[Bug]: RunPod clusters fail to provision due to minCudaVersion set

[r4victor]

Steps to reproduce

1. Try to provision a RunPod cluster via dstack, e.g. a two-pod H100 cluster is currently available
2. The provisioning fails as if there is no capacity with an obscure error from RunPod:

WARNING 2026-05-27T10:33:42.249 dstack._internal.server.background.pipeline_tasks.jobs_submitted
  job(50b3ad)tame-snake-1-0-0: NVIDIA H100 80GB HBM3 launch in runpod/EUR-IS-3 failed:
  RunpodApiClientError([{'message': 'Error creating cluster - cluster creation failed', 'path':
  ['createCluster'], 'extensions': {'code': 'RUNPOD'}}])

Actual behaviour

If I drop minCudaVersion: "12.8" from the request, then provisioning succeeds:

dstack/src/dstack/_internal/core/backends/runpod/api_client.py

Line 703 in d38ae9b

input_fields.append(f'minCudaVersion: "{RunpodProvider.MIN_CUDA_VERSION}"')

But the host does have an nvidia driver supporting cuda 12.8 so minCudaVersion seems to work incorrectly. Moreover specifying any lower value, e.g. "11", "11.1", all fail. My guess is that minCudaVersion does not work for CreateCluster even though it's listed in the reference.

Introduced in #3304 so RunPod clusters are not working since then.

As a workaround we can drop minCudaVersion from CreateCluster request until it's clarified/fixed on the RunPod side.

Expected behaviour

No response

dstack version

master

Server logs

Additional information

No response