Module binding 

2026-09-25: Typed weight binding for the GLM-5.3-Flash KDA attention blocks.

Owner: model-arch (GLM-5.3-Flash KDA). Invariants:

• bind_kda_weights returns weights only when a layer’s self_attn.* tensors are exactly KDA_TENSORS, each with its listed dtype and shape; anything else is an error.
• Bytes are uploaded unchanged: no cast, requantisation or dequantisation.

The MTP layer is DSA-shaped (it loads through glm5next_dsa, see weight_loader/glm5_next_mtp.rs) and is not bindable here; classify_attn_block tells the two apart from the tensor names alone.

The checkpoint stores three rank-3 conv tensors ([qkv, 1, kernel]). The binder concatenates them in q, k, v order and uses the result as [conv_dim, kernel]: dropping the singleton dimension moves no bytes. A_log (per head) and dt_bias (per channel) are F32; a BF16 copy of either is refused.

Structs

KdaBindReport

2026-09-25: Per-layer counts from bind_kda_weights. Tensors outside self_attn. are counted in non_attn_seen and never bound.

RawTensor

2026-09-25: One tensor as it sits in the checkpoint: dtype, shape and raw little-endian bytes.

TensorSpec

2026-09-25: One entry of KDA_TENSORS: a self_attn tensor’s name, dtype and shape.

Enums

AttnBlockKind

2026-09-25: What kind of attention block a layer’s tensor names describe.

KdaDtype

2026-09-25: The dtypes the KDA binder accepts.

Constants

DSA_MARKERS

2026-09-25: Names that identify a DSA (deepseek_sparse_attention) block, the MTP layer included, so a caller can classify a layer without its index.

KDA_TENSORS

Traits

KdaTensorSource

2026-09-25: A checkpoint slice scoped to one decoder layer. Names are layer-relative (self_attn.q_proj.weight), so the same binder works for any layer index.

Functions

bind_kda_weights

2026-09-25: Bind one KDA block: exact tensor set, dtypes, shapes and byte counts, or an error.

classify_attn_block

2026-09-25: Classify from tensor names alone: Dsa when every DSA_MARKERS name is present, else Kda when every KDA_TENSORS name is present, else Unknown.