Deploy NVIDIA NIM Models on PAI for High-Performance Inference - Platform for AI

NVIDIA Inference Microservice (NIM) provides pre-built containers with optimized AI model inference. These models deliver significant performance improvements over open source versions.

PAI Model Gallery offers multiple NIM models (Alibaba Cloud is an NVIDIA-authorized partner). Filter by Model Source: NIM to find available models. Deploy using either method:

Deploy in Model Gallery
Deploy locally

Available models

Deploy these NIM models directly in PAI Model Gallery:

Model name	Model Gallery page	Supported instance types
qwen2.5-7b-instruct-NIM	qwen2.5-7b-instruct-NIM	ecs.gn7e series ecs.gn8is series
MolMIM	MolMIM	General-purpose GPU instance types
Earth-2 FourCastNet	FourCastNet	General-purpose GPU instance types
NVIDIA Retrieval QA Mistral 7B Embedding v2	NV-EmbedQA-Mistral-7b-v2	ecs.gn7e series
Eye Contact	eyecontact	General-purpose GPU instance types
NV-CLIP	nvclip	ecs.gn7e series ecs.gn7i series
AlphaFold2-Multimer	AlphaFold2-Multimer	General-purpose GPU instance types
Snowflake Arctic Embed Large Embedding	Arctic-embed-l	ecs.gn7e series ecs.gn7i series
NVIDIA Retrieval QA Mistral 4B Reranking v3	NV-RerankQA-Mistral-4B-v3	ecs.gn7e series ecs.gn7i series
NVIDIA Retrieval QA E5 Embedding v5	NV-EmbedQA-E5-v5	ecs.gn7e series ecs.gn7i series
Parakeet CTC Riva 1.1b	Parakeet-CTC-1.1b-asr	General-purpose GPU instance types
FastPitch HifiGAN Riva	FastPitchHifi-GAN	General-purpose GPU instance types
VISTA-3D	Vista-3D	General-purpose GPU instance types
AlphaFold2	AlphaFold2	General-purpose GPU instance types
ProteinMPNN	ProteinMPNN	General-purpose GPU instance types
megatron-1b-nmt	megatron-1b-nmt	General-purpose GPU instance types

Deploy in Model Gallery

Go to PAI Model Gallery.
In the filter pane, set Model Source to NIM.
Select a model to open its details page. Click Deploy.

Prerequisite: You must be an NVIDIA AI Enterprise user or NVIDIA Developer Program user.
Configure runtime resources and click Deploy to create an online service. For invocation instructions, see the model introduction page.

Deploy locally

Download image and model files to deploy locally via Docker.

Prerequisite: You must be an NVIDIA AI Enterprise user or NVIDIA Developer Program user.

Configure the environment. See NVIDIA Getting Started.
On the model details page, click Download Address, accept the NIM terms and license, and obtain image and model addresses.
Pull the image. Replace ${image_address} with your actual address.
```
docker pull ${image_address}
```
Download the model file using ossutil.
Start the container. This example assumes the model file is in /local/model/. Replace ${model_mount_path} and ${image_address} with actual values.
```
docker run --rm \
--runtime=nvidia \
--gpus all \
-u $(id -u) \
-v /local/model/:${model_mount_path} ${image_address}
```

First-time setup

If this is your first time using PAI:

Go to Alibaba Cloud and click Log On. Log in or register a new account.
Complete identity verification, then go to Platform for AI (PAI).

On first use, complete identity verification and grant authorization (leave defaults, takes ~10 seconds). You can then deploy models in the default workspace.