- https://publish.obsidian.md/aixplore/Practical+Applications/dgx-lab-benchmarks-vs-reality-day-4#%20GPU%20Inference%20is%20Fundamentally%20Broken https://publi

- https://publish.obsidian.md/aixplore/Practical+Applications/dgx-lab-benchmarks-vs-reality-day-4#%20GPU%20Inference%20is%20Fundamentally%20Broken https://publish.obsidian.md/aixplore/Practical+Applications/... Does it work if you change to torch.bfloat16? - https://publish.obsidian.md/aixplore/Practical+Applications/dgx-lab-benchmarks-vs-reality-day-4#1.+ARM64+Architecture https://publish.obsidian.md/aixplore/Practical+Applications/... The PyTorch 2.9 wheels do work. You can pip install torch --index-url <whatever-it-is> and it just works. You do need to build flash attention from source, which takes an hour or so.