Responsibilities

• Integrating SW stack with DNN frameworks (e.g., PyTorch) to support eager mode, model export path, custom kernel interface and DNN compiler.
• Collaborate with cross-teams, including compiler and algorithm teams, to achieve the native integration between SW stack and DNN frameworks.
• Design and implement DNN model graph’s pre-processing modules, involving sub-graph pattern matching/replacement, and graph partitioning, to enable the model parallelism and optimize the end-to-end inference performance.
• Participate in designing and implementing the user-facing interface of LLM stack.
• Write and maintain API references and development documentation.

Minimum Qualifications

• Bachelor’s degree in Computer Science or equivalent work experience.
• Experience with ML/DNN frameworks, such as TensorFlow, PyTorch, or JAX.
• Strong programming skills with 3+ years of experience in Python.
• Excellent communication skills for gathering and clarifying requirements.
• Ability to identify inefficiencies in programs and processes, with a proactive approach to proposing sustainable solutions.

Preferred Qualifications

• Deep understanding of transformer-based DNN model architecture.
• 2+ years of experience working in AI research.
• Familiarity with PyTorch 2.0 technologies (e.g., TorchDynamo, TorchInductor) or DNN compiler technologies (e.g., Triton, MLIR).
• Proficient programming skills, particularly in CUDA or Rust.
• Experience developing production-grade software for customers.
• Understanding of LLM serving optimization techniques, such as mixture of experts, paged attention, speculative decoding, chunked prefill, or continuous batching
• Experience with Hugging Face Transformes, PEFT, TGI, vLLM, or TensorRT-LLM
• Proven track record of contributing to open-source projects.
• Knowledge of testing and CI/CD pipelines, such as Jenkins or similar tools.

Contact

• minsup.lee@furiosa.ai