Sky is Blue

Category : Testing & Debugging Code Training LLMs is expensive. Burning hours on an H100 only to discover your model was memory-bound the entire time is an expensive mistake and it happens constantly. Live Roofline Analyzer fixes that. It's a VS Code extension that shows you exactly what's bottlenecking your transformer — compute or memory — in real time, right inside your editor. Open any Python training file, and it auto-detects your model architecture, plots every transformer operation on a live roofline chart, and tells you precisely which ops are wasting your GPU. Drop one Python file next to your training script and the panel updates live as training runs, showing step count, tokens per second, and a continuously updated roofline. No more guessing. No more profiling after the fact. No more wasted compute. Supports TPU v4/v5e/v5p, A100, H100, and H200 out of the box, with one-click presets for LLaMA 3 8B/70B/405B, Mistral, Gemma, and GPT-2. Hit one button and GPT-4o streams back concrete optimization suggestions — batch size, precision, parallelism strategy — based on your actual numbers.