Skip to content

vllm.models.deepseek_v4.common

Modules:

  • mm_preprocess –

    Multimodal preprocessing for the DeepSeek-V4 vision variants

  • ops –
  • rope –

    DeepseekV4 rotary embedding initialization.

  • vision –

    DeepSeek-V4 vision tower (ViT + aligner) with TP-sharded linears.

  • vl_model –

    DeepSeek-V4 vision variant (e.g. DeepSeek-V4-Flash-Vision-Exp).