Most people running PyTorch on a GPU are already running Triton without knowing it. PyTorch’s compiler guide states that for NVIDIA, AMD and Intel GPUs, torch.compile uses OpenAI Triton as the key building block. The kernels your compiled model executes are generated into a language whose point is not being CUDA.
What Triton is
The Triton repository describes itself as the development repository of a language and compiler for writing custom deep learning primitives, with the stated aim of providing an open-source environment to write fast code at higher productivity than CUDA and with more flexibility than other domain-specific languages. The documentation site puts it as a Python-based programming environment for writing custom neural network compute kernels. Its foundations were published at MAPL 2019 as “Triton: An Intermediate Language and Compiler for Tiled Neural Network Computations”.
Installation is pip install triton, with binary wheels for CPython 3.10 to 3.14. Triton uses LLVM to generate code for GPUs and CPUs, and a source build of LLVM for it enables the NVPTX and AMDGPU targets.
The compatibility section is short and worth reading before planning around it. Supported platforms: Linux. Supported hardware: NVIDIA GPUs with compute capability 8.0 or higher, AMD GPUs with ROCm 6.2 or newer, and CPUs listed as under development. Some tooling is vendor-specific in both directions, with an address sanitizer currently supported only on the AMD backend and PTXAS_OPTIONS applying only on NVIDIA. Version 2.0 rewrote the backend to use MLIR. Intel maintains a Triton backend for its own GPUs in a separate repository.
So Triton is portable across the two main vendors, not across every device, and it is a kernel language rather than a full stack. There is no cuDNN or NCCL equivalent inside it.
What Modular claims
Modular’s position is more sweeping, and it is a vendor claim rather than a measured result, so it is worth separating from the documented requirements.
The company says its vision is to unify a fragmented software stack with one programming language that can target CPUs, GPUs, custom accelerators and ASICs, and that Mojo is already able to target CPUs and GPUs from different vendors, which it calls the first language built for the AI era while adding that there is a lot of work left to do. On MAX, the marketing page claims the same package runs on NVIDIA, AMD and Apple silicon with no CUDA version to match, and under the heading “One stack, zero vendor lock-in” it says MAX does not depend on PyTorch, CUDA or ROCm, so there is nothing to bundle, patch or sync. Modular also states it reimplemented these stacks itself, building a graph compiler and runtime without CUDA, after arguing that earlier attempts failed because they were built on vendor-specific libraries such as cuBLAS or rocBLAS.
Those are the claims. Here is what the documentation commits to.
What Mojo documents
Mojo runs on Mac, Linux and Windows with WSL, and GPU support is optional. Linux needs glibc 2.34 or later, with Ubuntu 22.04 LTS or later the officially tested distribution. macOS needs Sequoia 15 or later on Apple silicon, M1 to M5. Windows is not supported natively. The CPU floor is x86-64-v3 or ARM64 Neoverse N1 or newer on Linux, and RAM starts at 8 GiB for development.
GPU compatibility is split into two levels that are unusually honest: continuously tested, meaning it runs in Modular’s CI on every release, and known compatible, meaning confirmed by Modular or community members but not continuously tested. For NVIDIA, the continuously tested entry is B200. Known compatible covers B300, B100, DGX Spark, H200, H100, L4, L40, RTX 50 and 40 series, A100, A10, RTX 30 series, Jetson Orin and Thor, T4 and RTX 20 series. The NVIDIA driver floor is 580 or later, with an environment variable to point at a system ptxas for older drivers. AMD entries run from MI355X and MI300X down to Radeon RX 9070, 7900, 7800, 7700 and 7600, integrated Radeon parts and the Steam Deck. Apple silicon lists M5 through M2 as known compatible.
Licensing splits. The repository and its contributions are under Apache 2.0 with LLVM exceptions, and the FAQ says Mojo is open source under the same licence. MAX usage and distribution are licensed separately under the Modular Community License.
FAQ
Is Triton a replacement for CUDA?
Not as a whole stack. It replaces hand-written CUDA C++ at the kernel level and compiles for NVIDIA and AMD GPUs, but it does not replace the surrounding libraries, and its supported platform list is Linux only.
Do I need to learn Triton to use torch.compile?
No. PyTorch generates Triton kernels for you. Learning it matters when you want a custom kernel that the compiler will not produce.
Is Mojo open source?
The Mojo FAQ says Mojo is open source under Apache 2.0 with LLVM exceptions. Modular’s repository notes it is not yet accepting contributions to the Mojo compiler itself, and MAX is distributed under the separate Modular Community License.
Has anyone verified Modular’s no-CUDA claim?
We have not tested it, and the statements above are Modular’s own. The checkable part is its GPU compatibility table, which marks only B200 on the NVIDIA side as continuously tested in CI.