Inside Google Gemma 3n: my PyTorch Profiler insights
Hi everyone,
If you’ve ever wondered what really happens inside modern vision-language models, here’s a hands-on look. I profiled the Google Gemma 3n model on an NVIDIA GPU using PyTorch Profiler, asking it to describe a bee image.
I visualized the profiling results using https://ui.perfetto.dev/, as shown in the animated GIF below:
I’ve shared the full code, profiling scripts, and raw trace data, so you can dive in, reproduce the results, and explore the model’s internals for yourself.
👉 https://github.com/sbnb-io/gemma3n-profiling/
If you’re looking to better understand how these models run under the hood, this is a solid place to start. Happy to hear your thoughts or suggestions!
Hi @aospan1 ,
Welcome to Google Gemma family of open source models, thanks for sharing your script with us, we are really excited to try this from our end and will share the feedback and suggestion with you.
Thanks.