PyTorch Ecosystem Adds 10 Projects Spanning RL, Inference, and Data Tooling

The PyTorch Ecosystem Working Group added 10 projects to its Ecosystem Landscape on August 26, 2026. The new entries cover reinforcement learning, inference serving, data tooling, and model visualization, with several aimed at production-scale workloads.

RL and Post-Training Additions

AReaL is an open source, modular RL infrastructure that connects foundation model training with LLM and VLM agent applications. It uses a fully asynchronous RL training paradigm and decomposes RL into independent, composable services. The project integrates with vLLM, SGLang, custom PyTorch-native 5D parallel engines, and other ecosystem pieces, focusing on connecting existing components into an end-to-end loop rather than introducing new ones.

RLinf is an open source reinforcement learning framework for embodied and agentic AI. It provides reusable building blocks for rollout, reward computation, environment execution, and learning, while hiding the complexity of heterogeneous hardware. RLinf aims to make robots and GPUs first-class citizens in the same workflow, supporting post-training beyond isolated model loops.

Miles, built and maintained by RadixArk, is an open source post-training framework for large-scale models. It targets frontier scale open MoEs, multi-node clusters, and long-running jobs. Miles has deep SGLang integration, support for new model architectures and hardware platforms, and production capabilities including LoRA, TITO, and low-precision training. The framework is customizable, allowing teams to modify the training loop and swap components.

TorchJD is a library for training neural networks with multiple losses. It supports scalarization, which combines losses into a single scalar loss, and Jacobian Descent, which computes the Jacobian of the vector of losses and aggregates it into a common update direction. With proper aggregation, Jacobian descent can decrease all losses simultaneously, though scalarization may suffice for very aligned gradients.

Inference and Serving

SMG, or Shepherd Model Gateway, is an engine-agnostic model-routing gateway for large-scale LLM deployments. Written in Rust, it sits in front of self-hosted inference engines like vLLM, TensorRT-LLM, TokenSpeed, SGLang, and MLX, unifying them behind a single OpenAI-compatible endpoint. Its cache-aware routing tracks each worker's KV-cache state in radix trees to maximize prefix reuse and GPU utilization. A native streaming gRPC pipeline supports prefill and decode disaggregation and data-parallel-aware routing. SMG adds priority admission scheduling, multi-tenancy, API-key auth with OIDC, WebAssembly plugins, and self-hosted chat history, along with 90+ Prometheus metrics and OpenTelemetry tracing.

TokenSpeed is an open source LLM inference engine that separates the control plane from the execution plane. The control plane is implemented in C++ as a finite-state machine, using the type system to enforce safe resource management at compile time. The execution plane is in Python, enabling fast iteration. TokenSpeed treats kernels as a modular subsystem with a portable public API, centralized registry, and plugin mechanism for heterogeneous accelerators.

Data and Visualization Tools

FiftyOne is a multimodal data platform for physical AI, indexing data like images, video, LiDAR, and radar for search, curation, annotation, and model evaluation. It integrates natively with PyTorch, allowing teams to load pre-trained models from PyTorch Hub and use FiftyOne datasets in training pipelines without separate dataset definitions.

VisualTorch is a PyTorch model visualization library that traces an actual forward pass to capture branching architectures and custom forward logic. It renders in multiple styles, including animated GIF reveals, and has been used in published research in Nature, IEEE, Elsevier, and MDPI journals.

TorchSurv is a lightweight, PyTorch-native toolkit for survival analysis. It provides fully differentiable survival losses, including Cox proportional hazards and Weibull AFT, along with evaluation tools like the concordance index, time-dependent AUC, and Brier score. TorchSurv has been used in studies spanning oncology, medical imaging, and multimodal AI, and is included in the FDA's Regulatory Science Tools Catalog.

Perforated is a data-efficiency library that adds neuron-specific reinforcement learning signals during training. It applies a lightweight modification to backpropagation, helping models extract more value from available training data. The project reports that teams reduced error by 70% with existing data, and it has been most extensively validated on edge AI and computer vision workloads.