Packages
pyautogen is a proxy package that redirects to autogen-agentchat, a framework for building multi-agent AI applications where agents can operate autonomously or collaborate with humans.
However, note that the proxy package itself has not been released in 395 days—you are relying on the underlying autogen-agentchat package for active maintenance.
FlagEmbedding provides embedding and reranking models for semantic search and retrieval-augmented generation (RAG), supporting dense retrieval, multilingual queries, and document ranking.
ClearML is an ML/DL development and production suite that automates experiment tracking, captures environment and hyperparameter data, logs outputs and metrics, and provides orchestration and data management for machine learning workflows.
MMDetection is an object detection and instance segmentation framework built on PyTorch that provides modular, pre-trained detection models and tools for training and inference on custom datasets.
However, maintenance is dormant (last release 2024-01-05), so expect no new features or bug fixes—use it only if the current stable version meets your needs.
Provides NVIDIA's CUDA C++ Core Compute Libraries (CCCL) for GPU-accelerated computing on CUDA 12 platforms.
smolagents lets you build and run AI agents that write Python code to call tools and orchestrate other agents, with support for multiple LLM providers and sandboxed execution.
nemo-relay wraps a shared Rust runtime to give Python applications an agent execution framework with scope management, middleware, lifecycle events, and observability for tool and LLM calls.
Integrates Azure OpenAI language models into LlamaIndex applications, enabling text completion, streaming, and chat interactions through Azure's managed OpenAI service.
Python SDK for querying Claude Code and building interactive agents with custom tools and hooks, now deprecated in favor of claude-agent-sdk.
Chalk is a Python SDK for building and deploying machine learning feature pipelines that compute features from multiple data sources and serve them for both online inference and offline training.
However, the license is unclear, so verify terms before production use.
Liger Kernel provides optimized Triton kernels for LLM training, implementing fused operations like RMSNorm, RoPE, SwiGLU, and CrossEntropy to reduce memory usage and increase training throughput.
Unstructured Ingest is a local ETL pipeline that prepares unstructured data for retrieval-augmented generation (RAG) and other AI workflows by ingesting and transforming raw documents into structured, ready-to-use formats.
Provides utility functions for PyTorch and einops to accelerate tensor manipulation in ML workflows.
Decord2 decodes video and audio files with hardware acceleration, providing efficient random frame access and batch loading for deep learning workflows.
Computes cross-entropy loss for large-vocabulary language models with minimal memory footprint by materializing logits selectively rather than globally, using custom kernels and torch.compile fallbacks.
However, verify the license terms before production use (license treatment is unclear), and be aware that maintenance is dormant—no active development is expected.
Tinker Cookbook provides recipes and abstractions for fine-tuning language models via the Tinker API, supporting supervised learning, reinforcement learning, and specialized tasks like math reasoning, code generation, and multimodal training.
Labelbox is a Python SDK for interacting with the Labelbox data labeling platform, providing client-side access to projects, labels, annotations, and related data management through a GraphQL-backed API.
Install it if you need to programmatically interact with Labelbox projects, export annotations, or integrate labeling workflows into Python applications.
sklearn-crfsuite wraps CRFsuite (Conditional Random Fields) as a scikit-learn compatible estimator, letting you train sequence labeling models using scikit-learn's standard tools like cross-validation and hyperparameter optimization.
POT provides solvers for optimal transport problems, including Wasserstein distances, Gromov-Wasserstein distances, and related algorithms for signal processing, image processing, and machine learning applications.
DeepFace performs face recognition, verification, and facial attribute analysis (age, gender, emotion, race) using deep learning models, with support for real-time video streams and large-scale database searches.
Install it if your use case justifies the disk and memory overhead; avoid it if you need a lightweight, minimal-dependency solution or are uncertain whether you need…
iden manages machine learning datasets organized into shards with lazy loading, supporting multiple storage formats (JSON, YAML, Pickle, PyTorch, safetensors) and optional in-memory caching.
Evaluates and benchmarks large language models' ability to invoke functions correctly across diverse scenarios, from simple calls to multi-turn agentic interactions with web search and memory management.
Lepton AI is a Python framework for building and deploying AI services from research and modeling code, with built-in support for launching HuggingFace models, autobatching, background jobs, and a client for calling services as native Python functions.
Computes ML-relevant statistical summaries of datasets, integrating with Amazon SageMaker to surface data characteristics for model development and debugging.
No—not recommended for new projects.
Provides a Python library for Amazon SageMaker Data Wrangler, enabling data preparation and transformation workflows within the SageMaker ecosystem.
Reads and writes Kaldi binary archives, alignment files, and neural network training examples in Python, providing sequential and random-access interfaces to Kaldi's data formats.
Gin provides a lightweight configuration framework that lets you decorate Python functions and classes to accept parameter values from config files or command-line arguments, eliminating the need for manual configuration object management.
Install it if parameter management is a pain point in your workflow.
CLU provides utilities and abstractions for writing machine learning training loops in JAX, including metrics collection, checkpointing, and common patterns to keep loop code concise while retaining research flexibility.
Client library for NVIDIA Riva, a GPU-accelerated Speech AI SDK providing APIs for automatic speech recognition, text-to-speech synthesis, and natural language processing tasks.
Install it if you need to build Speech AI applications with NVIDIA Riva; the main prerequisite is having a Riva server running and, for audio I/O features, PyAudio…
Automates machine learning model training and prediction on tabular data with minimal code, handling feature engineering, algorithm selection, and hyperparameter tuning internally.
DVCLive logs machine learning metrics, parameters, and plots to simple file formats that integrate with DVC for experiment tracking and comparison.
Connects LangChain applications to OpenRouter, a unified API gateway for hundreds of AI models across multiple providers, enabling model selection and switching without changing application code.
Install it if you're building with LangChain and want the flexibility to choose from hundreds of models without rewriting your code.
EconML estimates heterogeneous treatment effects from observational data using machine learning, measuring how causal effects of interventions vary across different sample characteristics.
AutoGluon Core provides the foundational infrastructure for automated machine learning, enabling training of high-accuracy models on tabular, time series, image, and text data with minimal code.
Trains and deploys decision forest models (Random Forests, Gradient Boosted Trees) within TensorFlow for classification, regression, and ranking tasks.
Not recommended for Windows users without WSL, or for projects not already committed to TensorFlow.
Provides a unified interface to access local and remote files (HTTP, HTTPS, S3, Google Cloud Storage, HuggingFace Hub) with automatic caching and archive extraction.
Verifiers provides environments and evaluation harnesses for training and assessing large language models using reinforcement learning, with integration into the Prime ecosystem.
NIXL provides Python bindings to accelerate point-to-point communications in AI inference frameworks, abstracting CPU and GPU memory alongside file, block, and object storage through a modular plugin architecture.
Flytekit is the Python SDK for authoring, testing, deploying, and managing Flyte tasks, workflows, and launch plans—a framework for building data and ML pipelines that run on the Flyte backend.
However, note that Flyte 2 is now generally available—new projects should evaluate whether Flyte 2's Python SDK is a better fit, and existing Flyte 1.x users should…
FairScale extends PyTorch with distributed training primitives and optimizations for scaling model training across multiple machines or GPUs.