NVIDIA open-sources the Reshard collective, making RL post-training weight refit 2.7x to 5.7x faster than gather-and-broadcast
The step in RL post-training where training GPUs push new weights to inference GPUs now has an open implementation that moves only the minimum data.
// Key points
- With training and inference on separate GPUs, updated weights must be refit to the inference GPUs periodically; NVIDIA estimates about 10 TB per refit for today's largest open-weight models
- Reshard treats refit as a DTensor-to-DTensor transfer and moves only the theoretical minimum data between arbitrary source and destination layouts
- The open reference implementation, NCCL M2N Reshard, is in the NVIDIA/nccl-extensions repository
- In NeMo RL tests on DGX H100 and DGX GB200 it ran 2.7x to 5.7x faster across models including Qwen 3 235B and DeepSeek v3 671B
Builder's takeMost product teams will not run RL at this scale, but it shows post-training cost for open models is still being engineered down, so training your own specialist model keeps getting cheaper. Get your eval set and reward signals ready now, and start once this lands in NeMo RL's default path.
// Background · from #Open models
Full timeline →- Nov 8 Kimi K3.1 Pro (Highspeed) arrives on Amazon Bedrock: open-weight, natively multimodal, a 1-million-token context, and explicit prompt caching
- Nov 8 Gradium releases Phonon for on-device speech: from about 65 MB, 28 ms p95 time to first audio on CPU streaming, and an Apache 2.0 English model
- Nov 7 Foundry Local runs Qwen3.8-27B at 4-bit in about 18 GB, on 24 GB consumer GPUs with up to 64K context, as a GitHub Copilot model provider