Update Azure catalog (azure/vms.csv) to add H100.

[concretevitamin]

The new catalog is obtained by checking out skypilot-org/skypilot#2844 and running

python <sky repo>/sky/clouds/service_catalog/data_fetchers/fetch_azure.py

New catalog contains H100:

» grep H100 *csv                                                                                                      
Standard_NC40ads_H100_v5,H100,1,40.0,320,H100,,,eastus,V2
Standard_NC40ads_H100_v5,H100,1,40.0,320,H100,,,southcentralus,V2
Standard_NC80adis_H100_v5,H100,2,80.0,640,H100,,,eastus,V2
Standard_NC80adis_H100_v5,H100,2,80.0,640,H100,,,southcentralus,V2
Standard_ND48s_H100_v5,H100,6,48.0,950,H100,,,northcentralus,V2
Standard_ND96isr_H100_v5,H100,12,96.0,1900,H100,117.984,29.496,northcentralus,V2
Standard_ND96is_H100_v5,H100,12,96.0,1900,H100,106.186,26.5465,northcentralus,V2

Known problems

• First 5 lines above have Price and SpotPrice set to NaN -- Azure's pricing URLs currently do not return any pricing records for them, likely because they are in-preview
• This means for example, sky launch -t Standard_NC40ads_H100_v5 fails, because pricing is not available

Workaround

• If users want to use these instance types, they need to manually set the prices to 0 for these rows, in their azure/vms.csv
  • Tested: manually setting to 0 and running the above launch works

Note

• Last two lines have non-NaN prices, which means Standard_ND96isr_H100_v5 and Standard_ND96is_H100_v5 should be launchable by people with quotas
• On my laptop, SkyPilot recognizes them and proceeds to launch (then running into quota issues)

» sky launch --gpus H100:12                                                                                             
I 12-05 22:13:05 optimizer.py:694] == Optimizer ==
I 12-05 22:13:05 optimizer.py:717] Estimated cost: $106.2 / hour
I 12-05 22:13:05 optimizer.py:717]
I 12-05 22:13:05 optimizer.py:841] Considered resources (1 node):
I 12-05 22:13:05 optimizer.py:910] ---------------------------------------------------------------------------------------------------------
I 12-05 22:13:05 optimizer.py:910]  CLOUD   INSTANCE                  vCPUs   Mem(GB)   ACCELERATORS   REGION/ZONE      COST ($)   CHOSEN
I 12-05 22:13:05 optimizer.py:910] ---------------------------------------------------------------------------------------------------------
I 12-05 22:13:05 optimizer.py:910]  Azure   Standard_ND96is_H100_v5   96      1900      H100:12        northcentralus   106.19        ✔
I 12-05 22:13:05 optimizer.py:910] ---------------------------------------------------------------------------------------------------------

[Michaelvll]

Would it be good to add the same price of ND or just $0 for the preview H100 NC family? Otherwise, a normal user may have to change the catalog manually, which is a bit demanding.

[concretevitamin]

Would it be good to add the same price of ND or just $0 for the preview H100 NC family? Otherwise, a normal user may have to change the catalog manually, which is a bit demanding.

Good call, I think a fake price, like $0 or $9999, is better to signal that it's temporary. Going with $0 for now since most users won't have quotas anyway.

[Michaelvll]

Would it be good to add the same price of ND or just $0 for the preview H100 NC family? Otherwise, a normal user may have to change the catalog manually, which is a bit demanding.

Good call, I think a fake price, like $0 or $9999, is better to signal that it's temporary. Going with $0 for now since most users won't have quotas anyway.

$0 sounds good to me! I am wondering if it is possible to enable the GitHub Action that fetch for the Azure catalog automatically as well considering more people are going to use Azure.

[concretevitamin]

Updated, PTAL.

RE: Action. Maybe still ok to keep commit history smaller for now?