Sarthak Joshi

Julien Abadji, Marah Abdin, Connor Adams et al.

We present Laguna M.1 and Laguna XS.2, two Mixture-of-Experts foundation models built for long-horizon, agentic coding: M.1 has $225.8$B total parameters ($23.4$B activated per token) and XS.2 has $33.4$B total ($3$B activated). Both models were trained from scratch end-to-end inside the same internal system that we refer to as our Model Factory: a tightly-integrated stack of versioned data, training, evaluation, and inference components that turn model development into an industrial process. We describe the principles and design choices of the Model Factory and also detail the end-to-end training process of our models, throughout pre-training data and architecture, post-training stages, evaluation, and quantization. On agentic software engineering and terminal benchmarks (SWE-bench Verified, SWE-bench Multilingual, SWE-Bench Pro, and Terminal-Bench 2.0) M.1 and XS.2 are competitive with state-of-the-art open models in their respective weight classes. Laguna XS.2 weights are released under Apache~2.0 at https://huggingface.co/collections/poolside/laguna-xs2.

Sarthak Joshi, Sathish Vadhiyar

Faults in high-performance systems are expected to be very frequent in the current exascale computing era. To compensate for a higher failure rate, the standard checkpoint/restart technique would need to create checkpoints at a much higher frequency, resulting in an excessive amount of overhead, which would not be sustainable for many scientific applications. To improve application efficiency in such high-failure environments, the mechanism of replication of MPI processes was proposed. Replication allows for fast recovery from failures by simply dropping the failed processes and using their replicas to continue the regular operation of the application. In this paper, we have implemented FTHP-MPI (Fault Tolerance and High Performance MPI), a novel fault-tolerant MPI library that augments checkpoint/restart with replication to provide resilience from failures. The novelty of our work is that it is designed to provide fault tolerance in a native MPI library that does not provide support for fault tolerance. This lets application developers achieve fault tolerance at high failure rates while also using efficient communication protocols in the native MPI libraries that are generally fine-tuned for specific HPC platforms. We have also implemented efficient parallel communication techniques that involve replicas. Our framework deals with the unique challenges of integrating support for checkpointing and partial replication. We conducted experiments with three applications, HPCG, PIC, and CloverLeaf. We show that, for large-scale systems where failure intervals are expected to be within an hour, our replication-based library achieves higher efficiency and performance than checkpoint-based approaches. We show that, under failure-free conditions, the additional overheads from replication are negligible in our library.

Sarthak Joshi, Sathish Vadhiyar

As we have entered Exascale computing, the faults in high-performance systems are expected to increase considerably. To compensate for a higher failure rate, the standard checkpoint/restart technique would need to create checkpoints at a much higher frequency resulting in an excessive amount of overhead which would not be sustainable for many scientific applications. Replication allows for fast recovery from failures by simply dropping the failed processes and using their replicas to continue the regular operation of the application. In this paper, we have implemented PartRePer-MPI, a novel fault-tolerant MPI library that adopts partial replication of some of the launched MPI processes in order to provide resilience from failures. The novelty of our work is that it combines both fault tolerance, due to the use of the User Level Failure Mitigation (ULFM) framework in the Open MPI library, and high performance, due to the use of communication protocols in the native MPI library that is generally fine-tuned for specific HPC platforms. We have implemented efficient and parallel communication strategies with computational and replica processes, and our library can seamlessly provide fault tolerance support to an existing MPI application. Our experiments using seven NAS Parallel Benchmarks and two scientific applications show that the failure-free overheads in PartRePer-MPI when compared to the baseline MVAPICH2, are only up to 6.4% for the NAS parallel benchmarks and up to 9.7% for the scientific applications.