Module binding 

GLM-5.3 DSA checkpoint binding: the 14 self_attn tensors of a DSA block, their expected dtype and shape, and a verifier that fails loudly.

Scoped to LibertAIDAI/GLM-5.3-Flash-NVFP4@9e0d74e3. Mirrors crate::layers::glm5next_kda::binding and deliberately reuses its RawTensor / TensorSource / dtype types instead of growing a parallel set.

Shapes are expressed against Glm5NextDsaConfig, so a geometry change fails here rather than at kernel launch — the same contract the KDA binder holds.

🪤 What this exists to catch

• indexer.k_norm has a bias. It is an nn.LayerNorm, not an RMSNorm — the only other norm in GLM-5.3 with a bias. A binder that loads only .weight drops mean-subtraction and the bias, and nothing about the shapes says so. The spec table lists the bias as REQUIRED so its absence is an error, not a silent zero.
• index_kpool_compress_ape is [kpool, index_head_dim] — indexed by pool SLOT, not by head and not by pool. Its 1 KB size makes a wrong-axis bind easy to miss.
• q_b_proj and kv_b_proj carry different per-head widths — qk_head_dim (256) vs nope + v_head_dim (512). Both are [heads * w, lora], so a swapped width still yields a well-formed 2-D tensor.
• NoPE: there is no wkv_a_rope / wq_b_rope here at all, and kv_a_proj_with_mqa is kv_lora_rank wide (512), not + rope (576). A spec that expects 576 rejects the real checkpoint.

Structs

DsaBindReport

What a successful bind saw.

DsaSpec

One expected tensor: layer-relative name, dtype, and full (unsharded) shape.

Functions

dsa_tensor_specs

Every self_attn tensor a DSA block has — and the complete list of what it may have. full_heads is the pre-shard head count; the checkpoint is never sharded on disk, so binding always validates against the full geometry and slices after.

verify_dsa_block

Verify one DSA block against the spec table.