buptzyb

Investigate real PyTorch CUDA Graph tensor mismatches, replay drift, NaNs, gradients, allocator growth, pool behavior, or eager-versus-graph differences with torch-cudagraph-debug. Use when debugging CUDA Graph accuracy or GPU memory issues in an application workload, conducting a fresh investigation, or analyzing persisted bundles. Do not use for implementing the debug library itself or for generic CUDA questions that do not require its public APIs.

Operational guide for choosing and combining parallelism strategies in Megatron Bridge, including sizing rules, hardware topology mapping, and combined parallelism configuration.