Verify Distributed Deep Learning Model Implementation Refinement with Iterative Relation Inference
This addresses a critical reliability issue for developers implementing distributed training of large models, though it is an incremental improvement in verification tools.
The paper tackles the problem of bugs in distributed deep learning implementations by developing a static verification approach that checks whether a distributed model's outputs can reconstruct the sequential model's outputs, evaluating it on GPT and Llama-3 models.
Distributed machine learning training and inference is common today because today's large models require more memory and compute than can be provided by a single GPU. Distributed models are generally produced by programmers who take a sequential model specification and apply several distribution strategies to distribute state and computation across GPUs. Unfortunately, bugs can be introduced in the process, and a distributed model implementation's outputs might differ from the sequential model's outputs. In this paper, we describe an approach to statically identify such bugs by checking model refinement, that is, can the sequential model's outputs be reconstructed from the distributed model's outputs? Our approach, implemented in GraphGuard, uses iterative rewriting to prove model refinement. Our approach can scale to today's large models and deployments: we evaluate it using GPT and Llama-3. Further, it provides actionable output that aids in bug localization.