Jobs Companies Embedding VC 视频生成模型 · 训练 Infra 工程师

About this 视频生成模型 · 训练 Infra 工程师 role at Embedding VC

Embedding VC · Remote · San Francisco Bay Area

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。

Ready to apply to Embedding VC?
Apply to Embedding VC

Similar jobs

Sign up for suggestions tailored to the jobs you open and the searches you save.

More jobs at Embedding VC

See all jobs at Embedding VC →

Apply now
🤖

Whoa — hold up

JobsRadar was built for real people having a rough time in their job search — not for automated requests. You're clicking way too fast and you're now temporarily blocked.

Come back later. If you're genuinely job hunting, we've got your back — just act like a human.

Catch your next role the second it’s posted.

Create a free account and we’ll watch the boards for you — the instant a job matches your search, it lands in your inbox or Telegram. No digging, no refreshing.

Create free account

Free forever · takes 30 seconds · already have one?

Get an edge on your job hunt.

Join our Telegram channel for the stuff that helps you land the role — salary benchmarks, the weekly market pulse, and new-feature drops. No spam, just signal.

Join the channel — it's free