About The Role 

The Core ML team at Cerebras is looking for both senior and junior research scientists dedicated to the development of novel, state-of-the-art ML algorithms. Our team focuses on research areas that emphasize our hardware's unique capabilities: 

• Accelerated unstructured sparse matrix multiplications enable weight sparsity and activation sparsity.
• Unprecedented on-chip memory bandwidth enables >10x faster inference than GPU-based hyperscale cloud services.
• Weight streaming execution mode enables training 1T+ parameter models on 1 WSE, simple data parallel scaling instead of 3D+ parallelism required on GPUs.
• Flexible support to use low-precision numerics to improve performance of many aspects of training.

We also invest in advancing our fundamental understanding of ML training dynamics and distilling these insights into recipes which systematically improve existing customer ML workloads.  

Here is a sampling of our recent publications and releases:  

• Straight to Zero: Why Linearly Decaying the Learning Rate to Zero Works Best for LLMs, ICLR 2025 [paper](https://www.arxiv.org/abs/2502.15938) 
• The practitioner's guide to the maximal update parameterization, 2024, https://cerebras.ai/blog/the-practitioners-guide-to-the-maximal-update-parameterization 
• Sparse maximal update parameterization: A holistic approach to sparse training dynamics, NeurIPS 2024 [paper](https://arxiv.org/abs/2405.15743) 
• Normalization Layer Per-Example Gradients are Sufficient to Predict Gradient Noise Scale in Transformers, NeurIPS 2024 [paper](https://arxiv.org/abs/2411.00999) 
• Sparse-IFT: Sparse Iso-FLOP Transformations for Maximizing Training Efficiency, ICML 2024 [paper](https://arxiv.org/abs/2303.11525) 
• SPDF: Sparse Pre-training and Dense Fine-tuning for Large Language Models, UAI 2023 [paper](https://arxiv.org/abs/2303.10464) 
• BTLM-3B-8K: 7B Parameter Performance in a 3B Parameter Model https://huggingface.co/cerebras/btlm-3b-8k-base (most popular 3B model on Hugging Face), 2023 [paper](https://arxiv.org/abs/2309.11568) 
• SlimPajama: A 627B token cleaned and deduplicated version of RedPajama, 2023, https://www.cerebras.net/blog/slimpajama-a-627b-token-cleaned-and-deduplicated-version-of-redpajama 
• Cerebras-GPT: Open Compute-Optimal Language Models Trained on the Cerebras Wafer-Scale Cluster, 2023, [paper](https://arxiv.org/abs/2304.03208) 

Here is a sampling of our research directions: 

• Scaling laws to predict and analyze improvements in large-scale training: accuracy/loss, architecture scaling, and hyperparameter transfer. 
• Sparse and low-precision training algorithms for reduced training time and increased accuracy: weight and activation sparsity, mixture-of-experts, and low-rank adaptation. 
• Improving training dynamics and efficiency through advanced optimizers, initializers, and normalizers. 

As a member of the Core ML team, you'll engage in groundbreaking research alongside a collaborative and dedicated group. We deliver impactful demos of Cerebras capabilities and publish our findings at top-tier venues in order to engage and support the wider community. Senior roles will involve providing mentorship and guidance to other talented scientists and engineers.  

We offer competitive compensation, flexible working arrangements, ample computing resources, collaboration opportunities with leading academic groups, and extensive options for professional growth and advancement in your AI journey. 

If you join our team, you'll develop deep ML expertise, engage with our team and partners on impactful ML applications. Join us now to push the boundaries of AI---Scale up your career with Cerebras. 

Responsibilities 

• Develop novel training algorithms that advance the state-of-the-art in model quality and compute efficiency.
• Develop novel network architectures that address foundational challenges in language and multi-modal domains.
• Co-design ML algorithms that take advantage of our unique hardware, and collaborate with engineers to co-design next-generation architectures.
• Design and run research experiments that show novel algorithms are efficient and robust.
• Analyze results to gain research insights, including training dynamics, gradient quality, and dataset preprocessing techniques.
• Publish and present research at leading machine learning conferences.

Skills And Qualifications 

• Strong grasp of machine learning theory, fundamentals, linear algebra, and statistics.
• Experience with machine learning frameworks, such as PyTorch and Jax.
• Strong track record of research success through relevant publications at top conferences or journals (e.g. ICLR, ICML, NeurIPS), or patents and patent applications.

Preferred Skills And Qualifications 

• PhD in a relevant discipline.
• Experience with state-of-the-art transformer language models.
• Experience with distributed training concepts and frameworks, such as TorchTitan, Megatron/Deepspeed, or FairSeq FSDP.
• Experience with training speed optimizations, such as model architecture transformations to target hardware, or low-level kernel development (e.g., Triton).