Handling communication overhead in large-scale tensor-parallel training remains a critical challenge due to the dense, near-zero distributions of intermediate tensors, which exacerbate errors under frequent communication and introduce significant computational overhead during compression. To this end, we propose TACO (Tensor-parallel Adaptive COmmunication compression), a robust FP8-based framework for compressing TP intermediate tensors. First, we employ a data-driven reshaping strategy combined with an Adaptive Scale–Hadamard Transform to enable high-fidelity FP8 quantization, while its Dual-Scale Quantization mechanism ensures numerical stability throughout training. Second, we design a highly fused compression operator to reduce memory traffic and kernel launch overhead, allowing efficient overlap with communication. Finally, we integrate TACO with existing state-of-the-art methods for Data and Pipeline Parallelism to develop a compression-enabled 3D-parallel training framework. Detailed experiments on GPT models and Qwen model demonstrate up to 1.87 × end-to-end throughput improvement while maintaining near-lossless accuracy, validating the effectiveness and efficiency of TACO in large-scale training.

Collective communication is critical to scaling large language model (LLM) training across various parallelism strategies, including data, tensor, and pipeline parallelism on GPU clusters. However, as model sizes and training scales increase, communication overhead is emerging as a major performance bottleneck. While compression is a promising mitigation strategy, existing solutions often lack user-transparency, hinder deployment and extensibility, and are not co-designed with communication algorithms. To address these limitations, we present COCCL, a high-performance collective communication library built on top of NCCL. COCCL introduces a novel programming model that can easily integrate compression into communication workflows with flexible configurability. It features a suite of compression-aware collective algorithms and runtime overlap mechanisms that mitigate error propagation and reduce computational overhead. We integrate well-established compression techniques into COCCL and tune the compression configurations during 3D-parallel training on GPT and Qwen models with up to 7 billion parameters. Using the optimal configuration (COCCL-3D), we achieve 1.24× throughput improvement while maintaining training accuracy.