Module layer 

Glm5NextDsaLayer — the DSA block: NoPE MLA attention over indexer-selected tokens.

Decode, end to end:

hidden ─┬─ q_a_proj ─ RMSNorm ─┬─ q_absorb ────────────── Q (latent space)
        │                      └─ indexer.wq_b ────────── q_idx  ─┐
        ├─ indexer.wk ─ LayerNorm(w,b) ─ state.k_normed ──────────┤
        ├─ compress_gate ─────────────── state.gate ──────────────┼─ select_tokens
        ├─ weights_proj ──────────────── head weights ────────────┘        │
        └─ kv_a_proj ─ RMSNorm ─ FP8 ─── paged latent cache                │
                                                                           ▼
                                           glm5next_dsa_mla_decode_fp8 (gather)

🪤 Four silent-wrong-answer traps this file exists to hold

• Two RMSNorm kernels differ only by a +1. rms_norm computes x * rms * (1 + w); rms_norm_vanilla computes x * rms * w. Same signature, same shapes. GLM is plain, so every norm here takes the vanilla entry point.
• indexer.k_norm is an nn.LayerNorm with a bias, not an RMSNorm at all — mean subtraction plus a bias term. It takes nllb_layernorm_bf16(x, w, b, …).
• weights_proj output must already carry index_heads^-0.5. dsa_index_scores does not apply it. Folded into the weight at load — see Glm5NextDsaWeights.
• Q must be absorbed into latent space before it reaches the decode kernel. The kernel dots Q against the 512-dim latent directly, so q_absorb is q_b_proj pre-multiplied by kv_b_proj’s K half. A raw q_b_proj is the right shape per head (256 vs 512 is not) but the wrong space.

Structs

Glm5NextDsaLayer

Glm5NextDsaLayerKernels

Every kernel a DSA block launches, beyond the selection set.

Glm5NextDsaWeights

One DSA block’s weights, already sharded for this rank.

Glm5NextDsaWorkspace

Scratch reused across decode steps. Allocated once per layer.