Tag: On-device AI
All the articles with the tag "On-device AI".
The Six-Step Sentis Workflow: The Ownership and Blocking the Docs Leave Out
Unity's official Sentis workflow page is a six-step skeleton. Here is what it omits — the Resources folder, tensor ownership, and the GPU stall hiding in the line that reads your result — filled in against the rest of the same manual.
The Ollama Response Was Long Because of the Format, Not the Answer
I checked a set of notes on the Ollama API against the official docs. The cause of the too-long response was one parameter, but the notes changed three at once, so the credit got split three ways.
Reading the ML-Agents Install Docs: Python 3.10.12 Is a Ceiling, Not a Recommendation
Why ML-Agents installation is so particular, traced back to the toolkit being split in two. What Conda and the exact 3.10.12 are for, how the Unity and Python package versions pair up, and the two names for the inference layer.
Accessing Tensor Data Directly in Sentis: Why and How to Avoid Readback
The bottleneck in on-device inference is usually not inference itself but data movement between GPU and CPU. Here are the Sentis APIs that reduce it.