About MLPerf Training

Pretraining Large Language Models (LLMs) is the initial phase of LLM training where the model learns to understand and generate human-like text by training the model on vast amounts of unstructured textual data. A pre-trained model can then be fine-tuned on domain-specific datasets, which gives the model knowledge of specific tasks. Reasoning approaches have become popular recently, helping models better explain their thinking process. Because pre-training uses large amounts of data, it is generally the most compute-intensive phase of LLM training. 

MLCommons® added a GPT3 Pretraining benchmark to MLPerf® Training in 2022 to help academics and industry experts optimize LLM pretraining. It has been a huge success, as we have seen over 3x speedups on the results of this benchmark for large systems (over 10,000 accelerators).

Over this same period, newer state-of-the-art models have demonstrated greater scale and many architectural advancements. For example, Google’s PaLM model, released in 2022, contains 540B parameters, three times as many as GPT3’s 175B parameter count. Meta’s Llama 2 and Llama 3 series, released in 2023, utilize new algorithms such as Root Mean Square Layer Normalization (RMSNorm, Zhang et al., https://arxiv.org/abs/1910.07467), Rotary Position Embedding (RoPE, Su et al., 2024), and Grouped Query Attention (GQA, Ainslie et al., 2023). 

A MLPerf Training working group task force was created to investigate a July 2024 proposal to  replace GPT3 with Llama 3.1 405B. Task force members evaluated the proposal and industry options to recommend the best new pretraining reference. They agreed that Llama 3.1 405B would best represent the current state-of-the-art pretraining and adopted it as a new benchmark. Details on the review and selection process are included in this blog.

Model selection

The most important factor the task force considered when constructing this new pretraining benchmark was the model architecture. As mentioned above, the current GPT3 architecture does not include some of the recent algorithm updates and therefore is not competitive with other state-of-the-art models such as Nemotron-4, GPT4.5 and Claude 3.7 Sonnet. However, Meta’s Llama 3.1 405B model, with an architecture and scale similar to these top-tier models, demonstrates on-par performance with these models across multiple quality benchmarks, thus positioning itself as a competitive model representative of the current state-of-the-art models. 

Choosing a model with high community engagement was also an important consideration, and Llama 3.1 405B is the perfect candidate on that basis. With more than 300 million total downloads of all Llama versions to date, the Llama family of models is very widely adopted. The MLPerf community, for instance, has already included Llama 2 in the LoRA benchmark. Additionally, Meta’s released technical paper details the training process, model architecture, dataset and hyperparameters, which allowed us to easily create a high-quality benchmark based on their work.

One other requirement for the new MLPerf Training benchmark is that the model architecture must be publicly accessible so that all submitters can download it and reproduce the results on their end. This typically requires a permissive license. Thanks to Meta’s support, the current Llama 3.1 405B model is now available to all MLCommons members for benchmarking. 

Dataset selection

To help submitters transition easily from GPT3 to Llama 3.1 405B, we use the same AllenAI C4-en dataset (Colossal Cleaned Common Crawl dataset) in both benchmarks. This dataset contains 365M English-language paragraphs scraped from the internet, cleaned and preprocessed to convenient JSON formats for training the model. 

One key difference is that Llama 3.1 405B uses a context length of 8,192, which is 4x larger than the 2,048-token context of GPT3. The context length represents the amount of text that a model can process at once, thus a larger context length leads to better understanding. 

Benchmark technical details

The reference code is implemented in the NVIDIA NeMo Framework, an open-sourced and scalable AI framework built for large language models. By offering easy-to-use modular components and pre-trained model recipes, NeMo enables developers to efficiently build performant models tailored to specific use cases. The reference code functionality is tested on NVIDIA H100 GPUs. 

To ensure that the new benchmark model setup is easily understandable without complex knobs and hyperparameters scattered around the training script, it starts with NeMo’s integrated Llama 3.1 405B training recipe, which closely followed the original paper’s hyperparameters and introduced only the minimum necessary changes to our script, such as adjusting the optimizer’s learning rates. 

When researching how to construct this benchmark, we first explored the GPT3-style setup. For the GPT3 pretraining benchmark, the task force pretrained from random weights for about 12.6T tokens to generate a stable customized checkpoint. Submitters were required to continue pretraining from this checkpoint until they reached a predetermined quality metric. This methodology was used to ensure that the GPT3 pretraining benchmark showcased a reproducible slice of the actual GPT3 pretraining process.

Applying the same methodology to the Llama 3.1 405B pretraining benchmark, the task force generated a customized checkpoint trained on 32B tokens. But experiment results showed that resuming from this checkpoint had more variation in the convergence curves than we could tolerate. This meant that resuming from a custom-trained checkpoint in the benchmark was not feasible. The interpretation that we were still in very early phases of training after the 32B tokens, meant that the checkpoint generated was still quite unstable.

So the taskforce decided to train from the fully trained and stable HuggingFace checkpoint by giving it some new information, which led to a much more stable convergence trend. The benchmark now begins from the checkpoint and stops when the evaluation log perplexity reaches our defined target.

One noticeable difference between the task force’s implementation and Meta’s implementation is that we replaced the original Tiktokenizer, which used a vocab size of 128k, with the 32k vocab size Mixtral 8x22b tokenizer. The motivation behind this change was that Meta’s checkpoint is already well-trained, so it will not learn anything new on the current C4 dataset. Replacing the tokenizer forces the model checkpoint to adapt to a new, different token distribution, therefore continuing to pretrain on the current dataset. This difference implies that, when we resume from the pre-trained checkpoint, only the first 32,000 rows of the 128,256-row word embedding layer weights will be loaded to the model.

As mentioned before, the Llama 3.1 405B pretraining benchmark is a replacement for the GPT3 pretraining benchmark, and the following components will be unchanged between the two:

  • Validation dataset: To reduce the cost of validation, we chose to create a customized subset of the full C4 validation dataset so that the evaluation time would not be significant compared to training time. Specifically, we found out that using only 5,760 sequences is sufficient to measure model convergence, and since only 91,205 validation samples are needed to yield the required 5,760 validation sequences, the first 91,205 unshuffled rows of the full C4 validation dataset were selected as the customized validation dataset for this benchmark.
  • Loss and target accuracy metric: As with GPT3, log perplexity is used, computed on the customized validation dataset to evaluate the model’s convergence behavior. This metric is not costly to compute and is widely adopted as a versatile indicator of evaluating LLMs.
  • Limit number of submission logs: We recognize that the Llama 3.1 405B pretraining benchmark is very expensive to run, so to ensure fairness while reducing submitters costs, each submitter is required to only run the benchmark 3 times, the same as the GPT3 pretraining benchmark. The median result is then reported.
  • Restriction on hyperparameter searching: Because running the benchmark is expensive, we believe it is reasonable to disallow hyperparameter searches and borrowing. Most hyperparameters have been given static values. For a very few others, such as the learning rate, formulas are provided to compute them based on the global batch size.

Conclusion

In this blog post, we have provided an overview of a new LLM pretraining benchmark based on Meta’s Llama 3.1 405B model with twice as many parameters and four times as many training tokens as GPT3. This benchmark is massively more computationally intensive and contains its own new challenges. It is the task force’s belief that this new benchmark will encourage more innovation in the area of LLM pretraining.

About MLCommons

MLCommons is the world’s leader in AI benchmarking. An open engineering consortium supported by over 125 members and affiliates, MLCommons has a proven record of bringing together academia, industry, and civil society to measure and improve AI. The foundation for MLCommons began with the MLPerf benchmarks in 2018, which rapidly scaled as a set of industry metrics to measure machine learning performance and promote transparency of machine learning techniques. Since then, MLCommons has continued using collective engineering to build the benchmarks and metrics required for better AI – ultimately helping to evaluate and improve AI technologies’ accuracy, safety, speed, and efficiency.

For additional information on MLCommons and details on becoming a member, please visit MLCommons.org or email [email protected].