phi4_multimodal

Model type

phi4_multimodal

Class

Phi4MMMultiModalModel

Task

phi4mm-multimodal

Source

models/phi.py

Description

Phi-4 Multimodal model (4-model split).

Produces four separate ONNX models via Phi4MMMultiModalTask:

  • vision: SigLIP encoder + projection -> image_features

  • speech: Conformer encoder + projection -> audio_features

  • embedding: token embedding + InputMixer fusion -> inputs_embeds

  • decoder: LoRA text decoder + lm_head -> logits + KV cache

Replicates HuggingFace’s Phi4MMForCausalLM.

Usage

mobius build --model <MODEL_ID> output_dir/

from mobius import build

model = build("<MODEL_ID>")