Awesome

🚀 🔥 Fractional GPU! ⚡ 📣

Run multiple containers on the same GPU with driver level memory limitation ✨ and compute time-slicing 🎊

🌟 Leave a star to support the project! 🌟

🔰 Introduction

Sharing high-end GPUs or even prosumer & consumer GPUs between multiple users is the most cost-effective way to accelerate AI development. Unfortunately, until now the only existing solution applied for MIG/Slicing high-end GPUs (A100+) and required Kubernetes, <br>

🔥 🎉 Welcome To Container Based Fractional GPU For Any Nvidia Card! 🎉 🔥 <br>

We present pre-packaged containers supporting CUDA 11.x & CUDA 12.x with pre-built hard memory limitation! This means multiple containers can be launched on the same GPU, ensuring one user cannot allocate the entire host GPU memory! (No more greedy processes grabbing the entire GPU memory! Finally we have a driver level hard limiting memory option).

🚀 Offerings

ClearML offers several options to optimize GPU resource utilization by partitioning GPUs:

• Dynamic GPU Slicing: On-demand GPU slicing per task for both MIG and non-MIG devices (available under the ClearML Enterprise plan):
  • Bare Metal deployment
  • Kubernetes deployment
• Container-based Memory Limits (this repository): Use pre-packaged containers with built-in memory limits to run multiple containers on the same GPU (available as part of the ClearML open source offering).
• Kubernetes-based Static MIG Slicing: Set up Kubernetes support for NVIDIA MIG (Multi-Instance GPU) to define GPU fractions for specific workloads (available as part of the ClearML open source offering).

With these options, ClearML enables running AI workloads with optimized hardware utilization and workload performance. This repository covers container-based fractional GPUs. For more information on ClearML's fractional GPU offerings, see the ClearML documentation.

⚡ Installation

Pick the container that works for you and launch it:

docker run -it --gpus 0 --ipc=host --pid=host clearml/fractional-gpu:u22-cu12.3-8gb bash

To verify fraction GPU memory limit is working correctly, run inside the container:

nvidia-smi

Here is an example output from A100 GPU:

+---------------------------------------------------------------------------------------+
| NVIDIA-SMI 545.23.08              Driver Version: 545.23.08    CUDA Version: 12.3     |
|-----------------------------------------+----------------------+----------------------+
| GPU  Name                 Persistence-M | Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp   Perf          Pwr:Usage/Cap |         Memory-Usage | GPU-Util  Compute M. |
|                                         |                      |               MIG M. |
|=========================================+======================+======================|
|   0  A100-PCIE-40GB                Off  | 00000000:01:00.0 Off |                  N/A |
| 32%   33C    P0              66W / 250W |      0MiB /  8128MiB |      3%      Default |
|                                         |                      |             Disabled |
+-----------------------------------------+----------------------+----------------------+
                                                                                         
+---------------------------------------------------------------------------------------+
| Processes:                                                                            |
|  GPU   GI   CI        PID   Type   Process name                            GPU Memory |
|        ID   ID                                                             Usage      |
|=======================================================================================|
+---------------------------------------------------------------------------------------+

🐳 Containers

[!IMPORTANT]

You must execute the container with --pid=host !

[!NOTE]

--pid=host is required to allow the driver to differentiate between the container's processes and other host processes when limiting memory / utilization usage

[!TIP]

ClearML-Agent users add [--pid=host] to your agent.extra_docker_arguments section in your config file

🔩 Customization

Build your own containers and inherit form the original containers.

You can find a few examples here.

☸ Kubernetes

Fractional GPU containers can be used on bare-metal executions as well as Kubernetes PODs. Yes! By using one of the Fractional GPU containers you can limit the memory consumption of your Job/Pod and easily share GPUs without fearing they will memory crash one another!

Here's a simple Kubernetes POD template:

apiVersion: v1
kind: Pod
metadata:
  name: train-pod
  labels:
    app: trainme
spec:
  hostPID: true
  containers:
  - name: train-container
    image: clearml/fractional-gpu:u22-cu12.3-8gb
    command: ['python3', '-c', 'print(f"Free GPU Memory: (free, global) {torch.cuda.mem_get_info()}")']

[!IMPORTANT]

You must execute the pod with hostPID: true !

[!NOTE]

hostPID: true is required to allow the driver to differentiate between the pod's processes and other host processes when limiting memory / utilization usage

🔌 Support & Limitations

The containers support Nvidia drivers <= 545.x.x. We will keep updating & supporting new drivers as they continue to be released

Supported GPUs: RTX series 10, 20, 30, 40, A series, and Data-Center P100, A100, A10/A40, L40/s, H100

Limitations: Windows Host machines are currently not supported. If this is important for you, leave a request in the Issues section

❓ FAQ

• Q: Will running nvidia-smi inside the container report the local processes GPU consumption? <br> A: Yes, nvidia-smi is communicating directly with the low-level drivers and reports both accurate container GPU memory as well as the container local memory limitation.<br> Notice GPU utilization will be the global (i.e. host side) GPU utilization and not the specific local container GPU utilization.

• Q: How do I make sure my Python / Pytorch / Tensorflow are actually memory limited? <br> A: For PyTorch you can run: <br>

import torch
print(f'Free GPU Memory: (free, global) {torch.cuda.mem_get_info()}')

Numba example:

from numba import cuda
print(f'Free GPU Memory: {cuda.current_context().get_memory_info()}')

• Q: Can the limitation be broken by a user? <br> A: We are sure a malicious user will find a way. It was never our intention to protect against malicious users. <br> If you have a malicious user with access to your machines, fractional GPUs are not your number 1 problem 😃

• Q: How can I programmatically detect the memory limitation? <br> A: You can check the OS environment variable GPU_MEM_LIMIT_GB. <br> Notice that changing it will not remove or reduce the limitation.

• Q: Is running the container with --pid=host secure / safe? <br> A: It should be both secure and safe. The main caveat from a security perspective is that a container process can see any command line running on the host system. If a process command line contains a "secret" then yes, this might become a potential data leak. Notice that passing "secrets" in the command line is ill-advised, and hence we do not consider it a security risk. That said if security is key, the enterprise edition (see below) eliminate the need to run with pid-host and thus fully secure.

• Q: Can you run the container without --pid=host ? <br> A: You can! But you will have to use the enterprise version of the clearml-fractional-gpu container (otherwise the memory limit is applied system wide instead of container wide). If this feature is important for you, please contact ClearML sales & support.

📄 License

The license to use ClearML is granted for research or development purposes only. ClearML may be used for educational, personal, or internal commercial use.

An expanded Commercial license for use within a product or service is available as part of the ClearML Scale or Enterprise solution.

🤖 Commercial & Enterprise version

ClearML offers enterprise and commercial license adding many additional features on top of fractional GPUs, these include orchestration, priority queues, quota management, compute cluster dashboard, dataset management & experiment management, as well as enterprise grade security and support. Learn more about ClearML Orchestration or talk to us directly at ClearML sales.

📡 How can I help?

Tell everyone about it! #ClearMLFractionalGPU

Join our Slack Channel

Tell us when things are not working, and help us debug it on the Issues Page

🌟 Credits

This product is brought to you by the ClearML team with ❤️