vllm.models.deepseek_v4.amd.mono.attention

The mono layer's attention stages: front (K1: wqkv, q / kv norms and the KV insert, wq_b) and back (K2: split attention, combine, wo_a, wo_b).

Modules:

• back –

  K2 back: one attention layer from the rotated query to the rank's wo_b

• device –

  Device primitives of the mono layer's attention stages.

• front –

  K1 front: one attention layer (the counts below are TP8's) from its

• gemv –

  MXFP8 GEMV of the mega kernel: 16 rows of a row-major e4m3 weight [N, K]

• plan –

  Geometry of the attention stages (front, back): the KV record, the