Module weights

Module weights 

Source
Expand description

2026-09-25: Device weight handles for one GLM-5.3 MLP site, already sharded for this rank.

Owner: model-arch (GLM-5.3). Invariants: none beyond the types; build.rs fills them.

Structs§

Glm5NextDenseMlpWeights
2026-09-25: A BF16 SwiGLU MLP: the dense layers 0..first_k_dense_replace, and the shared expert of every routed layer.
Glm5NextExpertPtrTable
2026-09-25: Device pointer tables for one projection over the full expert set.
Glm5NextExpertWeights
2026-09-25: One routed NVFP4 expert, owned whole by one EP rank.
Glm5NextMoePtrTables
2026-09-25: The three projections’ pointer tables for one routed site.
Glm5NextMoeWeights
2026-09-25: A routed MoE site’s weights for this rank.
Nvfp4Proj
2026-09-25: One NVFP4 projection: packed e2m1 pairs, per-16-element e4m3 block scales, and one global f32 scale.