compile-trace-inductor

To debug torch.compile missed fusion problems, trace Inductor lowering, scheduling, and loop-body construction to investigate scheduler dependencies and identify why operations are not fusing into FusedSchedulerNodes.

Too many kernel launches in PyTorch Inductor compiled models are caused by missed fusion opportunities during scheduling, which you can diagnose by inspecting intermediate buffers and SchedulerNode dependencies using TORCH_LOGS.

Trace Triton code generation failures by following Inductor IR transformations from Pointwise and Reduction ops to generated Triton code, using config.trace.enabled to inspect progressive compilation artifacts and isolate compilation errors.

Yes, you can use TORCH_LOGS to diagnose incorrect output from a torch.compile model by comparing transformations across Inductor IR levels, checking ops.load and ops.store operations, and tracing execution to find dtype and indexing issues.

Interpret Inductor IR and SchedulerNode structures for performance analysis by tracing Pointwise and Reduction operations through lowering and fusion, examining generated Triton code, and evaluating intermediate buffer behavior to identify bottlenecks.