GLM-5.2 RL weight transfer in 4 seconds using NIXL and ModelExpress
How prime-rl transfers a 1.6 TB GLM-5.2 policy into DPEP=32 vLLM inference, overlaps online quantization, and reduces end-to-end synchronization from 86.1 seconds to 3.9 seconds.
GLM-5.2 RL weight transfer in 4 seconds using NIXL and ModelExpress In RL at 1T Scale, we detailed how prime-rl trains trillion-parameter models like GLM-5 with sub-5-minute step times on just 28 H200 nodes. That speedup exposed a new bottleneck: syncing updated weights from trainer to inference after every step, which stayed stuck at 60-90 seconds no matter how fast everything else got. Here, we rebuild that hand-off on NIXL and ModelExpress, cutting weight transfer for GLM-5.2 down to single-digit seconds and removing the static-process-group constraints that stood in the way of…
saved by
related reading
- Journey to 2-second Inter-node RL Weight Transferle.qun.ch
- Updating 1T parameters in seconds — P2P weight transfer in Large Scale Distributed RL - LMSYS Orglmsys.org
- RL at 1T Scale: prime-rl Performance Deep Diveprimeintellect.ai
- Keep the Tokens Flowing: Lessons from 16 Open-Source RL Librarieshuggingface.co
- Transformer Inference Arithmetic | kipply's blogkipp.ly
- Reiner Pope – The math behind how LLMs are trained and serveddwarkesh.com
- RL Post-Training on Macs | Pluralis Researchpluralis.ai
- Transformer Math 101 | EleutherAI Blogblog.eleuther.ai
- The 4-bitter Lesson | humans&humansand.ai
- LLM Inference Performance Engineering: Best Practices | Databricks Blogdatabricks.com
- Async RL in Pure JAXdivyamakkar0.github.io
- The upcoming GPT-3 moment for RL | Mechanize, Inc.mechanize.work