shadcn.io is not affiliated with official shadcn/ui
TPU Usage Monitor Block for Svelte
The last 20 training steps as stacked columns of compute, waiting for input, cross-chip communication and idle time, with a verdict from a typed rule and the four numbers of the step you choose.
Installation
Overview
See where each training step spends its time, with this profiler panel for Svelte with Ark UI. It is Steploom, a training-step profiler (made up), reading the last 20 steps of a run on a Google Cloud TPU slice. Each step is a stacked column of compute, waiting for input, cross-chip communication and idle time on one millisecond axis, told apart by weight and hatch. A typed rule gives the verdict: when waiting for input passes 25% of all step time the run is input-bound, and the header says so. The columns are a native radio group, so the arrow keys move between steps and choosing one reads its four numbers. Under the plot three figures give matrix unit use, memory use and the median step. Guides draw the frame and its bands; the entrance is CSS only. For training engineers looking for a slow input pipeline. The numbers are examples, not from any provider.
FAQ
How do I use my own steps?
Edit the STEPS list at the top of the file; in the demo it is built from small noise helpers. Each step has four times in milliseconds and a step number. The verdict, the input share, the median step and the highlighted worst step are derived. MATRIX_USE and MEMORY_USE are typed cluster figures.
What is the verdict rule?
If the time spent waiting for input is more than 25% of all step time across the 20 steps, the run is called input-bound and the header says how large the share is; otherwise it is called compute-bound with the compute share. The threshold is INPUT_BOUND_FROM.
How do I pick a step?
Click a column, or Tab into the group and use the arrow keys. The chosen column is outlined, its four numbers appear under the plot and a polite status line speaks them. The panel starts on the step that waited longest.
What is real here and what stays local?
Nothing is profiled and no cluster is contacted; the names and numbers are examples. The vendor is named once as the kind of hardware the made-up tool reads.