Module attend 

GLM-5.3 DSA attention — the launcher for the selected-index NoPE MLA paged decode.

Consumes what super::select produced: a [q_rows, out_width] i32 row of token ids with -1 holes. Gathers exactly those tokens from the paged FP8 latent cache.

Why this is not the masked kernel

dsa_mla_masked_attn is an oracle (see super::MASKED_ATTN_MAX_KEYS). The production path gathers per row, which is what HF says the reference cannot do (_supports_flash_attn = False, “cannot be mapped to FA without a custom kernel that can select on a per indices bases per row”) and what vLLM ships. The gather is exactly equivalent, not an approximation: the reference mask is pure set membership, with duplicates collapsed and no additive weighting.

🪤 NoPE, and why no common/ kernel would do

GLM-5.3 has qk_rope_head_dim == 0: the latent is the whole cache token. common/mla_paged_decode.cu declares kv_cache_dim and never reads it (its strides come from #define ROPE_DIM 64); common/mla_paged_decode_fp8.cu uses the runtime stride but then overwrites dims 448–511 with rope taken from the next token. Both fail silently. Hence a GLM-target kernel with no rope arm at all.

Structs

DsaDecodeInputs

Everything the decode reads, all caller-owned.

DsaDecodePaging

Paging geometry the decode needs and the selection does not.

Glm5NextDsaDecodeKernel

The selected-index MLA decode entry point.

Constants

DSA_DECODE_MODULE

Module name the DSA decode kernel resolves from — an unlisted .cu takes its file stem, and this one lives in the glm-5.3-flash target, not common/.

Functions

decode_attention

Launch the selected-index decode. Enqueued on stream, not synchronised.