Why your LLM is slow — a beginner's guide to reading PyTorch traces
huggingface launched a three-part series on profiling pytorch models with torch.profiler, starting with matmul basics. it aims to help engineers identify bottlenecks and squeeze more tokens per second out of llms.
huggingface released part one of a tutorial series titled 'profiling in pytorch', designed to teach developers how to use torch.profiler to diagnose and optimize model performance. the guide focuses on reading profiler tables and traces for simple matrix multiplication and addition operations, explaining the link between python calls and cuda kernels. by mastering these profiling tools, developers can identify bottlenecks in their pipelines and understand the performance impacts of torch.compile. future parts will scale these lessons to nn.linear and transformers.