Research map / Large language models / Instruction and fine-tuning
Distributed Inference and Fine-tuning of Large Language Models Over The Internet
Alexander Borzunov, Max Ryabinin, Artem Chumachenko, Dmitry Baranchuk, Tim Dettmers, Younes Belkada, Pavel Samygin, Colin Raffel
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 fault-tolerant dual-cache algorithm instead of caching with restarts | BLOOM 7.1B inference · 128 tokens, failure rate 1e-2 | steps/s: 0.18 -> 3.38 | Improved |
| use fault-tolerant algorithm instead of cache-less inference | BLOOM 7.1B inference · 1024 tokens, failure rate 1e-2 | steps/s: 0.89 -> 2.17 | Improved |
| use fault-tolerant algorithm instead of caching with restarts when no failures occur | BLOOM 7.1B inference · 128 tokens, failure rate 0 | steps/s: 17.1 -> 11.4 | Hurt |
| use Petals distributed inference instead of RAM offloading | BLOOM-176B on 3xA100 · sequential inference seq len 128 | steps/s: 0.0495 -> 1.71 | Improved |
| use Petals distributed inference instead of RAM offloading | Llama 2 (70B) on 3xT4 · sequential inference seq len 128 | steps/s: 0.139 -> 2.29 | Improved |
| use Petals instead of offloading for parallel forward with slow network | BLOOM-176B on 10xRTX 3090 · parallel forward 64x128, 100 Mbit/s | tokens/s: 109.3 (offloading) vs 28.1 | Hurt |
| increase RTT from under 5 ms to 100 ms (ablation) | Petals on Llama 2 (70B) · sequential inference seq len 128, 100 Mbit/s | steps/s: 2.29 -> 1.57 | Hurt |
| reduce bandwidth from 1 Gbit/s to 100 Mbit/s (ablation) | Petals on BLOOM-176B · sequential inference seq len 128 | steps/s: 1.71 -> 1.66 | No clear effect |
| quantize BLOOM-176B weights to 8-bit (ablation) | BLOOM-176B · zero-shot avg of three benchmarks | accuracy: 70.1 -> 70.3 | No clear effect |
| quantize OPT-175B weights to 8-bit (ablation) | OPT-175B · zero-shot avg of three benchmarks | accuracy: 75.3 -> 74.9 | No clear effect |
| quantize weights to 8-bit (ablation) | BLOOM-176B · generation throughput batch size 1 | tokens/s: 4.18 -> 3.95 | Hurt |
| use prefix tuning instead of prompt tuning (ablation) | BLOOM-176B fine-tuning · forward pass, batch 8, prompt length 16 | tokens/s: 195.6 -> 111.0 | Hurt |
Closest papers by meaning
- Federated Full-Parameter Tuning of Billion-Sized Language Models with Communication Cost under 18 Kilobytes (ICML 2024 · similarity 0.90)
- FlowerTune: A Cross-Domain Benchmark for Federated Fine-Tuning of Large Language Models (NeurIPS 2025 · similarity 0.89)
- Ferret: Federated Full-Parameter Tuning at Scale for Large Language Models (ICML 2025 · similarity 0.89)
- Hybrid LLM: Cost-Efficient and Quality-Aware Query Routing (ICLR 2024 · similarity 0.89)
- Inference Scaling Laws: An Empirical Analysis of Compute-Optimal Inference for LLM Problem-Solving (ICLR 2025 · similarity 0.89)
- Scaling Laws Meet Model Architecture: Toward Inference-Efficient LLMs (ICLR 2026 · similarity 0.89)
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.
