Module weight_loader 

Weight loading traits and per-model loader implementations.

Translates flat WeightStore into typed TransformerLayer objects. Each model architecture has its own ModelWeightLoader implementation that knows the HuggingFace weight name patterns.

Submodules contain per-family loaders:

• qwen3: Qwen3-Next (NVFP4, hybrid SSM+Attention+MoE)
• qwen35: Qwen3.5 MoE (35B, 122B)
• qwen35_dense: Qwen3.5 Dense (27B)
• qwen3_vl: Qwen3-VL (vision-language)
• nemotron: Nemotron-H (Mamba-2 + MoE + Attention)
• gemma4: Gemma-4 (pure attention, GeGLU, sliding + full attention)

Re-exports

pub use dflash_loader::DflashConfig;

pub use dflash_loader::DflashLayerWeights;

pub use dflash_loader::DflashSubConfig;

pub use dflash_loader::DflashWeights;

pub use dflash_loader::load_dflash_weights;

pub use dflash_loader::store_has_dflash_weights;

pub use glm5_next_load::Glm5NextWeightLoader;

pub use kimi_k3::KimiK3WeightLoader;

Modules

dflash_loader

DFlash drafter weight loader.

fp8_residency

The native-FP8 dense loader’s derived-copy decision table and the shape arithmetic that prices it (#915). Which derived weight copies the native-FP8 dense route actually needs, and a tally of the ones it still builds.

glm5_next

GLM-5.3-Flash tensor accounting (Slice 1: classification only). GLM-5.3-Flash (glm5_next) tensor accounting.

glm5_next_load

Glm5NextWeightLoader — assembles the 45-layer GLM-5.3 text stack from a WeightStore.

glm5_next_vision

Bind GLM-5.3-Flash’s vision tower: 347 tensors under model.visual.*.

kimi_k3

Kimi K3 weight loader — C1: BF16 0.40B twin bind. Packed experts bind to DSV4 E8M0 (quantized_k3_mxfp4_e8m0) only when K3_ALLOW_MXFP4=1. Default refuse so an accidental official download cannot silently land.

predicted_residency

That table evaluated BEFORE the checkpoint loads, so preflight can size the SSM decode ring against the predicted post-load KV headroom (#915). The derived-weight residency of the native-FP8 dense route, predicted from config.json alone — BEFORE the checkpoint loads.

Structs

DeepSeekV4WeightLoader

Gemma4WeightLoader

LagunaWeightLoader

LongcatWeightLoader

MinimaxM2WeightLoader

NemotronHWeightLoader

NllbWeightLoader

Qwen3VLWeightLoader

Qwen3WeightLoader

Qwen4ExpWeightLoader

Qwen35DenseWeightLoader

Qwen35WeightLoader

Step3p7WeightLoader

Enums

QuantFormat

Runtime quantization format for weight dispatch.

WeightFormat

Checkpoint weight format, detected from safetensors metadata.

Traits

ModelWeightLoader

Loads weights from a WeightStore into typed layer objects.