vllm.models.muse_spark.amd.mono.kernels

FlyDSL sources of the Muse Spark decode monokernel (see ../README.md).

k3.build_vllm(S, W, next_rope) (8 / 16 rows), k5.build_vllm (HA / HB) and k5m.build_vllm (HM) for 32-128 rows are the builds vLLM launches: TP = EP = 8, the router output multiplier folded into the router weight, q without the QK scale, bf16 KV pages in vLLM's LBHNC layout, and vLLM's shuffled MXFP4 expert weights and scales read in place. With attn="swa" / "full" the K3 / HA builds also run the layer's verify attention at their head (attn_frag.make: vLLM's bf16 KV cache, block table and seq_lens). Importing these modules needs FlyDSL and AITER's FlyDSL helpers.

Modules:

• attn_frag –

  The verify attention of one Muse Spark layer as a fragment of the persistent kernels: K3

• dev –

  Device-side primitives of the Muse Spark monokernel.

• k3 –

  K3: one Muse Spark decoder layer for S = 8 / 16 rows as one persistent launch.

• k5 –

  HA and HB: one Muse Spark decoder layer for S = 32 / 64 / 96 / 128 rows as two persistent

• k5m –

  HM: this rank's routed experts for S = 32 / 64 / 96 / 128 rows as one persistent launch