[peterschmidt85]

Kubernetes can infer the wrong VRAM from a GPU model name that has multiple capacities, while SSH fleet discovery rounds unmatched measurements. This can misrepresent available VRAM and affect memory-based scheduling.

• Read NVIDIA's nvidia.com/gpu.product and nvidia.com/gpu.memory labels and normalize measured VRAM through gpuhunt. Match both labels when scheduling so cards with different capacities stay distinct.
• Use the same normalization for SSH fleets, preserving measured MiB when no capacity matches. Both older string and newer numeric shim memory formats remain accepted.
• Warn with node or instance details when a model or capacity cannot be normalized. Kubernetes also warns when the measured-memory label is missing or invalid.

Slurm code and configuration syntax stay unchanged. Existing bare P100 mappings continue resolving to 16GB; A100 mappings continue specifying A100:40GB or A100:80GB.

Depends on dstackai/gpuhunt#279. Release that change and bump the gpuhunt dependency before merging this draft.

[peterschmidt85]

Created by mistake. Cloud-offer VRAM normalization is handled in gpuhunt; these dstack changes are not required.