Packages
Provides a C-based API for annotating events, code ranges, and resources in applications to enable capture and visualization through NVIDIA's Visual Profiler.
PyTorch Geometric is a library for building and training Graph Neural Networks (GNNs) on structured data, providing pre-built GNN layers, datasets, data loaders, and utilities for geometric deep learning.
It is the standard library for GNN work in PyTorch—install it if you need to build or train graph neural networks.
Provides a portable intermediate representation (TileIR) for CUDA kernels that abstracts away language-specific details, enabling kernel compilation and optimization across different contexts.
Integrates curated transformer models (ALBERT, BERT, CamemBERT, RoBERTa, XLM-RoBERTa) into spaCy pipelines via the curated-transformers library.
Exposes Unity Catalog functions as tools for OpenAI models, enabling agents to call UC functions through OpenAI's function-calling interface.
Provides pre-built schema inference artifacts and task sample inputs/outputs for Hugging Face models in SageMaker workflows.
However, dormant maintenance status means you should verify that included schemas match your target versions before relying on it in production.
Provides NVIDIA's collective communication library (NCCL) runtime for GPU-accelerated all-reduce, all-gather, reduce, broadcast, and reduce-scatter operations optimized for CUDA 11.
WhisperX performs fast automatic speech recognition with word-level timestamps and speaker diarization, using batched inference and forced phoneme alignment to improve accuracy over standard Whisper.
Neptune-scale is a Python client library for logging and monitoring experiment metadata—metrics, configurations, files, and histograms—during model training, with integration to the Neptune web platform for visualization and analysis.
However, the aging maintenance status (261 days since last release) suggests you should verify that updates and support align with your project timeline before…
Integrates Databricks AI features, particularly vector search, into OpenAI applications through tool definitions and retrieval workflows.
The package is narrowly scoped to a specific integration pattern, so install only if you have both Databricks infrastructure and OpenAI API access.
Provides a Python CLI and SDK for scaffolding, developing, and deploying AI agents on Amazon Bedrock AgentCore, though the package is now superseded by the AgentCore CLI.
No, not for new projects.
PyThaiNLP provides Thai-language natural language processing tools including tokenization, part-of-speech tagging, transliteration, spelling correction, and linguistic utilities, designed as a Thai counterpart to NLTK.
Install it if you need to process Thai text; the base package is lightweight and the optional extras allow you to add machine translation or WordNet support as needed.
nvdisasm disassembles NVIDIA CUDA cubin files into human-readable CUDA assembly code, extracting and presenting the compiled GPU kernel information.
Humming is a JIT-compiled GEMM kernel library for quantized matrix multiplication on NVIDIA GPUs, supporting mixed quantization formats (FP16, BF16, FP8, FP4, INT8, INT4) for both dense and MoE inference workloads.
However, verify the license terms before use, and confirm that your CUDA and PyTorch versions are compatible—the fact sheet does not specify exact version constraints…
Integrates Fireworks.ai language models with LangChain, enabling you to use Fireworks' inference API within LangChain applications and workflows.
MLX LM loads, generates text with, fine-tunes, and quantizes large language models on Apple silicon using the MLX framework and Hugging Face Hub integration.
Stable Baselines3 provides PyTorch implementations of reinforcement learning algorithms designed to work with Gymnasium environments, following a scikit-learn-like API for training and inference.
Install it if you're training RL agents or need a trusted baseline for comparison.
TF-Keras is a pure-TensorFlow implementation of Keras that provides a high-level deep learning API for building and training neural networks.
The nightly build is suitable for developers and researchers who can tolerate frequent changes and want early access to new features, but not for production systems…
Stagehand is a Python SDK for building browser agents with self-healing actions, agent-optimized APIs, and support for complex DOM structures like iframes and Shadow DOMs.
x-transformers provides modular transformer building blocks—encoder, decoder, and encoder-decoder architectures—with experimental features like Flash Attention, memory tokens, and persistent memory for research and production use.
Read-only API for querying experiment metadata, runs, metrics, and attributes with filtering and export to pandas DataFrames.
However, no bug fixes or compatibility updates are forthcoming—install only if your use case is stable and unlikely to require updates, or if you can manage potential…
Provides predefined Kubeflow Pipelines components for Google Cloud Vertex AI that handle model training, deployment, and data processing tasks without writing custom pipeline logic.
Install only if you have a GCP project with Vertex API enabled and authenticated credentials; it is not useful without Google Cloud infrastructure.
Supervision provides utilities for loading, annotating, and processing computer vision datasets and model outputs—detection, segmentation, and classification results from any model framework.
Install it if you work with detection, segmentation, or classification models and want to avoid reinventing dataset utilities.
Provides pre-compiled Python gRPC stubs for protocol buffers used by SMG (Shepherd Model Gateway) to communicate with LLM inference services including SGLang, vLLM, and TensorRT-LLM.
Install only if you actually need these specific proto stubs; it is a narrow-purpose dependency.
Unsloth Zoo provides utilities for fine-tuning large language models with reduced memory usage and faster training speeds.
Augments images and related data (heatmaps, segmentation maps, keypoints, bounding boxes, polygons) by applying transformations like rotations, noise, cropping, and color shifts to expand training datasets for machine learning.
However, the package is dormant—last release was 2020-02-05, and two security vulnerabilities are recorded.
Chonkie splits text into semantically meaningful chunks for RAG pipelines, offering multiple chunking strategies (recursive, semantic, token-based, code-aware) plus refinement and embedding integration.
Enables Python bots to send and receive messages through Microsoft Bot Framework channels by providing HTTP client authentication and conversation management.
Provides CMA-ES (Covariance Matrix Adaptation Evolution Strategy) optimization in an ask-and-tell interface, supporting continuous, integer, and categorical variable optimization with variants for mixed-variable and multi-objective problems.
Provides serialized data schemas and message types for communication between bots and users in the Microsoft Bot Framework ecosystem.
However, the repository is archived and abandoned, meaning no future updates or bug fixes will be released.
Provides model serving and deployment functionality for machine learning models on Amazon SageMaker, integrating with SageMaker's core training and inference infrastructure.
However, verify that its API and deployment model match your specific serving requirements before committing, as the fact sheet does not detail its exact interface or…
Truss is a CLI tool for packaging ML models with their serving logic and deploying them to production, handling containerization, dependency management, and GPU configuration automatically.
MLX is an array framework for machine learning on Apple silicon that provides NumPy-like Python APIs, automatic differentiation, lazy computation, and unified memory across CPU and GPU devices.
Connects Mistral AI language models to LangChain, enabling you to use Mistral's models within LangChain's agent and chain frameworks.
Trains and deploys machine learning models on Amazon SageMaker using popular frameworks like Apache MXNet, TensorFlow, and Amazon's built-in algorithms, or custom Docker containers.
However, verify the license terms (Apache 2.0 is mentioned in the description but not confirmed in metadata) and ensure your AWS account has the necessary IAM…
TorchRL is a PyTorch-native toolkit for building reinforcement learning systems with composable components for environments, policies, collectors, replay buffers, and loss functions.
However, verify the license status before committing to proprietary use, and ensure your target platform has compatible wheels (Python 3.10–3.14, macOS/Linux/Windows).
Integrates Cartesia's voice AI services (speech-to-text and text-to-speech) into LiveKit Agents for real-time voice applications.
Provides high-level orchestration for Amazon SageMaker workflows, including pipeline definitions, step implementations, and model building utilities that coordinate training, serving, and core SageMaker components.
Lhotse prepares multimodal (speech, audio, video, image, text) data for machine learning model training with flexible pipelines, on-the-fly augmentation, and efficient data loading.
Install it if you are building speech, audio, or multimodal training pipelines; skip it if you only need simple audio I/O without data augmentation or complex dataset…
Provides the NVIDIA CUDA nvcc compiler for building CUDA applications, packaged as a Python distribution for easy installation across Linux and Windows platforms.
However, verify that the proprietary license terms fit your use case, and confirm that CUDA 12.9.86 matches your GPU and toolkit requirements before committing to…