When training large scale LLMs, there is a large assortment of parallelization strategies which you can employ to scale your training runs to work on more GPUs. There are already a number of good resources for understanding how to parallelize your models: I particularly recommend How To Scale Your Model and The Ultra-Scale Playbook. The purpose of this blog post is to discuss parallelization strategies in a more schematic way by focusing only on how they affect your device mesh. The device mesh is an abstraction used by both PyTorch and JAX that takes your GPUs (however many of them you’ve got in your cluster!) and organizes them into a N-D tensor that expresses how the devices communicate with each other. When we parallelize computation, we shard a tensor along one dimension of the mesh, and then do collectives along that dimension when there are nontrivial dependencies between shards. Being able to explain why a device mesh is set up the way it is for a collection of parallelization strategies is a good check for seeing if you understand how the parallelization strategies work in the first place! (Credit: This post was influenced by Visualizing 6D Mesh Parallelism.) — Read More

#training

We introduce INTELLECT-2, the first globally distributed reinforcement learning (RL) training run of a 32 billion parameter language model. Unlike traditional centralized training efforts, INTELLECT-2 trains a reasoning model using fully asynchronous RL across a dynamic, heterogeneous swarm of permissionless compute contributors.

To enable a training run with this unique infrastructure, we built various components from scratch: we introduce PRIME-RL, our training framework purpose-built for distributed asynchronous reinforcement learning, based on top of novel components such as TOPLOC, which verifies rollouts from untrusted inference workers, and SHARDCAST, which efficiently broadcasts policy weights from training nodes to inference workers.

Beyond infrastructure components, we propose modifications to the standard GRPO training recipe and data filtering techniques that were crucial to achieve training stability and ensure that our model successfully learned its training objective, thus improving upon QwQ-32B, the state of the art reasoning model in the 32B parameter range.

We open-source INTELLECT-2 along with all of our code and data, hoping to encourage and enable more open research in the field of decentralized training. — Read More

Next token prediction has been the standard training objective used in large language model pretraining. Representations are learned as a result of optimizing for token-level perplexity. We propose Continuous Concept Mixing (CoCoMix), a novel pretraining framework that combines discrete next token prediction with continuous concepts. Specifically, CoCoMix predicts continuous concepts learned from a pretrained sparse autoencoder and mixes them into the model’s hidden state by interleaving with token hidden representations. Through experiments on multiple benchmarks, including language modeling and downstream reasoning tasks, we show that CoCoMix is more sample efficient and consistently outperforms standard next token prediction, knowledge distillation and inserting pause tokens. We find that combining both concept learning and interleaving in an end-to-end framework is critical to performance gains. Furthermore, CoCoMix enhances interpretability and steerability by allowing direct inspection and modification of the predicted concept, offering a transparent way to guide the model’s internal reasoning process. — Read More

#training

Regression, the task of predicting a continuous scalar target y based on some features x is one of the most fundamental tasks in machine learning and statistics. It has been observed and theoretically analyzed that the classical approach, mean squared error minimization, can lead to suboptimal results when training neural networks. In this work, we propose a new method to improve the training of these models on regression tasks, with continuous scalar targets. Our method is based on casting this task in a different fashion, using a target encoder, and a prediction decoder, inspired by approaches in classification and clustering. We showcase the performance of our method on a wide range of real-world datasets. — Read More

We present DeepSeek-V3, a strong Mixture-of-Experts (MoE) language model with 671B total parameters with 37B activated for each token. To achieve efficient inference and cost-effective training, DeepSeek-V3 adopts Multi-head Latent Attention (MLA) and DeepSeekMoE architectures, which were thoroughly validated in DeepSeek-V2. Furthermore, DeepSeek-V3 pioneers an auxiliary-loss-free strategy for load balancing and sets a multi-token prediction training objective for stronger performance. We pre-train DeepSeek-V3 on 14.8 trillion diverse and high-quality tokens, followed by Supervised Fine-Tuning and Reinforcement Learning stages to fully harness its capabilities. Comprehensive evaluations reveal that DeepSeek-V3 outperforms other open-source models and achieves performance comparable to leading closed-source models. Despite its excellent performance, DeepSeek-V3 requires only 2.788M H800 GPU hours for its full training. In addition, its training process is remarkably stable. Throughout the entire training process, we did not experience any irrecoverable loss spikes or perform any rollbacks. The model checkpoints are available at this https URL. — Read More

#training

A few months ago, we published an interpretability paper demonstrating our ability to learn interpretable features that correspond to various concepts (e.g., famous individuals, types of computer code, etc.) represented in Claude 3 Sonnet. To verify our feature interpretations, we ran qualitative feature steering experiments, where we artificially dialed up and down various features to see if they changed model outputs in intuitive ways. The results were promising – for example, turning up a feature that responded to mentions of the Golden Gate Bridge made the model talk about the Golden Gate Bridge. Such examples led us to hypothesize that feature steering might be a promising way to modify model outputs in specific interpretable ways. — Read More

#training

Instruction tuning is critical to large language models (LLMs) for achieving better instruction following and task adaptation capabilities but its success heavily relies on the training data quality. Many recent methods focus on improving the data quality but often overlook the compatibility of the data with the student model being finetuned. This paper introduces Selective Reflection-Tuning, a novel paradigm that synergizes a teacher LLM’s reflection and introspection for improving existing data quality with the data selection capability of the student LLM, to automatically refine existing instruction-tuning data. This teacher-student collaboration produces high-quality and student-compatible instruction-response pairs, resulting in sample-efficient instruction tuning and LLMs of superior performance. Selective Reflection-Tuning is a data augmentation and synthesis that generally improves LLM finetuning and self-improvement without collecting brand-new data. We apply our method to Alpaca and WizardLM data and achieve much stronger and top-tier 7B and 13B LLMs. — Read More

#training

Earlier this year, Sakana AI started leveraging evolutionary algorithms to develop better ways to train foundation models like LLMs. In a recent paper, we have also used LLMs to act as better evolutionary algorithms!

Given these surprising results, we began to ask ourselves: Can we also use LLMs to come up with a much better algorithm to train LLMs themselves? We playfully term this self-referential improvement process LLM² (‘LLM-squared’) as an homage to previous fundamental work in meta-learning.

As a significant step towards this goal, we’re excited to release our report, Discovering Preference Optimization Algorithms with and for Large Language Models. — Read More