Skip to content

Add Azure H100 NVL and H200 GPU support#220

Merged
peterschmidt85 merged 2 commits intodstackai:mainfrom
EzgiTastan:feature/azure-h100-h200-support
Mar 26, 2026
Merged

Add Azure H100 NVL and H200 GPU support#220
peterschmidt85 merged 2 commits intodstackai:mainfrom
EzgiTastan:feature/azure-h100-h200-support

Conversation

@EzgiTastan
Copy link
Copy Markdown
Contributor

Summary

Add support for Azure H100 NVL and H200 GPU virtual machines.

Changes

NC H100 v5-series:

  • Standard_NC40ads_H100_v5 (1x H100 NVL, 94GB)
  • Standard_NC80adis_H100_v5 (2x H100 NVL, 94GB each)
  • Pattern: NC(\d+)adi?s_H100_v5 — the i is optional (present only in NC80 for isolated instances)

ND H200 v5-series:

  • Standard_ND96isr_H200_v5 (8x H200, 141GB each)
  • Pattern: ND(\d+)isr_H200_v5

Correction to #101

The original suggestion used 80GB VRAM, but the H100 NVL has 94GB per GPU. The pattern also needed i? to match both NC40 (no i) and NC80 (has i).

References

Fixes #101

@peterschmidt85
Copy link
Copy Markdown
Contributor

The test errors above are fixed by #221

- Add NC H100 v5-series (Standard_NC40ads_H100_v5, Standard_NC80adis_H100_v5)
  with correct 94GB VRAM for H100 NVL GPUs
- Add ND H200 v5-series (Standard_ND96isr_H200_v5) with 141GB VRAM
- Update both gpu_vm_series and filter() method
Fixes dstackai#101
@peterschmidt85 peterschmidt85 force-pushed the feature/azure-h100-h200-support branch from b4798b9 to ec36633 Compare March 25, 2026 21:42
@peterschmidt85
Copy link
Copy Markdown
Contributor

@EzgiTastan also, did you have a chance to test the updated version of gpuhunt on your Azure?

@peterschmidt85
Copy link
Copy Markdown
Contributor

@EzgiTastan In order to try it with dstack, I suppose you need to run GPUHUNT_CATALOG_DIR=<path to the directory with generated azure.csv> dstack server

@peterschmidt85
Copy link
Copy Markdown
Contributor

peterschmidt85 commented Mar 25, 2026

The gpuhunt side looks good, but dstack won't show H100/H200 offers because of _SUPPORTED_VM_SERIES_PATTERNS in src/dstack/_internal/core/backends/azure/compute.py (line 420). It filters out any VM series not in the whitelist.

Add these two patterns:

r"NC(\d+)adi?s_H100_v5",    # NC H100 v5-series [H100 NVL 94GB]
r"ND(\d+)isr_H200_v5",      # ND H200 v5-series [8xH200 141GB]

To test end-to-end:

GPUHUNT_CATALOG_DIR=~/gpuhunt uv run dstack server
uv run dstack offer -b azure --gpu H100NVL
uv run dstack offer -b azure --gpu H200

@peterschmidt85

This comment was marked as resolved.

@peterschmidt85 peterschmidt85 merged commit 7ddfb24 into dstackai:main Mar 26, 2026
6 checks passed
peterschmidt85 pushed a commit that referenced this pull request Mar 26, 2026
Add H100NVL and H200 to expected GPU names in test_gpu_presented
following the addition of these VM series in #220.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
peterschmidt85 added a commit that referenced this pull request Mar 26, 2026
Add H100NVL and H200 to expected GPU names in test_gpu_presented
following the addition of these VM series in #220.

Co-authored-by: Andrey Cheptsov <andrey.cheptsov@github.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

Map H100 GPU on Azure

2 participants