DeServe is a offline-serving framework for decentralized inference of large language models. Benefiting from following technologies, DeServe can achieve up to 12.6x throughput improvement in high-latency network compared to the vLLM with pipeline parallelism. Following features are key to the performance:
- KV Cache Swapping: Maximizes GPU computation utilization by enlarging the KV cache size through swapping microbatch memory between CPU and GPU.
- Microbatch Scheduling: Allocates microbatches inside the pipeline for different network latency to maximize the throughput.
| Latency (ms) | real-world centralized | real-world east-west | sim 16 | sim 32 | sim 64 | sim 256 |
|---|---|---|---|---|---|---|
| vLLM (tp) | 253.0 | failed | / | / | / | / |
| vLLM (pp) | 89.1 | 37.3 | 68.8 | 55.3 | 36.1 | / |
| DeServe (pp) | 194.6 | 138.4 | 182.3 | 163.7 | 133.7 | / |
| DeServe (opt) | 445.2 | 434.1 | 458.5 | 457.3 | 456.8 | 442.9 |
To start the experiments, please refer to deserve_exp/readme.md.