Module binding 

Typed weight binding for the GLM-5.3-Flash KDA attention family.

Every KDA self_attn block in the checkpoint binds through bind_kda_weights, which is exhaustive and strict: the tensor set must be exactly the 15 names below, every dtype and shape is asserted, and any unrecognised self_attn.* tensor is a hard error. There is no “skip what we don’t know” path, silent or otherwise.

Why this can be strict

The family is structurally uniform. Audited across the checkpoint’s 34 KDA layers: one distinct (name, dtype, shape) signature, 0 quantisation artefacts, 0 missing or unexpected tensors — while all 510 tensor hashes are distinct, so the blocks share structure and nothing else. Layer 45 (MTP) is DSA-shaped, not KDA, and is not bindable here; classify_attn_block separates the two from the tensor names alone.

Traps this module exists to make impossible

• The checkpoint stores three conv tensors of rank 3 ([qkv, 1, kernel]); HF holds one fused depthwise conv. Binding is concat([q, k, v]) in that order, squeezed exactly once. Both the order and the squeeze are silent if wrong — the order because all three have identical shape, the squeeze because [dim, 1, ks] and [dim, ks] share their bytes.
• A_log is per head and F32; dt_bias is per channel and F32. Everything else is BF16. A loader that “helpfully” casts either to BF16 changes the gate.

Structs

KdaBindReport

Per-layer accounting, so “zero unknown, zero silent skips” is a reported number and not a claim. non_attn is counted but deliberately NOT bound — FFN, norms and mHC are other slices.

RawTensor

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

TensorSpec

The 15 self_attn tensors a KDA block has — and the complete list of what it may have.

Enums

AttnBlockKind

What kind of attention block a layer’s tensor names describe.

KdaDtype

The only two dtypes a KDA block contains.

Constants

DSA_MARKERS

Names that identify a DSA (deepseek_sparse_attention) block, including the MTP layer. Present so a caller can classify without guessing from the layer index.

KDA_TENSORS

Traits

KdaTensorSource

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 and any container.

Functions

bind_kda_weights

Bind one KDA block. Strict: exact tensor set, exact dtypes, exact shapes.

classify_attn_block

Classify from tensor names alone — never from the layer index, and never from layer_types, which Slice 1 has to strip and rebuild.