Research map / Large language models / Agents, code and math
DITRON: Distributed Multi-level Tiling Compiler for Parallel Tensor Programs
Size Zheng, Xuegui Zheng, Hanshi Sun, Qi Hou, Wenlei Bao, Shiyu Li, Haojie Duanmu, Jin Fang, Chenli Xue, Chenhui Huang, Yuanqiang Liu, Renze Chen and 7 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 |
|---|---|---|---|
| add cross-rank swizzle (ablation) | DITRON GEMM+AllReduce · Llama 8xH800 | tokens/s: 11,605 -> 12,174 | Improved |
| add multi-level tiling (ablation) | DITRON GEMM+AllReduce · Llama 8xH800 | tokens/s: 11,706 -> 12,174 | Improved |
| add computation-communication overlap (ablation) | DITRON GEMM+AllReduce · Llama 8xH800 | tokens/s: 10,320 -> 12,174 | Improved |
| add cross-rank swizzle (ablation) | DITRON AG+GEMM / GEMM+RS · Llama 8xH800 | tokens/s: 10,642 -> 10,928 | Improved |
| add multi-level tiling (ablation) | DITRON AG+GEMM / GEMM+RS · Llama 8xH800 | tokens/s: 9,494 -> 10,928 | Improved |
| add overlap with AG+RS (ablation) | DITRON · Llama 8xH800 | tokens/s: 10,320 -> 10,928 | Improved |
| use DITRON in vLLM at batch size below 128 | vLLM · Qwen3-32B/LLaMA3-70B | throughput | Hurt |
| use DITRON in vLLM at batch size above 128 | vLLM · Qwen3-32B/LLaMA3-70B | throughput: 5%-30% speedup | Improved |
| use AllReduce instead of AG+RS for attention module | DITRON · Qwen3-32B attention | speedup | Improved |
| use AG+RS instead of AllReduce for FFN beyond 2k tokens | DITRON · Qwen3-32B FFN, 128k tokens | speedup: 1.17x -> 1.27x | Improved |
| scale TP workloads across nodes (8-32 GPUs) | DITRON AG-GEMM/GEMM-RS · strong scaling | speedup over CuBLAS+NCCL: 0.80x to 1.71x | Mixed |
| distributed MegaKernel for single-batch inference | DITRON · Qwen3/LLaMA3 8xH800 | latency per token: 6.28x vs Torch Eager | Improved |
Closest papers by meaning
- SpeedLoader: An I/O efficient scheme for heterogeneous and distributed LLM operation (NeurIPS 2024 · similarity 0.88)
- SPD: Sync-Point Drop for Efficient Tensor Parallelism of Large Language Models (ICML 2025 · similarity 0.88)
- Synergistic Tensor and Pipeline Parallelism (NeurIPS 2025 · similarity 0.87)
- Ladder-Residual: Parallelism-Aware Architecture for Accelerating Large Model Inference with Communication Overlapping (ICML 2025 · similarity 0.86)
- Cost-Efficient LLM Training with Lifetime-Aware Tensor Offloading via GPUDirect Storage (NeurIPS 2025 · similarity 0.86)
- Practical Offloading for Fine-Tuning LLM on Commodity GPU via Learned Sparse Projectors (AAAI 2025 · similarity 0.86)
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.
