> This page is for version 1.0.0 (Latest) (default).
> For other versions, use one of these documentation indexes:
> - 1.0.0 (Latest) (default): https://docs.nvidia.com/nim/harmonizer/1.0.0/llms.txt

> For clean Markdown of any page, append .md to the page URL.
> For a complete documentation index, see https://docs.nvidia.com/nim/harmonizer/llms.txt.
> For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.nvidia.com/nim/harmonizer/_mcp/server.

# Harmonizer NIM Prerequisites

> Check the hardware, software, and verified configurations required to run Harmonizer NIM.

This page outlines the hardware and software requirements for running
Harmonizer NIM.

## Hardware Requirements

Confirm the following hardware requirements:

* **GPU**: One supported NVIDIA GPU per container. The smallest qualified
  configuration is specifically the 24 GB GeForce RTX 4090; this does not imply
  support for every GPU with 24 GB of memory. The supported 48 GB and
  80 GB-or-larger configurations use progressively larger startup
  specialization sets. Refer to the [Support Matrix](/nim/harmonizer/support-matrix) for the exact
  GPU SKUs.
* **CPU architecture**: x86_64 (amd64) and ARM64 are supported.
* **CPU memory**: Allocate at least 32 GiB of host memory per container for
  the recommended configuration.
* **Disk space**: Reserve 50 GiB of free space before the first pull. This
  accommodates the unpacked image, model checkpoint, generated compilation
  cache, compressed layers retained temporarily while Docker unpacks the image,
  and operating headroom.

### Release-Candidate Sizing Reference

The following measurements were collected with a pre-release container and
can differ from the published 1.0.0 image. Sizes are binary GiB. The runtime
measurements used an L40S on an x86_64 host, an empty compilation cache, cold
startup, readiness, and one `1024x576` HTTP inference request.

| Item                                                        | Measured Size            |
| ----------------------------------------------------------- | ------------------------ |
| Compressed registry download, amd64                         | 9.82 GiB                 |
| Compressed registry download, arm64                         | 10.49 GiB                |
| Unpacked Docker image, amd64                                | 22.89 GiB                |
| Release model checkpoint                                    | 4.70 GiB                 |
| Generated compilation cache after readiness and inference   | 0.45 GiB                 |
| Peak container CPU memory during cold startup and inference | 10.23 GiB                |
| Passing Docker memory limit in one controlled L40S run      | 16.00 GiB, swap disabled |
| Peak L40S GPU memory during cold startup and inference      | 35.4 GiB                 |

If Docker's data root and the persistent NIM cache are on separate file systems,
budget at least 35 GiB for Docker and 10 GiB for the NIM cache. The model and
generated TorchInductor and Triton artifacts are stored in the persistent cache
mounted at `/opt/nim/.cache`.

## Software Requirements

Install the following software:

* **Operating system**: A 64-bit Linux distribution [supported by NVIDIA Container Toolkit](https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/latest/supported-platforms.html).
  The Harmonizer image is published for `linux/amd64` and `linux/arm64` and
  uses Ubuntu 24.04 container userspace.
* **NVIDIA GPU driver**: Version 595.58.03 or later. The image bundles CUDA
  13.2.1.009 and performs TorchInductor and Triton PTX compilation during cold
  startup. For this reason, use the corresponding CUDA 13.2 Update 1 driver
  baseline instead of relying on the broader CUDA 13.x minor-compatibility
  floor. Refer to the [CUDA Toolkit release notes](https://docs.nvidia.com/cuda/archive/13.2.2/cuda-toolkit-release-notes/index.html)
  and [minor-version compatibility limitations](https://docs.nvidia.com/deploy/cuda-compatibility/minor-version-compatibility.html).
* **CUDA Toolkit**: Not required on the host. The container includes the CUDA
  runtime and libraries. Refer to the [NVIDIA PyTorch 26.05 release notes](https://docs.nvidia.com/deeplearning/frameworks/pytorch-release-notes/rel-26-05.html)
  for the bundled runtime stack.
* **Container runtime**: Docker 23.0.1 or later, or a compatible OCI runtime
  that meets the NVIDIA NIM platform requirements. Follow the [Docker installation guide](https://docs.docker.com/engine/install/).
* **NVIDIA Container Toolkit**: Version 1.16.2 or later. Follow the [NVIDIA Container Toolkit installation guide](https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/latest/install-guide.html#installing-the-nvidia-container-toolkit).

## Verified Configurations

The release-candidate validation run used Ubuntu 24.04.4 on x86_64, NVIDIA
driver 595.58.03, Docker 29.6.1, and NVIDIA Container Toolkit 1.19.1. These
versions are a validated configuration, not higher minimums than those listed
above.

## Configuration and Verification

After installing the NVIDIA Container Toolkit, configure Docker by following the
NVIDIA Container Toolkit documentation.

To verify that Docker can access the GPU, run:

```bash
docker run --rm --runtime=nvidia --gpus all ubuntu nvidia-smi
```

The command should list the available NVIDIA GPUs and driver version.

## Next Steps

After checking these requirements and verifying GPU access, follow the [Quickstart](/nim/harmonizer/quickstart-guide) to launch the NIM.