DeepView: A Generalizable Debugging Methodology for Enabling Large Language Models on Emerging AI Accelerators

Deploying large-scale generative AI models on new hardware accelerators presents significant and recurring engineering challenges. Failures arising from unsupported operations, graph compilation breaks, and numerical divergences between hardware backends are difficult to locate, reproduce, and resolve without deep manual expertise, often consuming days to weeks of engineering effort—and this cost is paid repeatedly for each model per platform. We present DeepView, a backend-agnostic debugging framework designed to systematically accelerate model enablement on any accelerator integrated through PyTorch’s torch.compile infrastructure. DeepView provides three complementary capabilities: (1) unsupported operations detection, which pinpoints unsupported PyTorch operators with rich metadata and auto-generates reproduction code; (2) layer debugging, which isolates the precise layer where compilation or runtime failures occur; and (3) layer I/O divergence analysis, which compares intermediate tensor outputs across backends using cosine similarity and mean absolute difference metrics against data-driven thresholds. By building on PyTorch’s forward hook mechanism, DeepView is model- and hardware-agnostic, requiring no modification to model source code and no backend-specific instrumentation. We demonstrate the methodology using IBM’s Spyre dataflow accelerator as a representative out-of-tree backend in PyTorch, showing reductions in manual debugging time from days or weeks to minutes across multiple production Large Language Models (LLMs) including Granite, Mistral, and Bamba. The generality of the approach makes it directly applicable to other emerging accelerators with custom torch.compile backends.