Source description
About the role
We are looking for an Engineer to build the training infrastructure, data pipelines, and inference optimization systems for state-of-the-art Diffusion Transformer (DiT)models. This role focuses on scaling the fine-tuning and deployment of models like Qwen , Wan , and LTX-2 . Key Responsibilities Training Infrastructure: Design and maintain scalable pipelines for training and fine-tuning Diffusion Transformer models on large-scale GPU clusters. Model Optimization: Optimize the inference performance of Wan , LTX-2 , and Qwen (Vision)using quantization, pruning, and hardware-aware tuning (e.g., TensorRT, FlashAttention). Data Engineering: Develop efficient ingestion and preprocessing pipelines for high-resolution image and video datasets used in generative tasks. Capability Expansion: Implement engineering workflows that allow researchers to rapidly fine-tune and expand the capabilities of open-weights diffusion models. Production Deployment: Transition experimental fine-tuned models into reliable, low-latency production services. Resource Management: optimize distributed training jobs (FSDP, DeepSpeed) to maximize GPU utilization and minimize costs. Required Qualifications Min 2 years of experience in Machine Learning Engineering with a focus on generative models. Core Tech: Strong proficiency in PyTorch , JAX , and distributed training frameworks. Model Expertise: Hands-on experience deploying or fine-tuning Diffusion Transformers (DiT)and specifically Qwen (Image) , Wan , or LTX-2 . Architecture: Deep understanding of Transformer-based diffusion backbones and flow matching (removing legacy reliance on CNNs/RNNs). Tooling: Proficiency in Python and modern ML ecosystem tools (e.g., Hugging Face, Diffusers, FFmpeg for video processing). Compute: Experience debugging and optimizing workloads in multi-node GPU environments. Preferred Qualifications Inference Optimization: Experience with techniques like KV-caching, compile-time optimizations, or kernel fusion for transformers. MLOps: Familiarity with experiment tracking (W&B) and model versioning tools in a generative media context. Streaming: Experience handling real-time video generation or streaming inference pipelines. Open Source: Contributions to libraries like diffusers or active experimentation with the latest open-source DiT implementations. We are looking for an Engineer to build the training infrastructure, data pipelines, and inference optimization systems for state-of-the-art Diffusion Transformer (DiT)models. This role focuses on scaling the fine-tuning and deployment of models like Qwen , Wan , and LTX-2 . Key Responsibilities Training Infrastructure: Design and maintain scalable pipelines for training and fine-tuning Diffusion Transformer models on large-scale GPU clusters. Model Optimization: Optimize the inference performance of Wan , LTX-2 , and Qwen (Vision)using quantization, pruning, and hardware-aware tuning (e.g., TensorRT, FlashAttention). Data Engineering: Develop efficient ingestion and preprocessing pipelines for high-resolution image and video datasets used in generative tasks. Capability Expansion: Implement engineering workflows that allow researchers to rapidly fine-tune and expand the capabilities of open-weights diffusion models. Production Deployment: Transition experimental fine-tuned models into reliable, low-latency production services. Resource Management: optimize distributed training jobs (FSDP, DeepSpeed) to maximize GPU utilization and minimize costs. Required Qualifications Min 2 years of experience in Machine Learning Engineering with a focus on generative models. Core Tech: Strong proficiency in PyTorch , JAX , and distributed training frameworks. Model Expertise: Hands-on experience deploying or fine-tuning Diffusion Transformers (DiT)and specifically Qwen (Image) , Wan , or LTX-2 . Architecture: Deep understanding of Transformer-based diffusion backbones and flow matching (removing legacy reliance on CNNs/RNNs). Tooling: Proficiency in Python and modern ML ecosystem tools (e.g., Hugging Face, Diffusers, FFmpeg for video processing). Compute: Experience debugging and optimizing workloads in multi-node GPU environments. Preferred Qualifications Inference Optimization: Experience with techniques like KV-caching, compile-time optimizations, or kernel fusion for transformers. MLOps: Familiarity with experiment tracking (W&B) and model versioning tools in a generative media context. Streaming: Experience handling real-time video generation or streaming inference pipelines. Open Source: Contributions to libraries like diffusers or active experimentation with the latest open-source DiT implementations.