Conquering FP16 Frustrations: One-Liner to FP8 Gold on H100s
·54 words·1 min
Below is super nice! I’ve spent fair chunk of my life to get FP16 working flawlessly. Now with amp integrated in PyTorch, things are easier but you can still spot missing/incorrect calls in lot of OSS. Below one liner makes all that go away. If you have H100s, FP8 is pure gold. https://x.com/thecharlieblake/status/1694281272527372349