About the job
Would you like to participate in creating the fastest Generative Models inference in the world? Join the Cerebras Inference Team to participate in development of unique Software and Hardware combination that sports best inference characteristics in the market while running largest models available. Cerebras wafer scale inference platform allows running Generative models with unprecedented speed thanks to unique hardware architecture that provides fastest access to local memory, ultra-fast interconnect and huge amount of available compute.
Responsibilities
Analysis of new models from generative AI field and understanding of impacts on compilation stack
Develop and maintain model definition framework that consists of model building blocks to represent large language models based on PyTorch and Cerebras dialects ready to be deployed on Cerebras hardware.
Develop and maintain the frontend compiler infrastructure that ingests PyTorch models and produces an intermediate representation (IR).
Extend and optimize PyTorch FX / TorchScript / TorchDynamo-based tooling for graph capture, transformation, and analysis.
Collaboration with other teams throughout feature implementation
Research on new methods for model optimization to improve Cerebras inference
Qualifications
Minimum
Degree in Engineering, Computer Science, or equivalent in experience and evidence of exceptional ability
Strong Python programming skills and in-depth experience with PyTorch internals (e.g., TorchScript, FX, or Dynamo).
Solid understanding of computational graphs, tensor operations, and model tracing.
Experience building or extending compilers, interpreters, or ML graph optimization frameworks.
Experience working with PyTorch and HuggingFace Transformers library
Knowledge and experience working with Large Language Models (understanding Transformer architecture variations, generation cycle, etc.)
Strong C++ programming skills.
Knowledge of MLIR based compilation stack
Preferred
Prior experience contributing to PyTorch, TensorFlow XLA, TVM, ONNX RT, or similar compiler stacks.
Knowledge of hardware accelerators, quantization, or runtime scheduling.
Experience with multi-target inference compilation (e.g., CPU, GPU, custom ASICs).
Understanding of numerical precision trade-offs and operator lowering.
Contributions to open-source ML compiler projects.