Auto uses chat_template when available, otherwise falls back to processor(text, media).
Media
Processor options
Input Flow
flowchart LR
A[Messages] --> B[Media paths]
B --> C[Chat template not loaded yet]
C --> D[Processor AutoProcessor]
B --> D
D --> E[input_ids tokenizer]
D --> F[media tensors]
E --> G[model forward]
F --> G
Mode: not run yet (requested Auto)
Text rule: raw prompt text
Processor path: click Inspect Processor Input
Media tensors: none
Metadata: none
Flowchart content updates only on Inspect; tab highlights are frontend-only.
Final messages after uploaded media paths and processor options are applied.