The prohibitive cost of training trillion-parameter large language models (LLMs) necessitates low-cost emulation tools for distributed system optimization. In modern large-scale clusters, communication often becomes the primary bottleneck to scalability. However, existing emulators, such as vTrain and ASTRA-Sim, overlook dynamic network factors that significantly impact performance at scale, resulting in limited emulation accuracy. This work offers an efficient and reliable tool for training system optimization and parallel strategy exploration, considerably lowering the barrier to large-scale AI research. We present LLMEmu, a distributed training emulator that combines real kernel profiling and actual communication execution. First, computation is profiled through real CUDA kernel traces on GPU nodes to construct an operator-level latency lookup table, enabling GPUlike execution on CPU clusters. Second, inter-node communication is executed using communication library primitives (e.g., AllReduce, Send/Recv), triggered by communication anchors embedded in the execution graph, and implemented using a pluggable communication backend. LLMEmu can seamlessly model hybrid parallelism strategies and supports multiple collective algorithms. Its lightweight design incorporates gradient bucketing with latency reuse to minimize overhead while maintaining extensibility to various network interconnects. The effectiveness of LLMEmu is validated through its performance results, demonstrating an average prediction error of only 2.17% on 24-GPU clusters, which outperforms vTrain by 21.09%, and confirming its scalability in modeling training cost distributions across 128- node CPU emulations under varying network conditions.
Yang et al. (2026) studied this question.