokenizer_2: Option<PathBuf>

CLIP-G / OpenCLIP tokenizer (SDXL only)

text_encoder_files: Vec<PathBuf>

Generic text encoder shard paths (Qwen3 for Z-Image)

text_tokenizer: Option<PathBuf>

Generic text encoder tokenizer (Qwen3 for Z-Image)

decoder: Option<PathBuf>

Stage B decoder weights (Wuerstchen only)

Implementations

impl ModelPaths

pub fn all_file_paths(&self) -> Vec<&Path>

Every resolved local artifact that can participate in engine creation. Policy fences consume this before inspecting or loading any artifact.

pub fn resolve(model_name: &str, config: &Config) -> Option<Self>

Resolve paths for a model. Checks config, then env vars. Returns None if transformer and VAE paths can’t be resolved. All other paths are optional (depend on model family).

pub fn resolve_from_model_config_exact(model_cfg: &ModelConfig) -> Option<Self>

Resolve only the paths captured in model_cfg.

This intentionally does not consult MOLD_*_PATH environment variables, manifests, sidecars, or model discovery. Durable chain jobs use it after admission so a restart cannot silently substitute different artifacts.

Trait Implementations

impl Clone for ModelPaths

fn clone(&self) -> ModelPaths

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for ModelPaths

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

Formats the value using the given formatter.

impl Eq for ModelPaths

impl PartialEq for ModelPaths

fn eq(&self, other: &ModelPaths) -> bool

Equality operator ==.

fn ne(&self, other: &Rhs) -> bool

Inequality operator !=.

impl StructuralPartialEq for ModelPaths