vllm.models.kimi_k3.amd.ops

Modules:

• flydsl_sparse_mla –

  FlyDSL sparse MLA forward on the fp8_ds_mla cache (gfx950).

• flydsl_sparse_mla_gfx950 –

  FlyDSL sparse MLA forward over fp8_ds_mla records, gfx950.

• fp4_indexer –

  No-RoPE FP4 indexer quantization for the AITER row-group layout.

• fp4_indexer_prefill –

  FP4 prefill: gather compact index keys once, then score global causal rows.

• kda_checkpoint –

  Conv half of the ROCm Kimi-K3 prefill checkpoint.

• kda_chunk –

  ROCm entry point for the fused Kimi-K3 KDA chunk kernel.

• kda_decode –

  ROCm entry points for the fused Kimi-K3 KDA decode kernel.

• kda_prefill –

  KDA prefill backend selection for ROCm.

• third_party –