Packages
pyRiemann provides scikit-learn-compatible machine learning algorithms for classifying multivariate data using Riemannian geometry of symmetric positive definite matrices, with support for both NumPy and PyTorch backends.
Implements vectorized Kalman filters that run many independent filters in parallel over multiple time series using NumPy's array operations, enabling fast filtering and smoothing of batched observations.
However, maintenance is dormant—no updates in 972 days—so evaluate whether you need ongoing support or can work with a stable but no-longer-maintained codebase.
Integrates Google Vertex AI's LLM capabilities with LlamaIndex, enabling use of Vertex models within LlamaIndex applications.
However, verify whether Google GenAI integration better suits your use case, as the package documentation suggests Gemini has moved to that path.
Python SDK for building, deploying, and managing AI agents on the xpander.ai Backend-as-a-Service platform, with support for agent lifecycle management, task execution, tool integration, knowledge bases, and event-driven workflows.
Enables PyTorch to use DirectML for GPU acceleration on DirectX 12-compatible hardware from AMD, Intel, NVIDIA, and Qualcomm.
However, operator coverage is still developing—verify that your specific PyTorch operations are supported via the operator roadmap before committing to production use.
Jraph provides data structures and utilities for building and working with graph neural networks in JAX, including a GraphsTuple representation for graphs and a model zoo with reference implementations of message-passing algorithms.
However, verify compatibility with your JAX version before adopting it for new projects, since there will be no maintenance updates.
A Python SDK for writing serverless endpoints that run on Cozy's worker pool, handling model discovery, scheduling, download, device placement, cancellation, and streaming.
TransformerLens loads and inspects the internal activations of transformer language models, letting you cache, edit, and analyze what happens inside the model as it processes text.
Helper library for integrating Python code with the Valohai machine learning platform, handling parameters, inputs, outputs, metrics logging, and distributed task coordination.
However, if you are not using Valohai, this package has no value.
Provides reference implementations for running OpenAI's gpt-oss-120b and gpt-oss-20b open-weight models locally, with support for PyTorch, Triton, and Metal inference backends.
Provides a safe wrapper around PyTorch's scan operator that checks backward compatibility across devices and PyTorch versions, with a fallback eager implementation for PyTorch 2.7.
However, it is early-stage (Alpha) with minimal real-world usage; test thoroughly in your environment before relying on it in production.
MemPalace stores conversation history and project files as verbatim text and retrieves them via semantic search, with a structured index (wings, rooms, drawers) and pluggable backends including ChromaDB, Qdrant, Milvus, and pgvector.
Install it if you work with Claude Code, need local RAG, or want to preserve and search your conversation history.
Extracts keyphrases from text documents using part-of-speech patterns and produces document-keyphrase matrices compatible with scikit-learn's vectorizer interface.
However, do not expect active development or rapid bug fixes; treat it as a research tool rather than a production library with ongoing support.
Integrates IBM watsonx.ai foundation models with LlamaIndex's LLM API, enabling you to use IBM's models for text completion and chat inference through a unified interface.
Estimates token usage for OpenAI Chat Completions API messages and builds message lists that fit within model token limits, accounting for text, images, tools, and few-shot examples.
Skope-rules learns interpretable logical rules from data to classify instances with high precision, balancing the simplicity of decision trees with the power of random forests.
However, verify compatibility with your current numpy, scikit-learn, scipy, and pandas versions before relying on it in production.
Automated machine learning for image classification and object detection, handling model selection, hyperparameter tuning, and training with minimal code.
However, the latest release is from 2023-01-11, so verify ongoing support and compatibility with your target environment before committing to production use.
Aegra API is a self-hosted Agent Protocol server that runs AI agent graphs on your own infrastructure, compatible with LangGraph SDK and providing streaming, human-in-the-loop workflows, and scheduled cron job execution.
Trainer is a PyTorch model training framework that handles the training loop, optimization, and experiment logging for machine learning models with support for mixed precision, distributed training, and callback customization.
Not recommended if you require active maintenance or support for the latest PyTorch features.
Integrates IBM watsonx.ai embedding models with LlamaIndex, allowing you to embed text strings using IBM's foundation models through the LlamaIndex API.
Amalgam is a Python wrapper that lets you call Amalgam, a tree-structured language designed for safe, sandboxed code generation and information-theoretic inference, directly from Python code.
However, verify that your OS and CPU architecture are in the supported set (Windows/Linux x86_64, Linux/macOS arm64), and be aware that the GNU Affero v3 license…
GraphRAG extracts structured knowledge graphs from unstructured text using LLMs, enabling retrieval-augmented generation (RAG) systems to reason more effectively over private data.
Not recommended if you need a lightweight, zero-configuration RAG solution.
vllm-tpu is a high-throughput LLM inference and serving engine optimized for Google TPUs, supporting 200+ model architectures with efficient memory management, continuous batching, and structured output generation.
Provides pretrained neural network models for image super-resolution (upscaling low-resolution images to higher resolution) using PyTorch, including EDSR, CARN, RDN, RCAN, and NinaSR architectures.
However, verify that pretrained weights remain accessible and that it works with your PyTorch version—no updates have landed since August 2022, so compatibility with…
Provides axial positional embeddings for transformer networks, enabling position encoding across multi-dimensional data like images, videos, or sequences using decomposed attention axes.
Automates machine learning for text data, training and deploying high-accuracy models with minimal code through AutoGluon's text prediction interface.
A local execution agent library for building and running AI agents with LLM integration, tool calling, guardrails, and skill management as part of the GL AIP ecosystem.
However, it depends on an internal registry that may not be accessible without special setup, and its license is undeclared.
Provides high-level APIs for training, evaluating, and exporting machine learning models, encapsulating the full model lifecycle.
Provides differentiable GPU-capable solvers for controlled differential equations (CDEs), enabling Neural Controlled Differential Equation models for irregular time series with support for backpropagation and memory-efficient adjoint methods.
However, proceed with caution: the project is aging (last release October 2021, no recent commits), so compatibility with the latest PyTorch versions and long-term…
Merges multiple ONNX neural network models into a single combined model by connecting specified output and input operators, with optional operator name prefixing to avoid conflicts.
Generates minimal ONNX operation graphs by specifying operation type, inputs, outputs, and attributes without external runtime dependencies.
However, dormant maintenance since April 2024 means you should verify compatibility with your target ONNX opset version before relying on it in production.
dtreeviz renders decision trees from scikit-learn, XGBoost, LightGBM, Spark MLlib, and TensorFlow as interactive visualizations to help understand and interpret how these models make predictions.
Adds or modifies output nodes in ONNX model graphs, allowing you to expose intermediate layer outputs as model endpoints via CLI or Python API.
However, maintenance is aging (last release 314 days ago), so verify compatibility with your ONNX runtime version before relying on it in production pipelines.
A self-hosted LLM console and OpenAI-compatible API server that runs model inference, chat, and a web UI in a single process, with support for GPU worker fleets and cross-machine model sharding.
Removes specified nodes from ONNX model graphs via CLI or Python API, outputting a modified model file or in-memory graph.
LMCache is a KV cache management layer that stores, reuses, and monitors transformer key-value caches across LLM inference requests, reducing time-to-first-token and improving throughput for long-context workloads.
Renames input, output, and operator nodes in ONNX model files using string matching (exact, partial, prefix, or suffix) via CLI or Python API.
However, maintenance is dormant (last update 808 days ago), so verify compatibility with your ONNX version before relying on it in production.
verl is a reinforcement learning training library for large language models that implements flexible RL algorithms like PPO and GRPO with modular APIs for integration into existing LLM frameworks.
Memsearch builds a searchable semantic memory index from markdown files, enabling AI coding agents to recall and reuse conversation context across sessions and platforms through vector search and hybrid retrieval.
Converts ONNX model files to JSON format, either writing to a file or returning a dictionary representation of the model structure.
However, do not rely on it for long-term maintenance or support for cutting-edge ONNX features—it is abandoned.