TikTok · Onsite
In processTikTok AI Infra Intern Kernel Round: 1 Experience
A TikTok AI Infra intern onsite round 1 in April 2026 was all GPU depth: kernel fusion, Hopper warp specialization, MoE inference and RL routing collapse.
At a glance
- Process: onsite round 1 in three blocks: kernel optimization, inference optimization, then project and internship deep dives (report 01).
- Asked recently: custom CUDA or CuTe DSL kernels for fusion, when fusion slows things down, Hopper warp specialization, MoE expert specialization and routing collapse under RL, and AI agents that generate CUDA kernels (report 01, Apr 2026).
- Difficulty: no coding problem; every question asked for hands-on experience with a concrete example.
Coverage: 1 report · AI Infrastructure Engineer intern · onsite round 1 · April 2026 · result pending when posted.
What came up
The round assumed real GPU kernel and MoE inference experience, and kept asking why an optimization helped or hurt.
Questions we found
GPU kernels and inference
| Question | Reports | Last seen |
|---|---|---|
| Kernel-level work with CuTe DSL or custom CUDA kernels for fusion | 1 | Apr 2026 |
| Which approach or framework for kernel fusion, and why | 1 | Apr 2026 |
| When has kernel fusion made performance worse, and why? | 1 | Apr 2026 |
| How does warp specialization on Hopper GPUs work? | 1 | Apr 2026 |
| Without warp specialization but with tiling and shared memory, where is the loss? | 1 | Apr 2026 |
| Optimizing MoE inference, especially with RL | 1 | Apr 2026 |
| Has an MoE model really learned expert specialization? | 1 | Apr 2026 |
| Routing collapse in RL + MoE training, and fixes | 1 | Apr 2026 |
| Using an AI agent to generate CUDA kernels | 1 | Apr 2026 |
1 detailed candidate timeline
01: kernels, MoE inference and production serving
Interview date: April 2026 · Role: AI Infrastructure Engineer intern · Result: In process
Kernel optimization
- Experience with kernel-level optimization, such as CuTe DSL or custom CUDA kernels for operator fusion, with a specific example.
- Which approach or framework you would choose for kernel fusion, and why.
- Whether fusion ever made performance worse, and what caused it.
- How warp specialization works on NVIDIA Hopper GPUs, in hardware and software.
- If warp specialization is removed but tiling and shared memory stay, where the main performance loss comes from.
Inference optimization
- Hands-on MoE inference optimization, especially with reinforcement learning (RL).
- How to tell whether an MoE model has learned meaningful expert specialization rather than acting like a dense model split across experts.
- Routing collapse in RL + MoE training, where the reward pushes most tokens to a few experts, and how to fix it.
Projects and internship
- The most relevant project: challenges and contributions.
- The internship, especially optimization work on production inference services.
- Whether they had tried an AI agent that generates CUDA kernels automatically, and the approach.
Common questions
What GPU questions does TikTok ask AI Infra interns?
Kernel fusion and when it backfires, Hopper warp specialization, and what tiling and shared memory do without it (report 01).
Does the TikTok AI Infra intern round 1 have coding?
Not in this report: it was all kernel, MoE inference and project questions (report 01).
Keep practicing
- TikTok question banks: algorithms, behavioral (sign-in needed).
- Everything TikTok in one place: the TikTok interview page and TikTok hiring insights.
- More firsthand reports: the Experience feed.
Sources: Anywhere useful interview info can be found.
Comments
0No comments yet — be the first to share what you know.