Reproduce “Delta-Matching: Closing the Final Gap of Native 8-bit Training for LLMs” 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
Haozhan Tang, Hao Kang, Han Cai, Song Han, Chenyan Xiong
Delta-Matching addresses a numerical mismatch in low-precision attention backpropagation by recomputing the softmax row correction from backward operands rather than saved forward output. Under stated assumptions, this restores the score gradient’s zero-row-sum invariant before its FP8 cast. On the tested 1.67B-parameter model, validation cross-entropy was 1.4162, close to 1.4178 for the BF16/FP32 reference, but the cast and other quantization errors remain.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.