Phase 13 — Cheatsheet: Multimodal Models

• Image -> vision encoder -> image embeddings -> placeholder token positions -> normal LLM.
• EncoderCacheManager reuses image features; don't recompute per step.
• Image tokens inflate seq length and KV usage; profile input processing.

Key upstream files

• vllm/multimodal/
• vllm/multimodal/processing.py
• vllm/v1/core/encoder_cache_manager.py
• vllm/model_executor/models/llava.py
• vllm/model_executor/models/qwen2_vl.py

Full reference: 00-guide.md · 01-deep-dive.md