Expand description
2026-09-25: Glm5NextWeightLoader: builds the GLM-5.3 text stack from a
WeightStore, plus the helpers the MTP loader shares.
Glm5NextTextSkeleton::from_config fixes each layer’s mixer (KDA or DSA)
and MLP (dense or routed MoE) from the config’s layer_types and
mlp_only_layers; the loader builds what it lists.
Under TP, DSA shards through DsaTpPlan, the MLP through the rank passed to
glm5next_mlp::build, and KDA through KdaShardedSource, which slices
the host bytes before bind_kda_weights sees them. At TP > 1 both mixer
plans report needs_output_all_reduce, and the layer all-reduces the mixer
output (Glm5NextLayer::mixer_all_reduce).
Owner: model-arch weight loader. Invariants: none beyond the types.
Structs§
Functions§
- upload_
bf16 - 2026-09-25:
upload_f32_as_bf16, for the MTP loader.