[Feature]: Support the Capacity Blocks for ML feature for AWS

[peterschmidt85]

Problem

With AWS, Capacity Blocks for ML is the main way to access H100 today.

Solution

No response

Workaround

No response

Would you like to help us implement this feature by sending a PR?

Yes

[r4victor]

@peterschmidt85, please describe what's meant by the support for Capacity Blocks. Do you suggest dstack provides an interface to purchase reserved instances or do you suggest dstack should be able to launch instances in capacity blocks purchased by users?

[r4victor]

Also note that testing EC2 Capacity Blocks for ML can be costly. The minimum reservation price is p4d.24xlarge for 1 day =. $14.75 * 24 = $354.

[peterschmidt85]

@peterschmidt85, please describe what's meant by the support for Capacity Blocks. Do you suggest dstack provides an interface to purchase reserved instances or do you suggest dstack should be able to launch instances in capacity blocks purchased by users?

@r4victor I think, we only need to add a way to specify the existing reservation ID. I believe this can be a field in YAML/Profile.

[peterschmidt85]

This issue is stale because it has been open for 30 days with no activity.