Research map / Architectures and efficiency / Quantization
ASTRA: Communication-Efficient Acceleration for Multi-Device Transformer Inference
Xiao Liu, Lijun Zhang, Deepak Ganesan, Hui Guan
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use mixed-precision attention with VQ codes for non-local tokens instead of full-precision sequence parallelism (ablation) | ASTRA · latency, 4 devices, 1024 tokens, 10 Mbps | latency speedup over SP: up to 171.82x | Improved |
| add Noise-Augmented Vector Quantization (lambda = 1.0) (ablation) | ASTRA VQ · validation accuracy under extreme compression | accuracy: +0.86% | Improved |
| use distributed class tokens instead of a single class token (ablation) | ASTRA ViT · CIFAR-100 and ImageNet-1K ablation | accuracy: +0.37% to +7.13% | Improved |
| increase VQ groups from 1 to 32 (ablation) | ASTRA · ImageNet-1K, ViT-Base | top-1 accuracy (%): 77.39 -> 80.28 | Improved |
| increase number of devices from 2 to 8 (ablation) | ASTRA 32 groups · CIFAR-100, ViT-Base | top-1 accuracy (%): 91.86 -> 91.14 | Hurt |
| apply ASTRA to zero-shot Wikitext-103 transfer | GPT2-M ASTRA 32 groups · Wikitext-103 zero-shot | perplexity: 43.22 -> 62.29 | Hurt |
| combine ASTRA with 4-bit post-training quantization (ablation) | ASTRA 32 groups, ViT-Base · ImageNet-1K | accuracy (%): 80.28 -> 79.78 | No clear effect |
| use ASTRA instead of multi-device baselines under low bandwidth | Llama-3-8B prefill, 4 devices · 10-100 Mbps | prefill latency speedup over BP Nb=4: 1.13-5.13x | Improved |
| increase device count from 2 to 8 with ASTRA 1 group at 20 Mbps (ablation) | ASTRA · latency speedup, 1024 tokens | speedup over single device: 1.72x -> 3.69x | Improved |
Closest papers by meaning
- IceFormer: Accelerated Inference with Long-Sequence Transformers on CPUs (ICLR 2024 · similarity 0.86)
- SwitchHead: Accelerating Transformers with Mixture-of-Experts Attention (NeurIPS 2024 · similarity 0.86)
- Ladder-Residual: Parallelism-Aware Architecture for Accelerating Large Model Inference with Communication Overlapping (ICML 2025 · similarity 0.86)
- StarTrail: Concentric Ring Sequence Parallelism for Efficient Near-Infinite-Context Transformer Model Training (NeurIPS 2025 · similarity 0.86)
- Understanding the Expressive Power and Mechanisms of Transformer for Sequence Modeling (NeurIPS 2024 · similarity 0.86)
- FrameQuant: Flexible Low-Bit Quantization for Transformers (ICML 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
