I wrapped up my internship at
@zml_ai in Paris, where I spent the summer working on inference.
- Added the Vulkan Graphics API as a hardware target for the ZML/XLA compiler
- Owned model bring-up for Stepfun MoE LLMs
- Added tensor primitives to ZML’s core library
- Learned the importance of hermetic builds
The hardest bug I hit was a silent predicate-layout mismatch between LLVM IR and Vulkan. I wrote up what I learned here:
reesechong.com/blog/zml?v=2
Thanks to
@steeve,
@GuillaumeLST, and the ZML team for a great summer!
My X account is still based in France :)