Configure NeMo Retriever Text Embedding NIM#

NeMo Text Retriever NIM use docker containers under the hood. Each NIM is its own Docker container and there are several ways to configure it. The remainder of this section describes the various ways to configure a NIM container.

Use this documentation to learn how to configure NeMo Retriever Text Embedding NIM.

GPU Selection#

The NIM container is GPU-accelerated and uses NVIDIA Container Toolkit for access to GPUs on the host.

You can specify the --gpus all command-line argument to the docker run command if the host has one or more of the same GPU model. If the host has a combination of GPUs, such as an A6000 and a GeForce display GPU, run the container on compute-capable GPUs only.

Expose specific GPUs to the container by using either of the following methods:

  • Specify the --gpus argument, such as --gpus="device=1".

  • Set the NVIDIA_VISIBLE_DEVICES environment variable, such as -e NVIDIA_VISIBLE_DEVICES=1.

Run the nvidia-smi -L command to list the device IDs to specify in the argument or environment variable:

GPU 0: Tesla H100 (UUID: GPU-b404a1a1-d532-5b5c-20bc-b34e37f3ac46)
GPU 1: NVIDIA GeForce RTX 3080 (UUID: GPU-b404a1a1-d532-5b5c-20bc-b34e37f3ac46)

Refer to GPU Enumeration in the NVIDIA Container Toolkit documentation for more information.

Memory Footprint#

You can configure the memory footprint of the NeMo Retriever Text Embedding NIM by adjusting the model’s maximum allowed batch size and sequence length.

For ONNX model profiles, memory is allocated dynamically according to the requests. A maximum batch size and sequence length limit memory usage. You can specify a value for maximum batch size and sequence length from 1 up to the maximum supported limit for a given model and GPU.

For TensorRT model profiles, memory is allocated statically based on the optimized static inference graph which has a defined maximum input shape. You must specify a value from a discrete set of options. Refer to the support matrix for the valid values and corresponding approximate memory footprint.

By default, the NIM uses the largest possible value (given the model and GPU constraints) for both the maximum batch size and the maximum sequence length. If you specify only one of these parameters, the NIM uses the largest possible value for the unspecified parameter. For example, if you only specify a limit for maximum batch size, the NIM uses the largest possible sequence length.

Performance Mode#

The NeMo Retriever Text Embedding NIM can run in modes optimized for low latency or high throughput. The performance mode is controlled by the NIM_TRITON_PERFORMANCE_MODE environment variable.

When set to latency optimized mode (default), the NIM optimizes for minimum latency on short sequence length and low batch size workloads. When running in this mode, the NIM provides good performance on all workloads but is not as performant as throughput mode on the highest batch sizes and longest sequence lengths.

When set to throughput optimized mode, the NIM optimizes for maximum throughput on long sequence length and high batch size workloads. This mode provides upwards of 2x performance when used with large batches containing 64-1000 long and diverse length sequences. This mode has higher latency when processing short sequences in small batches.

PID Limit#

In certain deployment or container runtime environments, default process and thread limits (PID limits) can interfere with NIM startup. These set limits are set by Docker, Podman, Kubernetes, or the operating system.

If the PID limit is too low, you might see symptoms such as:

  • NIM starts up partially, but fails to reach ready state, and then stalls.

  • NIM starts up partially, but fails to reach ready state, and then crashes.

  • NIM serves a small number of requests, and then fails.

To verify that PID limits are impacting the NIM container, you can remove or adjust the PID limit at the container, node, and operating system level. Removing the PID limit and then checking for success is a useful diagnostic step.

  • To increase the PID limit in a docker run command, set --pids-limit=-1. For details, see docker container run.

  • To increase the PID limit in a podman run command, --pids-limit=-1. For details, see Podman pids-limit.

  • To increase the PID limit in Kubernetes, set the PodPidsLimit on the kubelet on each node. For details, see your Kubernetes distribution specific documentation.

  • To increase the PID limit at the operating system level, see your OS-specific documentation.

Shared Memory Flag#

Tokenization uses Triton’s Python backend capabilities that scales with the number of CPU cores available. You may need to increase the available shared memory given to the microservice container.

Example providing 1g of shared memory:

docker run ... --shm-size=1g ...

Volumes#

The following table identifies the paths that are used in the container. Use this information to plan the local paths to bind mount into the container.

Container Path

Description

Example

/opt/nim/.cache or NIM_CACHE_PATH

Specifies the path, relative to the root of the container, for downloaded models.

The typical use for this path is to bind mount a directory on the host with this path inside the container. For example, to use ~/.cache/nim on the host, run mkdir -p ~/.cache/nim before you start the container. When you start the container, specify the -v ~/.cache/nim:/opt/nim/.cache -u $(id -u) arguments to the docker run command.

If you do not specify a bind or volume mount, as shown in the -v argument in the preceding command, the container downloads the model each time you start the container.

The -u $(id -u) argument runs the container with your user ID to avoid file system permission issues and errors.

-v ~/.cache/nim:/opt/nim/.cache -u $(id -u)