Embedding Vc
视频生成模型 · 训练 Infra 工程师
San Francisco Bay Area
Get past the screening software and onto a recruiter's desk
hirly rewrites your resume for this job — matching the keywords and skills in the posting, moving your most relevant experience to the top, and writing a cover letter to fit. About 30 seconds.
- Keywords matched to this posting
- Fit score before you apply
- Cover letter included
Matched against 2.4M live jobs from 200,000+ employers in 200+ countries.
Tailor my resume for this job →Apply from your AI assistant
Connect hirly to Claude and ask it to apply to this job. hirly tailors your resume, fills the employer’s form and asks before sending. ChatGPT: manual setup today.
Some employer sites stop an application at a CAPTCHA or sign-in and hand it back with a link. Applying needs a paid plan. Works with any assistant that supports MCP.
hirly's read of this role
- Seniority
- Mid level
- Country
- US
- Work mode
- Remote-friendly
- First seen by hirly
- 10 Sept 2026
Derived automatically from the posting. Upload your resume above to see how the role scores against it.
the posting
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。
Browse similar roles
Want this one?
Upload your resume and hirly rewrites it for this job and writes the cover letter — in about thirty seconds, before you sign up.
Tailor my resume for this job