I went through a GPU compiler loop last year. For me it was two technical screens then an onsite with a few deep dives, but it varied by team. I barely saw LeetCode style questions. Most coding was C or C++ plus compiler flavored graphs and systems thinking. What helped was building a tiny LLVM pass and being ready to explain CFGs, dominators, SSA, and a simple register allocation via graph coloring. I also drilled topo sort and parsing a small IR while talking out loud. I used timed mocks with Beyz coding assistant alongside prompts from the IQB interview question bank. Keep behavioral answers to about 90 seconds using STAR. You got this.