Research map / Architectures and efficiency / RNNs and normalization
CONTINUUM: Restoring the Contiguous Tensor Abstraction Efficiently for Dynamic AI Workloads via Hardware Virtualization
Yangyu Zhang, Shuoming Zhang, Chunwei Xia, Shuaijiang Li, Zhicheng Li, Ruiyuan Xu, Zheming Yang, Lei Chen, YUAN WEN, Guangli Li, Xiaobing Feng, Huimin Cui and 1 more
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 |
|---|---|---|---|
| replace standard CUDA VMM control plane with batched transactional PTE updates in driver | GPU VMM mapping · mapping 2 MB pages up to 2GB | mapping latency: up to 671.3× speedup, ms to us | Improved |
| use elastic on-demand adapter mapping instead of fixed adapter slots | multi-LoRA serving vs vLLM · Llama-2-7B, 1 to 16 adapters | throughput: up to 3.29× | Improved |
| dispatch to cuBLAS on virtually contiguous adapters instead of S-LoRA Triton kernels | multi-LoRA serving · Llama-2-7B | throughput speedup: 1.16–1.32× | Improved |
| zero-copy aliasing for tree branching instead of vLLM block-table reconstruction | tree decoding · batch size 1, K=8, Lstep=4 | latency speedup: 7.65× | Improved |
| zero-copy aliasing with token-by-token branching | tree decoding · batch size 1, Lstep=1 | latency speedup over vLLM: 0.97× to 1.03× | No clear effect |
| increase Lstep (sparser branching) from 4 to 256 (ablation) | CONTINUUM tree decoding · batch size 1, K=8 | latency speedup over vLLM: 7.65× -> 3.64× | Hurt |
| use driver-level VMM backend instead of standard cuMemMap (Ours w/o VMM) (ablation) | CONTINUUM · context caching TTFT, 4K-64K | TTFT | Improved |
| virtual contiguity with on-demand remapping instead of PyTorch contiguous allocation | Engram on DeepSeekMoE-16B config · request rate sweep | throughput | Improved |
Closest papers by meaning
- SpeedLoader: An I/O efficient scheme for heterogeneous and distributed LLM operation (NeurIPS 2024 · similarity 0.88)
- Cost-Efficient LLM Training with Lifetime-Aware Tensor Offloading via GPUDirect Storage (NeurIPS 2025 · similarity 0.87)
- Beyond the Trial-and-Error Loop: Hybrid Projection and Automated Tuning for Distributed Training (NeurIPS 2026 · similarity 0.86)
- 70% Size, 100% Accuracy: Lossless LLM Compression for Efficient GPU Inference via Dynamic-Length Float (DFloat11) (NeurIPS 2025 · similarity 0.85)
- Practical Offloading for Fine-Tuning LLM on Commodity GPU via Learned Sparse Projectors (AAAI 2025 · similarity 0.85)
- SPD: Sync-Point Drop for Efficient Tensor Parallelism of Large Language Models (ICML 2025 · similarity 0.84)
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.
