Tag
The post explains why LLM inference is nondeterministic due to non-associative floating-point addition and batch-dependent kernel selection, then presents DetLLM, a system that runs all LLM operations (matmuls, softmax, normalization, RoPE) in integer-only arithmetic to achieve bit-identical outputs across any GPU, CPU, or batch size. DetLLM became a paper accepted at two NeurIPS 2026 workshops.
Taco 提出一种跨平台 LLM 压缩方法,通过纯整数推理、精确的 float32 到整数转换与一致性测试向量,确保编码端与解码端在不同 GPU/CPU 硬件上产生完全相同的输出,并在 7 个后端上验证了哈希一致。
The cryptoint library (2025) introduces a library focused on cryptographic integer arithmetic, providing constant-time, portable operations for security-sensitive applications.
本文提出代数整数量化(AIQ)作为大语言模型推理的有损压缩格式,利用代数整数可实现无误差乘加运算的特性构建矩阵乘法引擎,在等比特预算下达到与业界标准 MX 格式竞争的精度。
博客文章指出整数与浮点算术遵循不同范式:整数运算需在灾难发生前预防溢出,而 IEEE-754 浮点运算中的 NaN 与无穷值会自然传播,不应照搬预防式检查(如用零分母或 FLT_EPSILON 守卫),这些检查在多数场景下无效且被 LLM 频繁写错。
This GitHub repository provides CMake scripts to execute GPT-2 using Q16.16 integer arithmetic, including instructions for running both full and toy models.
The article argues for using unsigned integers more often in programming, refuting common arguments against them.