Skip to content

vllm.models.deepseek_v41.amd.mono

DeepSeek-V4.1-Flash mono decode layer: one decoder layer of a decode step in two persistent FlyDSL launches, its TP all-reduces in-kernel (README.md). The host side is runner; vLLM docks it in ..mono_decode.

Modules:

  • attention –

    The mono layer's attention stages: front (K1: wqkv, q / kv norms and the

  • common –

    The mono kernels' shared mechanisms: ops (wave / math primitives, the TP

  • layer –

    The mono decode layer's two launches (README): K1 = the attention seam

  • runner –

    DeepSeek-V4.1-Flash mono decode layer: host side.

  • sources –

    The mono layer's kernel sources, in every build's JIT cache key

  • stages –

    The V4.1 stages the mono layer adapts from ROCm/ATOM: the mHC seams