Institution profile

Cerebras Systems

Industry researchnorthamerica · us
Official website
Research library39linked papers
Opportunities41open roles
Selected work

Representative Papers

Straight to Zero: Why Linearly Decaying the Learning Rate to Zero Works Best for LLMs

Feb 21, 2025

This work addresses the optimization of learning rate (LR) decay scheduling in large language model (LLM) training. We systematically evaluate diverse LR schedulers and find that linear decay to zero (D2Z) consistently outperforms mainstream alternatives—including cosine decay to 10%—under computation-optimal token-per-parameter (TPP) scaling. Theoretically, we provide the first analysis showing that D2Z achieves an optimal trade-off: accelerating escape from suboptimal initial parameter regions early in training while effectively suppressing gradient noise in later stages; we further reinterpret AdamW’s exponential moving average of weight updates through this lens. Empirically, under TPP-controlled scaling across models from 610M to 7B parameters, D2Z enables the 610M model to achieve lower loss at just 80 TPP—reducing compute cost by 60% versus a 10× decay baseline requiring 200 TPP—and delivers substantial training efficiency gains for larger models including Llama2-7B.

1 citationsRead paper

RecurrentGPT: Expressive Depth through Recurrent Modulation in Transformers

Aug 15, 2026

This study addresses the loss of representational diversity caused by deep weight sharing in Transformers by proposing RecurrentGPT. The method employs fixed initial and final modules encapsulating an iterable core, incorporating a gating recurrent modulation mechanism based on hidden states and noise to enable functional specialization within a few shared layers during iteration. This approach overcomes the limitations of traditional parameter reuse. Experiments demonstrate that a three-layer model matches the accuracy of a 12-layer GPT-2 Small. At larger scales, RecurrentGPT reduces parameters by 63% and peak memory usage by 59% while achieving significantly lower validation loss than non-recurrent baselines, effectively balancing expressive capacity with memory efficiency.

0 citationsRead paper
Recent publications

Latest Papers

RecurrentGPT: Expressive Depth through Recurrent Modulation in Transformers

Aug 15, 2026

This study addresses the loss of representational diversity caused by deep weight sharing in Transformers by proposing RecurrentGPT. The method employs fixed initial and final modules encapsulating an iterable core, incorporating a gating recurrent modulation mechanism based on hidden states and noise to enable functional specialization within a few shared layers during iteration. This approach overcomes the limitations of traditional parameter reuse. Experiments demonstrate that a three-layer model matches the accuracy of a 12-layer GPT-2 Small. At larger scales, RecurrentGPT reduces parameters by 63% and peak memory usage by 59% while achieving significantly lower validation loss than non-recurrent baselines, effectively balancing expressive capacity with memory efficiency.

0 citationsRead paper

Unstructured Hydrodynamics on Spatial Dataflow Architectures: A Joint Code and Data Decomposition Approach

Jul 20, 2026

This work addresses the challenge of efficiently mapping unstructured-mesh multiphysics applications onto spatial dataflow architectures originally designed for structured grids, which suffer from irregular memory access patterns and high-dimensional communication overhead. The authors propose an automated co-decomposition approach that jointly partitions code and data by modeling communication and memory characteristics, leveraging space-filling curves to enable high-dimensional partitioning that minimizes inter-partition communication. Integrated memory optimization techniques further adapt the execution to memory-constrained hardware. The methodology is the first to systematically deploy unstructured-mesh applications on the Cerebras Wafer-Scale Engine, demonstrating superior performance over GPUs on large-scale LULESH hydrodynamics benchmarks and validating both the effectiveness and scalability of the proposed approach.

0 citationsRead paper