mlx-bun library API
    Preparing search index...
    GraphMediaInput: "embeddings" | "embeddings+positions" | "pixels"

    How prepared media reaches the graph.

    • embeddings: spliced prompt embeddings (image/audio soft tokens, optional bidirectional and multimodal masks) enter the ordinary forward.
    • embeddings+positions: spliced embeddings with request-owned multi-axis rotary positions and a decode delta.
    • pixels: raw image pixels are the denoising method's own prefill input.