artyom-beilis/pytorch_dlprim

DLPrimitives/OpenCL out of tree backend for pytorch

What it solves

It provides an OpenCL backend for PyTorch, allowing users to run and train neural networks on a wider variety of hardware devices (including AMD, Nvidia, and Intel GPUs) that may not have native CUDA support.

How it works

The project implements an "out-of-tree" backend based on the dlprimitives library, which handles the actual operator implementations. By importing the pytorch_ocl package, users can target their OpenCL-compatible devices using the device string ocl (or privateuseone for older PyTorch versions) instead of cuda.

Who it’s for

Developers and researchers who want to utilize OpenCL-compatible GPUs from various vendors (AMD, Intel, Nvidia) for PyTorch workloads, particularly for vision-based networks.

Highlights

  • Broad Hardware Support: Tested on AMD (RX 6600XT, RX 560), Nvidia (GTX 960), and Intel (HD 530, UHD 770, Arc A380) devices.
  • Vision Network Validation: Validated against CPU references for a wide range of torchvision models, including ResNet, MobileNet, EfficientNet, and ViT.
  • PyTorch Compatibility: Supports PyTorch versions 1.13 and 2.4+.
  • Dedicated API: Includes specific functions for memory management and synchronization, such as torch.ocl.empty_cache() and torch.ocl.synchronize().

Related

  • Project
    MAX-API-Next/MAX-API

    MAX API is an open‑source gateway that unifies access to many LLM providers, adds governance (auth, rate‑limits, quotas), recoverable billing, and a SmartOps dashboard for evidence‑driven monitoring. It’s built with Go, Gin, React and supports SQLite, MySQL, PostgreSQL and Redis. Ideal for teams needing a secure, cost‑controlled, multi‑vendor model layer for agents and AI applications.

  • Project
    respanai/respan-gateway

    Respan Gateway is a universal adapter that lets you use any AI SDK to talk to any AI provider, with built-in translation, routing, tools, and observability.

  • Project
    ROCm/rocBLAS

    rocBLAS is a Basic Linear Algebra Subprograms library implemented in HIP and optimized for AMD GPUs to provide high-performance linear algebra operations.

  • Project
    zhiyu1998/Gemi2Api-Server

    Gemi2Api‑Server is a FastAPI‑based bridge that lets you call Google Gemini via an OpenAI‑compatible REST API. It authenticates with Gemini using your browser cookies, forwards chat‑completion requests, optionally proxies generated images, and provides a simple admin UI. Deployable locally, via Docker, or on Render/HuggingFace, it’s useful for developers who want Gemini access without a formal public API.