vllm.models.dots3_note.nvidia

NVIDIA-specific Dots3Note model implementations.

Modules:

• attention –

  Dots3 NOTE sliding-window MLA attention backends for Hopper.

• audio_encoder –

  Dots-path speech encoder for inference only (single GPU).

• model –

  NVIDIA implementation of the Dots3Note language model.

• mtp –

  NVIDIA multi-token predictor for Dots3Note.

• multimodal –

  vLLM composition layer for Dots3Note image and audio encoders.

• vision –
• vision_attention –

  Shared vision attention stack for Dots dense / MoE ViT encoders.

• vision_moe –

  NOTE vision MoE execution matching the native encoder's FP8 semantics.