Research Scientist — Privacy-Preserving Large-Scale Model Training & Architecture Optimization

TikTok
San Jose
Workplace: OnsiteFull timeFunction: Data Science & Machine LearningSkills: []

Build next-generation generative foundation model training architectures with a focus on diffusion-based and unified generation-understanding models. Own end-to-end design for large-scale, privacy-aware training infrastructure, including model-parallel execution, GPU-first performance optimization, and robust distributed strategies. Develop fault-tolerant, self-healing systems with fast recovery, checkpointing, and production readiness. Optimize diffusion transformer pipelines (noise schedules, timesteps, memory-efficient attention) and partner with research teams on architecture tradeoffs.

Loading

Loading job details...

Preparing the role view and application actions.

FursaFursa
TikTok
TikTok
1 month ago

Research Scientist — Privacy-Preserving Large-Scale Model Training & Architecture Optimization

✓ Verified Job

Canonical indexed version, validated from employer's careers page.

Source: Company careers pageValidated by: Fursa AI
Last checked: 21 hours agoStatus: Live

Job Summary

Build next-generation generative foundation model training architectures with a focus on diffusion-based and unified generation-understanding models. Own end-to-end design for large-scale, privacy-aware training infrastructure, including model-parallel execution, GPU-first performance optimization, and robust distributed strategies. Develop fault-tolerant, self-healing systems with fast recovery, checkpointing, and production readiness. Optimize diffusion transformer pipelines (noise schedules, timesteps, memory-efficient attention) and partner with research teams on architecture tradeoffs.
Location: San Jose
Workplace: Onsite
Employment Type: Full time
Job Function: Data Science & Machine Learning
Seniority: Mid level

Key Responsibilities

  • •Design and optimize large-scale training architectures for diffusion-based and unified generative models (e.g., DiT, Rectified Flow, hybrid AR + diffusion).
  • •Lead GPU-centric performance optimization across thousands of accelerators, including memory layout, communication overlap, and kernel fusion.
  • •Develop distributed training strategies (DP / TP / PP / ZeRO / FSDP-style sharding) for long-running, multi-stage foundation model training.
  • •Build fault-tolerant, self-healing training systems with fast failure detection, recovery, checkpointing strategies, and restart policies.
  • •Optimize diffusion transformer training pipelines (noise schedules, timestep strategies, memory-efficient attention) and support unified generation-and-understanding multimodal models.

Key Requirements

  • •Strong background in large-scale deep learning systems and distributed training.
  • •Hands-on experience with GPU optimization, including memory management, performance profiling, and communication/computation overlap.
  • •Experience training diffusion models (e.g., DiT) or large foundation models at scale.
  • •Proficiency in PyTorch and modern distributed training stacks.
  • •Working knowledge of parallelism strategies (DP / TP / PP / ZeRO / FSDP or equivalents) and training stability for long-running jobs.
Experience:Large-scale deep learningDistributed trainingFoundation modelsPrivacy-preserving ML
Tech Stack:PyTorchGPU optimizationDPTPPPZeROFSDPCUDA kernelsKernel fusionMemory-efficient attentionCommunication libraries

Company Brief

TikTok
Short-form video platform that lets users create, share, and discover entertainment content through algorithmic recommendations. It also offers advertising and creator tools for brands, influencers, and businesses.
Industry: Digital Media
Company Size: Enterprise (1,001+ employees)
Growth: Scaleup
Headquarters: Singapore, Singapore
Founded: 2016
WebsiteLinkedIn