Large language models (LLMs) have taken the tech world by storm over the past couple of years, dominating headlines with their ability to generate convincing human writing given simple prompts. But LLMs have also shown promise well outside of poetry and prose, with researchers leveraging their power to understand trends in other fields. Now (speaking of things that have taken the world by storm), one Gordon Bell Prize nominee is applying LLMs to transform how new and emergent variants of viruses like SARS-CoV-2 are identified and classified. The research leveraged several major supercomputers, as well as a major AI-focused system from Cerebras Systems. The work is nominated for the Gordon Bell Special Prize for High Performance Computing-Based Covid-19 Research. The three finalist teams for the prize presented their work at SC22 in Dallas ahead of the reveal of the winners at the conference tomorrow.
The paper’s 34 authors hail from a wide range of institutions: universities include California Institute of Technology, Harvard, Northern Illinois University, Technical University of Munich, University of Chicago, University of Illinois Chicago; others include Cerebras, Nvidia and Argonne National Laboratory.
The broad partnership adapted LLMs to understand genomic data, building what they instead call GenSLMs – short for “genome-scale language models.” They then trained a GenSLM on 110 million prokaryotic gene sequences (open-source data from the Bacterial and Viral Bioinformatics Resource Center) before fine-tuning it for SARS-CoV-2 using 1.5 million genomes. In comments to Nvidia, the team stressed the importance of tackling this data at the nucleotide level rather than the small molecule or protein level.
“We hypothesized that moving from protein-level to gene-level data might help us build better models to understand Covid variants,” said Arvind Ramanathan, head of the project and a computational biologist at Argonne. “By training our model to track the entire genome and all the changes that appear in its evolution, we can make better predictions about not just Covid, but any disease with enough genomic data.”
But this is no easy task, with billions of nucleotides in the human genome and tens of thousands of nucleotides in viruses like SARS-CoV-2. And, Ramanathan said: “The meaning of a nucleotide sequence can be affected by another sequence that’s much further away than the next sentence or paragraph would be in human text. It could reach over the equivalent of chapters in a book.” To ameliorate this problem, Nvidia said that it helped to design a hierarchical diffusion method that treated strings of a couple thousand nucleotides as “sentences.”
Training and developing models of this rigor and scale, of course, poses an enormous computational burden. The team brought to bear major supercomputers, including Argonne’s Polaris system (AMD Epyc “Rome” CPUs, Nvidia A100 GPUs, 25.8 Linpack petaflops), Nvidia’s Selene system (also Epyc Rome and A100s, 63.5 Linpack petaflops), and NERSC’s Perlmutter system (AMD Epyc “Milan” CPUs, Nvidia A100 GPUs, 70.9 Linpack petaflops).
However, the team also leveraged more experimental hardware – namely, four Cerebras Systems CS-2 systems, each powered by one of the company’s “Wafer-Scale Engine” chips, which are considered the largest computer chips ever built (with 850,000 7nm cores) and which are aimed squarely at specialization for deep learning. The systems, hosted at Argonne National Laboratory, were used both in standalone mode and as an interconnected cluster.
The researchers say that, using these tools, they were able to “develop some of the largest biological LMs … to date” – models with “2.5 and 25 billion trainable parameters[.]” The researchers scaled up to 4,096 GPUs for their science runs, utilizing what they’re calling an “aggregate of 1.54 zettaflops” of compute power over time. On Selene, the 25 billion-parameter model sustained 121 petaflops in mixed precision across those 4,096 GPUs with a peak of 850 petaflops. (See graphic below.)
Interestingly, the researchers found that Cerebras’ CS-2 systems performed remarkably well at this task. “We note … that these training runs frequently take >1 week on dedicated GPU resources (such as Polaris@ALCF),” the paper reads. “To enable training of the larger models on the full sequence length … we leveraged AI-hardware accelerators such as Cerebras CS-2, both in a stand-alone mode and as an inter-connected cluster, and obtained GenSLMs that converge in less than a day.” Further, they add: “All GenSLM training with full genomes on CS-2s converged within 12 hours.”
“Due to the time and resource constraints involved in training the 25 billion parameter model, we are unable to train this model to convergence,” said Kyle Hippe, a researcher at Argonne National Laboratory and co-author of the paper, during the finalist presentations at SC22 in Dallas. “But using the CS-2 Wafer-Scale [Engine] cluster, we are able to achieve convergence when training on the full SARS-CoV-2 genomes[.]”
“The CS-2 system allows for really fast time to solution,” Hippe added. “We were able to train from scratch on the SARS-CoV-2 genomes and reach convergence in less than a day. So this provides a really unique opportunity to keep pushing the next step of this workflow without necessarily having to train large models to convergence on other platforms.”
The work was also scaled up to 16 CS-2 systems at a Colovore facility. Hippe said that the model showed near-linear scaling to the full 16 systems across three models sizes, ranging from 250 million parameters to 25 billion parameters.
This is an impressive, high-profile, real-world win for the young company, which has delivered systems to several high-profile research customers over the past few years as it worked to establish a reputation for itself.
The SARS-CoV-2 version of GenSLM was able to both identify the genome sequences of SARS-CoV-2 variants and – more importantly – to predict potential mutations by generating nucleotide sequences. “Every gene was understood in the context of the entire genome,” said Andrew Feldman, co-founder and CEO of Cerebras Systems, in an interview with HPCwire. “And we did this with the initial virus and we predicted mutations – and the set of mutations we predicted included Omicron, included Delta.” (“Initial virus,” in this case, means that the model was only trained with data from the first year of the pandemic, encompassing the Alpha and Beta variants of SARS-CoV-2.)
Feldman, of course, also suggested that Cerebras’ chips were particularly suited for this work relative to GPUs thanks to the extreme amounts of memory that the Wafer-Scale Engine can dedicate to the large analysis sizes of GenSLMs. He also touted the CS-2 systems’ ability to handle sparsity better than GPUs. “We are way better at sparsity than [Nvidia is], not by a little bit but by orders of magnitude,” Feldman said. “So we can run models 90% sparse, we can handle sparsity that is structured or unstructured.” In sparse flops, he said, each CS-2 delivered around 750 petaflops; in dense flops, just 7.5.
In the preprint paper, the researchers identify areas for expansion for the work: integrating GenSLMs with popular protein structure prediction workflows like AlphaFold; incorporating experimental data to better constrain the model to variants of concern; and rigorously addressing issues like noise and bias in the data.