AI · Official
NVIDIA Introduces Reshard Technique to Cut Weight Transfer Time in Reinforcement Learning
NVIDIA Technical blog ·
At a glance
In reinforcement learning post-training, the latest model weights must be continuously moved from training GPUs to inference GPUs. This transfer is called refit, and as models have grown, the amount of data that must be moved at once has increased sharply. NVIDIA introduced Reshard, a new communication method that moves only the necessary data.
Key points
- The existing method combined all shards once and then broadcast them whole, causing many unnecessary transfers.
- Reshard compares the layouts of the sending and receiving sides and moves only the parts that are needed.
- On NVIDIA DGX H100 and GB200 systems, it was measured to be 2.7 to 5.7 times faster than the existing method.
- A hybrid approach is also used: Reshard is applied to a subset of weights that make up most of the total data, while the existing method is kept for the rest.
Why it matters
This technology reduces the time and cost of fine-tuning large AI models with reinforcement learning, so it could affect the speed of AI development.
Original source
Accelerating Reinforcement Learning Weight Refit with the Reshard Collective
NVIDIA Technical blog · Nov 7, 03:03 KST
Read on NVIDIA Technical blog ↗AI summarized the original article. Check the original for accuracy. Copyright belongs to NVIDIA Technical blog.
More from this field
- AWS Marketplace Lets AI Agents Buy Software on Their Own
AWS has updated its AWS Marketplace so AI agents can buy software, APIs, data, and content on their own, not just find them. Buyers pay in dollars, and charges are added to their AWS bill.
디지털투데이 · Nov 21, 07:37 KST - Amazon SageMaker HyperPod Global Inference Gateway Routes AI Requests Across Regions
Amazon SageMaker HyperPod now has a Global Inference Gateway that sends each request to the best cluster across regions through one endpoint. Before this, clients had to choose a cluster themselves, and a busy cluster could leave requests waiting while idle GPUs sat elsewhere.
Amazon AWS Machine Learning blog · Nov 21, 07:35 KST - Google Drops Search Button and Adds AI Sidebar to Its Search Homepage
Google has officially rolled out a redesigned desktop search homepage after a successful test. The Google Search button next to the search bar is gone, and a row of AI buttons now sits under the bar. Logged-in users also get a sidebar showing their recent AI Mode conversations.
Ars Technica · Nov 21, 06:58 KST - Meta Rethinks 2030 Net-Zero Goal as AI Data Center Emissions Hit Record High
Meta's 2026 sustainability report said the company is re-evaluating its 2030 net-zero goal. Its emissions rose about 40 percent from 2024 to 2025, the largest jump among the big US tech companies.
Ars Technica · Nov 21, 06:47 KST