视频生成模型 · 训练 Infra 工程师
Embedding VCUpload your resume once and Scout applies for you
Apply with ScoutYou will build and operate the training infrastructure for an in-house video-generation foundation model, including distributed training across hundreds of GPUs and large-scale video data pipelines. The role also focuses on training performance, memory use, recovery, and stability at cluster scale.
- Full-time
- Fully Remote
Skills and expertise
Key responsibilities
- Build reusable training infrastructure for job scheduling, checkpointing, monitoring, and data and weight pipelines.
- Run and optimize distributed training across hundreds of GPUs using parallelism strategies.
- Improve video data pipelines, memory use, performance, and recovery from training failures.
How this job compares
Compared with 362 open mid-level infrastructure jobs on November 21, 2026.
- Pay
- This job does not publish its pay. The 96 open mid-level infrastructure jobs that publish US pay start at $170K a year on average.
- Similar open jobs
- We Love AI Jobs lists 362 open mid-level infrastructure jobs at 144 companies.
- Remote work
- This job is remote. 57 of the 362 mid-level infrastructure jobs (16%) are remote.
- AI in this role
- High. AI is the core of the work: building, training or deploying AI systems. The role builds infrastructure for training the company’s video-generation foundation model and works directly with distributed model-training jobs. Its responsibilities include training performance, stability, and video data pipelines.
Full description
我们在训练自研视频生成基础模型(DiT / Flow Matching),需要一位既能搭起训练平台、又能把研究代码变成数百卡集群上稳定结果的工程师。你不只是用平台的人,更是建平台的人。
你会做
• 训练平台搭建:从作业调度、断点续训、监控告警到数据 / 权重流水线,把分散的脚本沉淀为团队可复用的训练基础设施。
• 数百卡规模的分布式训练:FSDP、张量并行、Context Parallel、Ulysses,把 MFU 推到合理水位。
• PB 级视频数据 pipeline:NVDEC 解码、VAE latent 缓存、变分辨率 bucket sampling。
• 显存与性能:FlashAttention、FP8 混合精度、Triton kernel、activation checkpoint 策略。
• 训练稳定性:loss spike 根因分析、断点秒级恢复、慢节点自动剔除。
硬性要求
• 精通 PyTorch distributed 与 CUDA 体系结构。
• 至少一个主流训练框架(Megatron / DeepSpeed / FSDP / TorchTitan)的源码级理解。
• ≥ 256 卡训练实战经验。
• 有从零或半程搭建训练平台 / 集群调度 / 训练工具链的经验。
加分
• DiT / Diffusion / 视频数据处理经验。
• 写过 Triton / CUTLASS kernel。
• 对 HunyuanVideo / Wan / CogVideoX 等开源项目有源码级了解。
我们提供
• 真实数百卡算力、把训练当工程问题的团队、合规边界内的开源 / 发表空间。
Upload your resume once and Scout applies for you
Apply with Scout