Packages
Automatic Speech Recognition using ONNX models with minimal dependencies, supporting multiple modern ASR architectures and running on CPUs, GPUs, and edge devices.
Install it if you need ASR inference in Python without framework overhead.
Defines step types for building computation units in Azure Machine Learning pipelines, enabling orchestration of training, data processing, and automated ML workflows.
Integrates Soniox speech-to-text and text-to-speech APIs into LiveKit Agents via WebSocket streaming, enabling real-time transcription and speech synthesis in agent applications.
Install it if you are building a LiveKit agent and want to use Soniox for speech services; it is the intended integration point for that use case.
Reduces transformer model vocabulary to only tokens used during training, cutting embedding memory and compute overhead while preserving the full model after training completes.
cuEquivariance provides CUDA-accelerated operations for building geometric neural networks that respect symmetries, with segmented polynomial kernels and bindings for PyTorch and JAX.
However, note that the package is in Beta, so expect potential API changes.
Adds training telemetry collection to Lightning applications via callback hooks, capturing training loop events, checkpointing, and lifecycle metrics automatically or on-demand.
Exports and converts PyTorch neural network models to PNNX format, an open standard for model interoperability that generates optimized graph definitions, weights, and inference code.
Polygraphy is a toolkit for running inference across multiple deep learning backends (TensorRT, ONNX-Runtime, etc.), comparing results, converting models to different formats, and debugging model behavior.
Install it if you need to prototype, compare, or debug deep learning inference across frameworks.
BlackJAX provides a collection of MCMC and other sampling algorithms built on JAX that run on CPU and GPU, designed for users who have a log-probability function and need a performant sampler.
Provides Python bindings for NVIDIA TensorRT 11, enabling high-performance deep learning inference on CUDA 12 GPUs through optimized model compilation and execution.
Converts ONNX models to PyTorch modules with a simple API, supporting a limited but growing set of operations and popular model architectures.
However, verify that your specific ONNX model's operations are supported before committing to it—the converter does not support all ONNX operations, and maintenance…
GEM is a reinforcement learning environment suite that provides a standardized API for training agentic LLMs through interactive experience in diverse task domains including games, math, code, QA, and reasoning.
However, the aging maintenance status (last commit 2026-01-21, no recent releases) and beta stability warrant caution for production use.
Implements a gradient reversal layer for PyTorch neural networks, enabling adversarial training by reversing gradients during backpropagation.
However, verify that the package works with your specific PyTorch version first, given the dormant maintenance status and last release in 2021.
Speeds up Stable Diffusion image generation by merging redundant tokens in the transformer, reducing computation and memory without requiring model retraining.
Provides OAuth 2.0 token management for H2O Python clients, handling refresh token exchange and access token lifecycle with both sync and async interfaces.
Install only if you need OAuth token management for H2O services; it is not a general-purpose OAuth library.
Qwen-TTS generates speech from text using Qwen3-TTS models, supporting voice cloning, voice design, and instruction-based voice control across 10 languages with streaming and non-streaming output.
However, the package is very recent (first release 2026-01-22) with aging maintenance status, so expect potential API changes and monitor for updates.
VisualDL is a visualization tool for deep learning training that logs metrics, model structures, and tensor distributions to files, then displays them in a web-based dashboard for real-time monitoring and analysis.
Integrates Exa's web search API with LangChain, enabling AI applications to search the web and retrieve clean, ready-to-use content from pages.
Install it if you are building a LangChain agent or RAG system and need to incorporate live web search results.
Open Interpreter lets language models execute code (Python, JavaScript, Shell) locally on your machine through a natural-language chat interface, with user approval required before each command runs.
However, the package is dormant (no releases in 657 days), carries an unclear license, and depends on 32 libraries—audit the license and dependencies before use in…
llmcompressor optimizes large language models for efficient deployment by applying quantization, pruning, and other compression techniques compatible with vLLM inference.
Detects faces in images and extracts facial landmarks (eyes, nose, mouth) using a deep learning model, with optional face alignment for recognition pipelines.
Install it if you are building a face recognition pipeline or need facial landmarks for alignment; skip it if you only need bounding boxes and can tolerate heavier…
Ax is a platform for adaptive experimentation that uses machine-learning-guided optimization to iteratively explore parameter spaces and identify optimal configurations. It supports Bayesian optimization and bandit optimization strategies powered by BoTorch.
Install it if you need a general-purpose optimization platform; the main constraint is the Python 3.11+ requirement.
Builds, optimizes, and manages machine learning workflows in Azure by composing data preparation, model training, and deployment steps into executable pipelines that can be submitted and tracked.
SeqIO provides task-based data pipelines for sequence models, combining raw data sources, preprocessing, vocabularies, and evaluation metrics into reusable registered tasks that output tf.data.Dataset instances compatible with JAX, PyTorch, and other frameworks.
However, this is a nightly build (0.0.18.dev20250227); verify that the development version is appropriate for your use case.
Provides Python bindings for NVIDIA TensorRT 11.2.1.2 compiled for CUDA 13, enabling deep learning inference acceleration on compatible GPUs.
Implements ring attention with flash attention for distributed multi-GPU training, enabling efficient sequence parallelism across devices with variants for different compute patterns and sequence packing strategies.
However, it is specialized infrastructure code—not a general-purpose library—and requires a distributed training setup with multiple GPUs to be useful.
H2O is a distributed machine learning and statistical analysis platform that scales algorithms over large datasets, offering Python interfaces to build and deploy predictive models.
However, the high install friction and large package footprint require adequate disk and memory resources.
OpenDataLab Python SDK provides programmatic and CLI access to datasets hosted on the OpenDataLab platform, with login, search, listing, and download capabilities.
However, be aware that the last release was over a year ago, the maintainers flag ongoing compatibility concerns with the platform API, and you may need to stay on…
Provides a Model Context Protocol (MCP) server that enables AI assistants and LLMs to query and manage OpenSearch clusters through a standardized interface supporting both stdio and streaming transports.
loralib provides PyTorch modules that implement Low-Rank Adaptation (LoRA), enabling efficient fine-tuning of large language models by training only small rank-decomposition matrices while freezing original weights.
However, note that loralib has been dormant since August 2023 and may not track the latest library changes, so verify compatibility with your current dependencies…
Aligns tokenizations between spaCy and transformer models by mapping token boundaries between different tokenization schemes, enabling seamless integration of spaCy NLP pipelines with transformer-based models.
Workspace MCP is an MCP server that exposes Google Workspace services (Calendar, Gmail, Drive, Docs, Sheets, Slides, Forms, Tasks, Contacts, Chat, Custom Search, and Apps Script) as tools for AI assistants and LLM clients through a single standardized interface.
River provides online machine learning algorithms that learn from streaming data one sample at a time, including linear models, decision trees, anomaly detection, clustering, and time series forecasting.
Integrates enterprise agents with data sources and tools across any agentic framework, treating agents and workflows as composable function calls.
Not recommended for first-time NeMo Agent Toolkit users; the documentation suggests installing from the source repository first to learn the toolkit.
Translates human hand motion to robot hand motion through retargeting optimizers, supporting both video-based and pose-dataset-based workflows.
However, the aging maintenance status and reliance on compiled dependencies like nlopt mean you should verify platform compatibility and be prepared to troubleshoot…
Provides a development toolkit for building, debugging, evaluating, and deploying LLM application flows with integrated tracing and observability.
Audiomentations applies randomized audio transformations—noise injection, pitch shifting, time stretching, filtering, and distortion—to augment training data for machine learning models that work with audio.
Install it if audio data augmentation is part of your training pipeline.
scikit-fuzzy provides fuzzy logic algorithms and operations for scientific Python, implementing fuzzy sets, membership functions, and fuzzy inference systems.
However, verify the license terms first, confirm that its dependencies meet your requirements, and be aware that long-term compatibility with future Python versions…
MNN is a lightweight deep learning inference and training framework that runs neural network models efficiently on mobile devices, embedded systems, and servers with support for multiple hardware backends including CPU, GPU, and NPU.
Python client for querying and managing documents in H2OGPTe, an enterprise RAG platform that combines document ingestion, semantic search, and LLM-powered question-answering.