I'm utterly shocked at the article saying GPU inference (PyTorch/Transformers)isn't working. Numerical instability produces bad outputs, Not viable for real-time serving, Wait for driver/CUDA updates!

My job just got me and our entire team a DGX spark. I'm impressed at the ease of use for ollama models I couldn't run on my laptop. gpt-oss:120b is shockingly better than what I thought it would be from running the 20b model on my laptop.

The DGX has changed my mind about the future being small specialized models.