InternVL 2.5 Multimodal Model: Dynamic High-Resolution Slicing & LMDeploy Serving Guide

🔍 Dynamic image slicing algorithm (Dynamic High-Resolution) principle

For engineering drawings and PDF long screenshots up to 4K/8K pixels, InternVL 2.5 will dynamically calculate the optimal tile grid (such as 2x2, 3x3), send local high-definition tiles to the visual encoder, and retain global thumbnails. This dual-channel design takes into account the macro view of global composition and ultra-high-resolution recognition of tiny text.

⚡ Image pulling and LMDeploy extremely fast inference

# Set up mirroring
export HF_ENDPOINT="https://hf-mirror.net"
huggingface-cli download OpenGVLab/InternVL2_5-8B \
  --local-dir /data/models/InternVL2_5-8B \
  --local-dir-use-symlinks False

# Use LMDeploy to start the high-performance API interface (single card 16G running)
lmdeploy serve api_server /data/models/InternVL2_5-8B \
  --backend turbomind \
  --server-port 8000