diff --git a/src/dstack/_internal/core/backends/verda/compute.py b/src/dstack/_internal/core/backends/verda/compute.py index e5b988e69..279c73c7e 100644 --- a/src/dstack/_internal/core/backends/verda/compute.py +++ b/src/dstack/_internal/core/backends/verda/compute.py @@ -234,14 +234,13 @@ def update_provisioning_data( def _get_vm_image_id(instance_offer: InstanceOfferWithAvailability) -> str: # https://api.datacrunch.io/v1/docs#tag/images/GET/v1/images - if len(instance_offer.instance.resources.gpus) > 0 and instance_offer.instance.resources.gpus[ - 0 - ].name in ["V100", "A6000"]: - # * A6000: Verda returns "Operating system is not valid for this instance type" for newer images - # Ubuntu 22.04 + CUDA 12.0 + Docker - return "2088da25-bb0d-41cc-a191-dccae45d96fd" + gpus = instance_offer.instance.resources.gpus + if len(gpus) > 0 and gpus[0].name == "V100": + # V100: Verda lists only this image as supported + # Ubuntu 24.04 + CUDA 12.9 (V100, proprietary driver 570) + Docker + return "cf310c99-b3f8-4227-baa4-d473c16f9ae1" # Ubuntu 24.04 + CUDA 13.0 Open + Docker - return "2404f580-1300-1111-82b3-f199fb9b701b" + return "59cd1ccb-499e-4053-9690-3ca4c898351b" def _create_ssh_key(client: VerdaClient, name: str, public_key: str) -> str: diff --git a/src/tests/_internal/core/backends/verda/test_compute.py b/src/tests/_internal/core/backends/verda/test_compute.py index 8edabaf0c..ab3edcc7b 100644 --- a/src/tests/_internal/core/backends/verda/test_compute.py +++ b/src/tests/_internal/core/backends/verda/test_compute.py @@ -9,10 +9,12 @@ VerdaInstanceBackendData, _create_ssh_key, _create_startup_script, + _get_vm_image_id, ) from dstack._internal.core.errors import BackendError, NoCapacityError from dstack._internal.core.models.backends.base import BackendType from dstack._internal.core.models.instances import ( + Gpu, InstanceAvailability, InstanceConfiguration, InstanceOffer, @@ -48,6 +50,12 @@ def _instance_offer() -> InstanceOfferWithAvailability: ) +def _gpu_instance_offer(gpu_name: str) -> InstanceOfferWithAvailability: + offer = _instance_offer() + offer.instance.resources.gpus = [Gpu(name=gpu_name, memory_mib=16384)] + return offer + + def _instance_config(public_keys: list[str]) -> InstanceConfiguration: return InstanceConfiguration( project_name="main", @@ -121,6 +129,17 @@ def test_raises_backend_error_on_api_exception(self): ) +class TestGetVMImageId: + def test_v100_uses_separate_image(self): + assert _get_vm_image_id(_gpu_instance_offer("V100")) != _get_vm_image_id(_instance_offer()) + + @pytest.mark.parametrize("gpu_name", ["A6000", "H100"]) + def test_other_gpus_use_default_image(self, gpu_name): + assert _get_vm_image_id(_gpu_instance_offer(gpu_name)) == _get_vm_image_id( + _instance_offer() + ) + + class TestCreateInstance: def test_cleans_up_created_ssh_keys_if_later_ssh_key_create_fails(self): compute = VerdaCompute.__new__(VerdaCompute)