Meta AI Updates: July 1, 2026

1. PyTorch Details Miles, a Native Stack for Large-Scale LLM RL Post-Training

PyTorch. The PyTorch blog published Miles, an open-source framework from RadixArk that treats large-scale LLM reinforcement learning post-training as a distributed systems problem rather than a single training loop. Built natively on PyTorch, it composes SGLang for high-throughput rollout generation, Megatron-LM for distributed training, and Ray for cluster orchestration, while handling weight synchronization, mixture-of-experts routing alignment, and low-precision recipes spanning BF16, FP8, MXFP8, and INT4-QAT. It matters because it abstracts away placement, fault tolerance, and precision decisions so teams can focus on RL algorithms and product logic. Source

2. PyTorch Introduces Cross-Repository CI Relay for Out-of-Tree Backends

PyTorch. PyTorch announced the Cross-Repository CI Relay (CRCR), an automated system that connects the upstream pytorch/pytorch repository with downstream out-of-tree backend repositories such as Intel XPU, AMD ROCm, and Apple MPS. CRCR automatically triggers CI in those downstream repos on each PR or commit, collects results, and surfaces them on a unified dashboard, using a tiered L1 to L4 participation model and five layers of security validation. The goal is to close visibility and coordination gaps so maintainers can catch downstream breakage before merging, with downstream repos needing only a single workflow file and an allowlist entry to opt in. Source