Communication compression isacrucial technique formodern distributedlearning systems to alleviate their communication bottlenecks over slower networks.

Large Language Models (LLMs) are now integral across various domains and have demonstrated impressive performance. Progress, however, rests on the premise that benchmark scores are both accurate and reproducible. We demonstrate that the reproducibility of LLM performance is fragile: changing system configuration, such as evaluation batch size, GPU count, and GPU version, can introduce significant differences in the generated responses. This issue is especially pronounced in reasoning models, where minor rounding differences in early tokens can cascade into divergent chains of thought, ultimately affecting accuracy. For instance, under bfloat16 precision with greedy decoding, a reasoning model like DeepSeek-R1-Distill-Qwen-7B can exhibit up to 9% variation in accuracy and 9,000 tokens difference in response length due to differences in GPU count, type, and evaluation batch size. We trace the root cause of this variability to the non-associative nature of floating-point arithmetic under limited numerical precision. This work presents the first systematic investigation into how numerical precision affects reproducibility in LLM inference. Through carefully controlled experiments across various hardware, software, and precision settings, we quantify when and how model outputs diverge. Our analysis reveals that floating-point precision - while critical for reproducibility - is often neglected in evaluation practices. Inspired by this, we develop a lightweight inference pipeline, dubbed LayerCast, that stores weights in 16-bit precision but performs all computations in FP32, balancing memory efficiency with numerical stability. Code is available at https://github.com/nanomaoli/llm_reproducibility.

The pursuit of computational efficiency has driven the adoption of low-precision formats for training transformer models. However, this progress is often hindered by notorious training instabilities. This paper provides the first mechanistic explanation for a long-standing and unresolved failure case where training with flash attention in low-precision settings leads to catastrophic loss explosion. Our in-depth analysis reveals that the failure is not a random artifact but caused by two intertwined phenomena: the emergence of similar low-rank representations within the attention mechanism and the compounding effect of biased rounding errors inherent in low-precision arithmetic. We demonstrate how these factors create a vicious cycle of error accumulation that corrupts weight updates, ultimately derailing the training dynamics. To validate our findings, we introduce a minimal modification to the flash attention that mitigates the bias in rounding errors. This simple change stabilizes the training process, confirming our analysis and offering a practical solution to this persistent problem. The pursuit of training ever-larger and more powerful transformer models is a relentless drive for computational efficiency (Brown et al., 2020; Hoffmann et al., 2022). A key strategy in this endeavor is the adoption of low-precision numerical formats (Micikevicius et al., 2017; Wang et al., 2018; Kalamkar et al., 2019; Liu et al., 2024), which promise substantial reductions in memory footprint and significant boosts in training speed. In industrial practice, it is common to use BF16 for memory-bound operations like flash attention while pushing compute-bound operations like FFNs to even lower precisions such as FP8 (Liu et al., 2024; Qwen-Team, 2025). This highlights the heightened sensitivity of attention mechanisms to numerical precision.

While ensembling deep neural networks has shown promise in improving generalization performance, scaling current ensemble methods for large models remains challenging. Given that recent progress in deep learning is largely driven by the scale, exemplified by the widespread adoption of large-scale neural network architectures, scalability emerges an increasingly critical issue for machine learning algorithms in the era of large-scale models. In this work, we first showcase the potential of low precision ensembling, where ensemble members are derived from a single model within low precision number systems in a training-free manner. Our empirical analysis demonstrates the effectiveness of our proposed low precision ensembling method compared to existing ensemble approaches.

Training Large Language Models (LLMs) is costly in terms of energy, hardware, and annotated data, often resulting in a positionality rooted in predominant cultures and values (Santy et al., 2023). Domain adaptation has emerged as a promising strategy to better align models with diverse cultural and value contexts (Hershcovich et al., 2022), but its computational cost remains a significant barrier, particularly for research groups lacking access to large-scale infrastructure. In this paper, we evaluate how the use of different numerical precision formats and data parallelization strategies impacts both training speed (as a proxy to energy and hardware consumption) and model accuracy, with the goal of facilitating domain adaptation in low-resource environments. Our findings are relevant to any setting where energy efficiency, accessibility, or limited hardware availability are key concerns.

Model merging enables efficient multi-task models by combining task-specific fine-tuned checkpoints. However, storing multiple task-specific checkpoints requires significant memory, limiting scalability and restricting model merging to larger models and diverse tasks. In this paper, we propose quantizing task vectors (i.e., the difference between pre-trained and fine-tuned checkpoints) instead of quantizing fine-tuned checkpoints. We observe that task vectors exhibit a narrow weight range, enabling low precision quantization (up to 4 bit) within existing task vector merging frameworks. To further mitigate quantization errors within ultra-low bit precision (e.g., 2 bit), we introduce Residual Task Vector Quantization, which decomposes the task vector into a base vector and offset component. We allocate bits based on quantization sensitivity, ensuring precision while minimizing error within a memory budget. Experiments on image classification and dense prediction show our method maintains or improves model merging performance while using only 8% of the memory required for full-precision checkpoints.