Reproduce “Efficient, VRAM-Constrained xLM Inference on Clients” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Aditya Ukarande, Deep Shekhar, Marc Blackstein, Ram Rangan
This research introduces pipelined sharding and VLM-specific optimizations (VLMOpt) within the `llama.cpp` framework to enable efficient, lossless inference of large language and vision models (xLMs) on client systems with limited VRAM. The solution delivers up to 30x Tokens Per Second (TPS) speedups for LLMs and reduces VRAM demand for VLMs by up to 10x, making high-accuracy xLMs feasible on consumer hardware.
Key insights
- A detailed benchmark-profile database, incorporating CPU and GPU performance metrics, can accurately predict the optimal hybrid execution strategy for xLMs across varying VRAM budgets and system configurations.
- Dynamically selecting execution plans based on the current 'new token' tier (e.g., context prefill vs. token decoding) significantly improves both Time-To-First-Token (TTFT) and Tokens Per Second (TPS) compared to static manual offloading.
- Specialized VLM optimizations, such as FlashAttention for vision encoders and VRAM allocation overlap avoidance, are essential to enable high-resolution VLM inference on memory-constrained client devices without sacrificing accuracy.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.