Monday, November 16, 2026 Edition No. 240
Search
Daily News Fully automated · 3 min read

NVIDIA AI Updates: November 8, 2026

Fully automated. Found, summarized and published by an automated AI pipeline. How posts are labeled

1. NeMo RL 0.8 Adds a Streaming Async RL Runtime, TensorRT-LLM and Dynamo Rollouts, and NVFP4 End-to-End RL on Blackwell

NVIDIA released NeMo RL v0.8.0 on November 6, headlined by the Single Controller (beta), a new async RL runtime in which one CPU-only Ray actor runs a rollout loop and a training loop at the same time over a shared TransferQueue data plane: finished prompt groups go into a replay buffer, training can start as soon as a configurable number of groups is ready instead of waiting for the full batch, and five samplers (in_order, weight_fifo, ready_first, windowed, custom) set how much staleness a run tolerates. The Single Controller supports GRPO and PPO on the Megatron training backend for now, retries or drops failed rollouts, rebuilds the refit group when a generation shard dies, and checkpoints rollouts that have not been trained on yet. TensorRT-LLM and Dynamo join vLLM, SGLang and Megatron Inference as rollout engines, and the NCCL reshard refit ships with two more transports: NIXL over UCX/RDMA (1.54x faster than NCCL refit on DeepSeek-V3) and sparse delta refit through ZeroMQ or S3 for slow links; in a full DeepSeek-V3 async recipe on H100, reshard cut refit time from 15.3 s to 3.1 s and step time from 33.2 s to 21.0 s. On Blackwell, routed-expert MLPs can now train in NVFP4 and roll out in NVFP4, which over 900 steps on Qwen3-30B-A3B-Base tracked BF16 on reward while giving 1.40x rollout throughput, 1.15x end-to-end throughput and 2.62x the KV-cache capacity, at the cost of 0.84x training throughput. The release also moves the Nemotron 3 Ultra (550B-A55B) post-training recipe to the main branch, adds Megatron-backend RL for Nemotron Omni models, and breaks existing configs: the DTensor v1 backend is removed and DTensor v2 is renamed Automodel, so dtensor_cfg becomes automodel_cfg and DTensor v1 checkpoints cannot be resumed. Source