Modern large language model (LLM) training frameworks widely employ operator optimizations to improve efficiency. These optimizations are typically assumed to be numerically equivalent to their standard counterparts. However, due to the inherent imprecision of floating-point arithmetic, even mathematically equivalent implementations can introduce small per-step numerical deviations that silently accumulate during training. Such deviations constitute silent errors: they trigger no crashes or exceptions, yet cause optimized and unoptimized training runs to diverge over time. This paper presents \textsc{DFuzz}, a differential fuzzing framework for systematically detecting and quantifying silent errors in operator optimizations. \textsc{DFuzz} searches for worst-case numerical deviations by exploring both the training configuration space and the input space. It employs a two-stage pipeline: \emph{configuration fuzzing}, which uses a genetic algorithm to identify training settings that amplify deviations, followed by \emph{input fuzzing}, which further maximizes deviation via projected gradient-based search. We evaluate \textsc{DFuzz} on 44 operator optimizations across three major LLM training frameworks. The results show that \textsc{DFuzz} reveals silent errors in 16 operators (12 previously undocumented), and that both fuzzing stages significantly outperform random baselines by up to $13.33\times$. \textsc{DFuzz} uncovers three classes of silent errors: floating-point precision accumulation, undocumented algorithmic approximations, and implementation defects in LLM-generated code.