FastV + Impromptu-VLA Reproduction

This repository documents my reproduction and analysis of FastV-style attention-based image token pruning, and its applicability to Impromptu-VLA (Qwen2.5-VL) under different inference backends.

What is included

• ✅ FastV reproduced under HuggingFace / LLaVA path
• ✅ Impromptu-VLA baseline inference (Qwen2.5-VL, sglang)
• ✅ DivPrune-style static image token truncation
• ❌ True FastV-style in-place pruning under sglang (analyzed)

Key findings (TL;DR)

• FastV requires decoder-level attention access and dynamic sequence length
• sglang pre-allocates KV cache slots, preventing mid-forward token pruning
• DivPrune works because it truncates tokens at preprocessing stage
• True KV-dropping FastV is infeasible in sglang without kernel-level changes

Repository Structure

fastv-impromptu-vla/ ├── README.md ├── REPRODUCTION.md ├── ANALYSIS.md ├── scripts/ ├── src/ └── configs/

How to run

See scripts/ for runnable examples.

References

• FastV:https://github.com/pkunlp-icler/FastV
• Impromptu-VLA: Impromptu VLA paper