Skip to content

vllm.models.deepseek_v4.amd

Modules:

  • dspark –

    DSpark draft model for DeepSeek-V4 on ROCm/AMD (gfx950).

  • mega_moe –

    AITER MegaMoE experts (fused dispatch, MXFP4 GEMMs, combine) for DeepSeek V4.

  • model –
  • mono –

    DeepSeek-V4-Pro mono decode layer: one decoder layer of a decode step in two

  • mono_decode –

    DeepSeek-V4-Pro mono decode layer on ROCm CDNA4 (--monokernel).

  • mtp –

    MTP draft model for DeepSeek V4 (internal codename: DeepseekV4).

  • rocm –