RepostedTikTok logo

TikTok · Onsite

In process

TikTok AI Infra Intern DeepSpeed Round: 1 Experience

AI infrastructure engineer · InternshipMaster's degreeApr 2026
Shared by

A TikTok AI Infra intern onsite round 1 in April 2026 covered DeepSpeed and ZeRO-3, BLIP-2 and Q-Former, chart-to-story data, and LeetCode 69.

At a glance

  • Process: onsite round 1: basic knowledge, training optimization, a project, then one coding problem (report 01).
  • Asked recently: DeepSpeed and the ZeRO Stage 3 memory-saving formula; other distributed training frameworks; BLIP versus BLIP-2 and Q-Former; preparing data for a chart-to-story model; integer square root (report 01, Apr 2026).

Coverage: 1 report · AI Infrastructure Engineer intern · onsite round 1 · April 2026 · result pending when posted.

What came up

Distributed training memory and multimodal model design made up most of the round, followed by an easy coding problem with an optimization discussion.

Questions we found

Algorithms / Data Structures

Question Reports Last seen
Integer square root (LeetCode 69), then optimizations 1 Apr 2026

ML systems and models

Question Reports Last seen
Main idea and use of DeepSpeed 1 Apr 2026
Why ZeRO Stage 3, and derive its memory saving 1 Apr 2026
Other distributed training frameworks and how they differ 1 Apr 2026
BLIP versus BLIP-2; common modern multimodal architectures 1 Apr 2026
How Q-Former works and what it is for 1 Apr 2026
Content understanding approaches; data for a chart-to-story model 1 Apr 2026

1 detailed candidate timeline

01: DeepSpeed, multimodal models and integer square root

Interview date: April 2026 · Role: AI Infrastructure Engineer intern · Result: In process

Basic knowledge

  1. The main idea behind DeepSpeed and what it is used for.
  2. Why use ZeRO Stage 3; how it saves GPU memory, with the memory-saving formula derived.
  3. Other common distributed training frameworks and their main differences.
  4. BLIP versus BLIP-2, and the architecture most modern multimodal models use.
  5. How Q-Former works and its purpose.

Training optimization

  1. A typical approach for content understanding tasks, and how to prepare training data for a chart-to-story model.

Project

  1. A key project and the main work involved.

Coding

  1. Implement integer square root (LeetCode 69) and discuss ways to optimize it.

Common questions

What ML systems questions does TikTok ask AI Infra interns?

DeepSpeed and the ZeRO-3 memory formula, distributed training frameworks, and multimodal architectures such as BLIP-2 and Q-Former (report 01).

Keep practicing

Sources: Anywhere useful interview info can be found.

0 total reactions. Your reactions: none.
0 comments

Comments

0

No comments yet — be the first to share what you know.