About this role
Data AML is ByteDance’s machine learning platform team, building systems for recommendation, advertising, search, and other business applications. We develop high-performance, scalable, and reliable training and inference frameworks for large-scale recommendation models and large language models, drawing on the open-source model and framework ecosystem. Our technical challenges span training with massive sparse and dense parameter sets, long-sequence modeling, online learning, post-training paradigms such as SFT, RL, and OPD, and efficient coordination between training and inference. We translate advances in models and training methods into reusable framework capabilities that accelerate model iteration and production deployment.
Responsibilities
- Develop training and inference frameworks for large-scale recommendation models and large language models. Support massive sparse feature vocabularies and embeddings, joint training of sparse and dense parameters, long-sequence training, online learning, and post-training paradigms such as SFT, RL, and OPD to accelerate model iteration and enable efficient execution at scale.
- Evaluate, adapt, and extend open-source models and frameworks to meet business needs. Develop core functionality and improve framework architectures, working with projects such as vLLM, SGLang, verl, AutoModel, slime, ms-swift, Dspark, DFlash, and others.
- Develop efficient inference and model generation capabilities. Optimize coordination across training, generation, reward computation, and evaluation, including model weight synchronization, data flow management, and synchronous or asynchronous execution, to support flexible and efficient training workflows.
- Optimize distributed execution, parallelism strategies, computation and communication, GPU memory management, tiered storage, data loading, and resource scheduling to improve end-to-end throughput, utilization, and cost efficiency.
- Partner closely with modeling and infrastructure teams to understand business-specific models, data characteristics, and workloads. Deliver solutions from system design and implementation through model-quality and performance validation to production deployment, turning these solutions into reusable framework capabilities.
- Build testing, benchmarking, observability, fault diagnosis, checkpointing, and recovery capabilities to support reliable and maintainable training and inference at scale. Track relevant technologies and open-source developments, drive practical adoption, and contribute to the community where appropriate.
Minimum Qualifications:
- A background in computer science, software engineering, artificial intelligence, or a related field, with solid foundations in data structures, algorithms, operating systems, and computer architecture.
- Proficiency in Python or C++, with experience developing and debugging software in Linux environments and strong practices in code design, testing, and software engineering.
- Familiarity with at least one deep learning framework, recommendation model training system, or large-model post-training and inference framework, with an understanding of its execution workflow and experience in framework development, model integration, or system optimization.
- An understanding of distributed systems or parallel computing fundamentals, including how computation, communication, memory, and scheduling interact in multi-GPU training or inference.
- Strong technical problem-solving skills, with the ability to use code analysis, profiling, and experiments to identify bottlenecks and implement and validate improvements.
- Strong collaboration and ownership skills, with the ability to understand business needs, deliver solutions with modeling and infrastructure teams, and make practical trade-offs among efficiency, reliability, development cost, and long-term architecture.
Preferred Qualifications:
- Experience developing core components of training, post-training, or inference frameworks, with a deep understanding of relevant open-source designs and implementations or substantive open-source contributions.
- Experience adapting open-source models or frameworks for production use, including model integration, functionality extensions, performance tuning, compatibility management, or ongoing version upgrades.
- Experience developing training systems for recommendation, advertising, or search, with familiarity in massive sparse feature vocabularies and embeddings, parameter servers, joint training of sparse and dense parameters, long-sequence training, or online learning.
- Experience developing large-model post-training systems for paradigms such as SFT, RL, or OPD, with familiarity in coordination between training and generation, model weight synchronization, reward computation, data flow management, or asynchronous execution.
- Experience optimizing GPU workloads in one or more areas such as CUDA, Triton, kernel development, memory optimization, high-performance communication, or distributed parallelism.
- Experience with reliability or performance engineering for large-scale training or inference systems, including checkpointing and recovery, elastic execution, observability, tiered storage, resource scheduling, or end-to-end benchmarking.