Responsibilities

• Develop novel LLMs and primitives that take advantage of Tenstorrent’s breakthrough accelerator architecture to deliver orders of magnitude performance & efficiency improvements.
• Model benchmarking, Performance Analysis, Characterization on novel hardware accelerator.
• Fine tune/Retrain emerging LLM models after architectural modification to improve performance.
• Implement ideas from latest research papers and validate new optimizations on novel hardware, suggest improvements on the existing ideas and validate new ideas through experiments.
• Implement ML models on Tenstorrent’s PyBUDA Compiler framework and analyze results, suggest novel architectural changes to make them more efficient.

Experience & Qualifications

• Deep understanding of Transformer based model architecture (e.g., BERT, GPT, LLaMA, FALCON) - should be able to analyse performance bottlenecks in training or inference.
• Expertise (through evidence in prior projects or research) in large-scale foundation model (CNN/ Transformer/Generative) training and optimisations including quantisation, pruning, low-rank approximation (e.g., LoRA, SVD), sparsity aware training, early exit models, conditional computing, sparse attentions, NAS, fine-tuning (e.g., RLHF, PEFT)
• Excellent working knowledge of Python, PyTorch or TensorFlow, (optional C/C++ if interested in compiler and software stack)
• Familiarity with and passion for any of the following -- ML compilers, high-performance and massively parallel systems, distributed model training and inference, GPU or AI accelerator architecture -- is a plus.