视频生成模型 · 训练 Infra 工程师
This is the employer's own posting, not a copy on a job board.
What we know
Is it still open?
Confirmed still open
Last checked 1d ago — checked against the employer's own applicant tracking system, which is the company answering directly.
We re-read the employer's own applicant tracking system and the posting was still there. That is the company answering directly.
How old is it?
Posted 49d ago
The date the source published, not the day we noticed it (2026-07-28). Last seen at its source 1h ago.
Is it remote?
Marked remote on the employer's board
Their board carries a remote setting on this posting — a field they filled in, not wording we read. The location field names somewhere specific, which is usually where the team or the entity sits.
Who may apply?
San Francisco, Bay Area
The description states no restriction of its own. This is the source's own tag.
Skills named in the ad
Recognised terms only, from a fixed vocabulary — this is what CV matching compares against.
Carried by 1 source
-
ashby employer's own board first seen 11d ago · last seen 1h ago
The listing
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。