Large Transformer Model Inference Optimization

Source: https://lilianweng.github.io/posts/2023-01-10-inference-optimization/ Author: Lilian Weng Date: 2023

Summary

Comprehensive survey of inference optimization techniques for large transformer models. Covers quantization, knowledge distillation, pruning, KV cache optimization, speculative decoding, and hardware-level techniques.

Key Claims

  • Four main optimization categories: model compression, parallelism, memory optimization, hardware efficiency
  • KV cache is the primary memory bottleneck for long-context inference
  • Speculative decoding: small draft model proposes tokens, large model verifies in parallel
  • Quantization-aware training vs post-training quantization tradeoffs
  • Inference is bottlenecked by memory bandwidth, not compute (at small batch sizes)

Entities

Concepts