Softmax Reparameterization for Output-Head Quantization

  • 类型:arxiv
  • 标识:2609.31291
  • 链接:https://arxiv.org/abs/2609.31291
  • 主分类:llm-infra
  • 形态:method
  • TLDR:Large vocabularies make output heads a substantial inference cost in small language models. We propose softmax reparameterization, a post-training method that selects a functionally equivalent output head before quantization. The method subtracts a scalar multiple of the vocabulary-row mean from every output row and selects the coefficient by validation KL separately for RTN, activation-weighted MSE, and full-Hessian GPTQ. This one-dimensional search includes the original head and fixed mean-centering, preserves the full-precision softmax distribution, and leaves the trained decoder unchanged;
  • 待LLM分类:否
  • 来源文件:
  • /inbox/tom/_candidates/2026-09-29-rag-retrieval-reranking-candidates.json
  • /inbox/tom/_candidates/2026-09-29-agent-rag-longcontext-candidates.json