ClearML/clearml-docs
1
0
Fork 0
mirror of https://github.com/clearml/clearml-docs synced 2025-02-07 13:21:46 +00:00
Code Issues Actions 1 Packages Projects Releases Wiki Activity

Rewrite K8s static MIG fraction section (#876)

Browse Source
This commit is contained in:
pollfly 2024-07-16 14:18:11 +03:00 committed by GitHub
parent ce6cfecc8a
commit cbb65cb974
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 10 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

11
docs/clearml_agent/clearml_agent_fractional_gpus.md
View File

@ -317,9 +317,10 @@ See example custom Dockerfiles in the [clearml-fractional-gpu repository](https:
Set up NVIDIA MIG (Multi-Instance GPU) support for Kubernetes to define GPU fraction profiles for specific workloads Set up NVIDIA MIG (Multi-Instance GPU) support for Kubernetes to define GPU fraction profiles for specific workloads
through your NVIDIA device plugin. through your NVIDIA device plugin.
The ClearML Agent Helm chart lets you specify a pod template for each queue which describes the resources that the pod The standard way to configure a Kubernetes pod template to use specific MIG slices is for the template to specify the
will use. The template should specify the requested GPU slices under `Containers.resources.limits` to have the pods use requested GPU slices under `Containers.resources.limits`. For example, the
the defined resources. For example, the following configures a K8s pod to run a `3g.20gb` MIG device: following configures a K8s pod to run a 3g.20gb MIG device:
``` ```
# tf-benchmarks-mixed.yaml # tf-benchmarks-mixed.yaml
apiVersion: v1 apiVersion: v1
@ -340,6 +341,10 @@ spec:
nvidia.com/gpu.product: A100-SXM4-40GB nvidia.com/gpu.product: A100-SXM4-40GB
``` ```
The ClearML Agent Helm chart lets you specify a pod template for each queue which describes the resources that the pod
will use. The ClearML Agent uses this configuration to generate the necessary Kubernetes pod template for executing
tasks based on the queue through which they are scheduled.
When tasks are added to the relevant queue, the agent pulls the task and creates a pod to execute it, using the When tasks are added to the relevant queue, the agent pulls the task and creates a pod to execute it, using the
specified GPU slice. specified GPU slice.