NCCL
NVIDIA says Reshard can cut reinforcement learning weight refit time
NVIDIA has introduced the Reshard collective, a communication method designed to accelerate the transfer of updated model weights between training and inference GPUs…
NVIDIA Open-Sources Attribution Tool to Trace Root Causes of AI Workload Failures
NVIDIA has made NVIDIA Attribution open source, a system that correlates logs, metrics, telemetry and scheduler state to explain why a distributed AI workload failed.…
PyTorch 2.15 makes rewritten NCCL backend the default and adds FP8 on Apple Silicon
PyTorch 2.15 is out, and the rewritten NCCL backend now runs every nccl process group by default. The release also adds one-sided transports to torch.distributed, finer…
NVIDIA Unifies GPU-Initiated Networking Under DOCA GPUNetIO Across Its Stack
NVIDIA says its DOCA GPUNetIO framework has become a shared GDA-KI foundation that lets CUDA kernels drive network operations directly, with libraries such as NCCL…
Modal Clusters reach general availability with one-line multi-node GPU access
Modal has made Modal Clusters generally available, letting developers run multi-node GPU workloads through a single decorator, @modal.clustered. According to the…