LLMs on Aurora: Overview

Sam Foreman

ALCF INCITE GPU HACKATHON
May 20–22, 2025

LLMs on Aurora: 🌌 AuroraGPT

Sam Foreman

2025-05-21

ALCF Incite Hackathon 2025

2025 ALCF INCITE GPU Hackathon (20-May 22, 2025)
LLMs on Aurora¹:
- 🍋 Hands-On: ezpz
- 🌌 Overview: AuroraGPT

🎯 AuroraGPT: Goals

AuroraGPT: General purpose scientific LLM
Broadly trained on a general corpora plus scientific
{papers, texts, data}

Explore pathways towards a “Scientific Assistant” model
Build with international partners (RIKEN, BSC, others)
Multilingual English, 日本語, French, German, Spanish
Multimodal: images, tables, equations, proofs, time series, graphs, fields, sequences, etc

Figure 1: Image from Hannibal046 / `Awesome-LLM`

Figure 2: Credit to the entire AuroraGPT team for slides.

🦙 Issues with “Publicly Available” LLMs

Trust and Safety:
- Skepticism about deployment in critical infrastructure
- Correctness and reliability of model outputs
Transparency:
- Data governance, what was used for pre-training? fine-tuning?
  - generally unknown
- What is open source?
  - Model weights?
  - Pre-training {code, logs, metrics} ?

🧪 AuroraGPT: Open Science Foundation Model

Figure 3: High-level overview of AuroraGPT project

AuroraGPT will be a publicly distributed, open source foundation model for open science
Is being trained on:
- Scientific / engineering structured data
- General text, media, news, etc.
- Large amounts of low to medium quality data
- Much less high quality data (that is publicly available for use)
This data is then cleaned, processed, de-duplicated and used for the initial pre-training phase of the model
The vast majority of the overall compute is spent during this initial pre-training phase
- This is the group I help to lead and will be talking a bit about today
The initial pre-training phase is currently underway
- Eventually, given a bit of time, effort and magic, the model will be ready for fine-tuning and additional training for a variety of downstream tasks
The pretrained model will then be handed off for additional fine-tuning on a variety of downstream tasks
- Scientific discovery
- Accelerate scientific tasks
- Digital twins
- Inverse design
- Code optimization
- Accelerated simulations
- Autonomous experiments
- Co-design
Becoming increasingly clear that LLMs have the potential to drastically accelerate computational science
- We’ve seen this already for {GenSLMs, Weather / Climate / Earth Systems Modeling, Particle Physics, etc.}

📊 AuroraGPT: Outcomes

Datasets and data pipelines for preparing science training data
Software infrastructure and workflows to train, evaluate and deploy LLMs at scale for scientific resarch purposes
- argonne-lcf/Megatron-DeepSpeed
  End-to-end training and inference, on any GPU cluster
- argonne-lcf/inference-endpoints
  Inference endpoints for LLMs, hosted @ ALCF
Evaluation of state-of-the-art LLM Models:
- Determine where they fall short in deep scientific tasks
- Where deep data may have an impact

📚 What do we hope to get?

Assessment of the approach of augmenting web training data with two forms of data specific to science:
- Full text scientific papers
- Structured scientific datasets (suitably mapped to narrative form)
Research grade artifacts (models) for scientific community for adaptation for downstream uses¹
Promotion of responsible AI best practices where we can figure them out
International Collaborations around the long term goal of AGI for science

🌌 Aurora

Table 1: Aurora Specs


Racks	166
Nodes	10,624
CPUs	21,248
GPUs	63,744
NICs	84,992
HBM	8 PB
DDR5c	10 PB

🏆 Fastest AI system in the world

🤖 ALCF AI Testbed

ALCF AI Testbed Systems are in production and available for allocations to the research community
Significant improvement in time-to-solution and energy-efficiency for diverse AI for science applications.
NAIRR Pilot

Up to $≈$ 25 $\times$ throughput improvement for genomic FMs with 6.5 $\times$ energy efficiency

Figure 5: **SambaNova SN-30** 2nd Gen, 8 nodes with 64 AI Accelerators

Figure 6: **Graphcore Bow**: Pod-64 configuration with 64 accelerators

Figure 7: **Cerebras**: 2x CS-2 WSE with Memory-X and Swarm-X technologies

Figure 8: **GroqRack**: 9 nodes, 8 GroqChip v1.5 Tensor streaming processors accelerators per node

👥 Team Leads

Planning

Data

Training

Evaluation

Post

Inference

Comms

Distribution

🤝 Teams

Planning
Data Prep
- Accumulate 20+ T tokens of high-quality scientific text and structured data
Models / Training¹
- Train (entirely from scratch) a series of models on publicly available data
Evaluation
- Skills, trustworthiness, safety, robustness, privacy, machine ethics

Post-Training
- Fine-tuning, alignment
Inference
- Model serving, API development / public-facing web services
Distribution
- Licensing, generating and distributing artifacts for public consumption
Communication

📚 Data

✅ Goal: Assemble a large corpus of documents (general and scientific) to train and fine-tune AuroraGPT models

Challenges: Avoid / detect contamination with benchmarks
- Respect copyright (ACM Digital Library), privacy, and ethical considerations
Performance Challenges: High throughput data processing
- Converting PDF $\rightarrow$ text (math formula, figures)
- Convert science information (data) into text (narratives)
- De-duplication (syntactic and semantic) of scientific documents (to avoid memorization, bias)
Quantity: Considering 20+ Trillion tokens $\rightarrow\approx$ 100M papers
Domains: All (long-term) scientific domains, starting with:
- Material science, Physics, Biology, Computer Science, Climate Science

⏱️ Dataset Processing

To train a fixed model on trillions of tokens requires:
1. Aggregating data from multiple different corpora
  (e.g. ArXiv, Reddit, StackExchange, GitHub, Wikipedia, etc.)
2. Sampling each training batch according to a fixed distribution across corpora
3. Building indices that map batches of tokens into these files (indexing)
The original implementation was slow:
- Designed to run serially on a single device
- Major bottleneck when debugging data pipeline at scale

🚀 Accelerating Dataset Processing: Results

Original implementation:
- Slow!
- 🐌 ~ 1 hr/2T tokens
Fix:
- Wrote asynchronous, distributed implementation
- significantly improves performance (30x !!)
- 🏎️💨 ~ 2 min/2T tokens

Figure 9: Time spent preparing 2T tokens

🦜 Model Training

✅ Goals

Want training runs at scale to be:
- efficient
- stable
- reproducible
This requires:
- robust data pipelines / file IO
- effectively overlapping compute with communication
- stability across {network, filesystem, machine}
3D / Multi-dimensional Parallelism strategies
Large batch training
Second order optimizers
Sub-quadratic attention
State space models
Highly optimized GPU kernels

❌ Challenges

Looong time to train, can be:
- weeks (even months) of continuous training
- order of magnitude longer than typical NN training jobs
Stability issues:
- failures are expensive (but inevitable)
- stragglers common at scale
Individual jobs are:
- fragile
- only as good as the worst rank
- one hang or bad worker can crash job
- network / filesystem / other-user(s) dependent
Cost / benefits of different collective communication algorithms
- depend on optimized / efficient implementations
Network performance
Highly optimized GPU kernels

📉 Loss Curve: Training AuroraGPT-7B on 2T Tokens

Figure 10: Loss curve during training on 2T tokens.

🤔 Evaluating FM Skills for Science

What to measure?
- Knowledge Extraction, Retrieval, Distillation, Synthesis: LLM is provided a question or instruction and a truthful answer is expected
- Text Grounded: Answers are expected to be fully grounded on peer-reviewed references to support responses
- Reasoning: LLMs are expected to solve deductive (prove a theory or hypothesis from formal logic and observations), inductive (validate / explain observations from theories) problems
- Creativity: A creative answer is expected from a question or instruction
  - thoughtful dialogue, coding, etc.

⚖️ Evaluating FM Skills for Science: Criteria

Criteria for all of the above:
- Correctness of facts
- Accuracy of solutions and inferences
- Reliability consistently good in quality or performance
- Speed how fast to produce a response
- # shots how many examples are needed for good quality
  - Extent of prompt engineering

🧬 MProt-DPO: Scaling Results

Figure 11: Scaling results for `3.5B` model across ~38,400 GPUs

~ 4 EFLOPS @ Aurora
38,400 XPUs
= 3200 [node] x 12 [XPU / node]
🔔 Gordon Bell Finalist¹:
- MProt-DPO: Breaking the ExaFLOPS Barrier for Multimodal Protein Design Workflows

📓 References

argonne-lcf / Megatron-DeepSpeed
For the largest of large language models.
saforem2 / ezpz
Distributed training, ezpz. 🍋
📊 See my other slides at samforeman.me/talks:

👀 See also:

❤️ Thank you!

Organizers
Feel free to reach out!

🙏 Acknowledgements

This research used resources of the Argonne Leadership Computing Facility, which is a DOE Office of Science User Facility supported under Contract DE-AC02-06CH11357.

📑 Bibliography

Refs:
- Wei et al. (2022)
- Animations from The Illustrated Transformer

Dharuman, Gautham, Kyle Hippe, Alexander Brace, Sam Foreman, Väinö Hatanpää, Varuni K. Sastry, Huihuo Zheng, et al. 2024. “MProt-DPO: Breaking the ExaFLOPS Barrier for Multimodal Protein Design Workflows with Direct Preference Optimization.” In Proceedings of the International Conference for High Performance Computing, Networking, Storage, and Analysis. SC ’24. Atlanta, GA, USA: IEEE Press. https://doi.org/10.1109/SC41406.2024.00013.

Hosseini, Ryien, Filippo Simini, Venkatram Vishwanath, Rebecca Willett, and Henry Hoffmann. 2025. “Quality Measures for Dynamic Graph Generative Models.” In The Thirteenth International Conference on Learning Representations. https://openreview.net/forum?id=8bjspmAMBk.

Song, Shuaiwen Leon, Bonnie Kruft, Minjia Zhang, Conglong Li, Shiyang Chen, Chengming Zhang, Masahiro Tanaka, et al. 2023. “DeepSpeed4Science Initiative: Enabling Large-Scale Scientific Discovery Through Sophisticated AI System Technologies.” https://arxiv.org/abs/2310.04610.

Wei, Jason, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, et al. 2022. “Emergent Abilities of Large Language Models.” https://arxiv.org/abs/2206.07682.

Yang, Jingfeng, Hongye Jin, Ruixiang Tang, Xiaotian Han, Qizhang Feng, Haoming Jiang, Bing Yin, and Xia Hu. 2023. “Harnessing the Power of LLMs in Practice: A Survey on ChatGPT and Beyond.” https://arxiv.org/abs/2304.13712.

🧬 MProt-DPO: Scaling Results

🚂 Loooooooooong Sequence Lengths

Working with Microsoft/DeepSpeed team to enable longer sequence lengths (context windows) for LLMs
- See my blog post for additional details

Figure 15: **Pre-training**: Virtually all of the compute used during pretraining phase

🍎 Training LLMs

Figure 18: Visualization from Yang et al. (2023)