Reposted
TikTok · Onsite
In processTikTok AI Infra Intern Onsite Round 2: 1 Experience
AI infrastructure engineer · InternshipMaster's degreeMar 2026
Shared by
A TikTok AI Infra intern's second onsite round in March 2026 asked for attention, MoE and MLA from scratch, DeepSeek-V3 math, heap sort and a fast Reduce.
At a glance
- Process: onsite round 2, almost entirely implementation from scratch with compute and memory-access analysis (report 01).
- Asked recently: Linear Attention, MoE, MHA and MLA from scratch; DeepSeek-V3 pseudocode with parameter and memory-access estimates; heap sort; an optimized Reduce (report 01, Mar 2026).
- Difficulty: heavy: every model component also needed a hand calculation of compute and memory traffic.
Coverage: 1 report · AI Infrastructure Engineer intern · onsite round 2 · March 2026 · result pending when posted.
What came up
The round tested whether the candidate could write modern transformer pieces without libraries and reason about their cost on hardware.
Questions we found
ML implementation
| Question | Reports | Last seen |
|---|---|---|
| Implement Linear Attention from scratch | 1 | Mar 2026 |
| Implement a Mixture of Experts layer (routing and experts) | 1 | Mar 2026 |
| Implement Multi-Head Attention from scratch | 1 | Mar 2026 |
| Implement Multi-Head Latent Attention from scratch | 1 | Mar 2026 |
| Compute and memory access of each implementation | 1 | Mar 2026 |
| DeepSeek-V3 pseudocode, parameter count and per-inference memory access | 1 | Mar 2026 |
Algorithms / Data Structures
| Question | Reports | Last seen |
|---|---|---|
| Implement heap sort from scratch | 1 | Mar 2026 |
| Implement and optimize a Reduce operation | 1 | Mar 2026 |
1 detailed candidate timeline
01: transformer components from scratch
Interview date: March 2026 · Role: AI Infrastructure Engineer intern · Result: In process
Onsite round 2
- Write Linear Attention without high-level library implementations.
- Write a Mixture of Experts (MoE) layer, including routing and the expert computation.
- Write Multi-Head Attention (MHA), then Multi-Head Latent Attention (MLA), from scratch.
- For each of these, calculate the computation and the memory access required.
- Write pseudocode for the DeepSeek-V3 architecture, count its parameters by hand, and estimate the memory access per inference pass.
- Coding: heap sort from scratch.
- Implement a Reduce operation, then discuss and implement optimizations to speed it up.
Common questions
What coding does TikTok ask AI Infra interns in round 2?
Model components from scratch (Linear Attention, MoE, MHA, MLA) with compute and memory analysis, plus heap sort and an optimized Reduce (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.
0 total reactions. Your reactions: none.
0 comments
Comments
0No comments yet — be the first to share what you know.