pytorch-inductor

Debug torch.compile failures by diagnosing TorchInductor compilation behavior, tracing FX graph lowering paths, inspecting generated Triton or C++ kernels, and analyzing fusion and memory planning to pinpoint missing decompositions or inefficient code generation.

The TorchInductor lowering pipeline works by applying operator decompositions, translating FX graphs into Inductor IR, executing scheduling and memory planning, and performing backend code generation to produce optimized Triton or C++ kernels.

Add a custom operator lowering in PyTorch Inductor by implementing the appropriate decomposition or lowering, defining layout constraints, integrating Triton templates, inspecting generated code, and creating validation tests for the new code generation path.

Yes, you can use Triton templates for custom backend integrations in torch.compile by developing custom lowerings, defining template selection logic, and configuring the Inductor IR to generate specialized Triton kernels.

The best way to optimize kernel performance in TorchInductor is to analyze fusion behavior, configure autotuning, inspect generated Triton code, and evaluate memory planning to improve CPU or GPU execution efficiency.

Your new operator fails under torch.compile because it likely lacks a required decomposition or lowering within TorchInductor; fix it by implementing the appropriate integration, inspecting the generated code, and creating validation tests.