dstack 0.19.28__py3-none-any.whl → 0.19.30__py3-none-any.whl

dstack/_internal/core/backends/cloudrift/compute.py

@@ -1,7 +1,8 @@
 from typing import Dict, List, Optional
 
-from dstack._internal.core.backends.base.backend import Compute
 from dstack._internal.core.backends.base.compute import (
+    Compute,
+    ComputeWithAllOffersCached,
     ComputeWithCreateInstanceSupport,
     get_shim_commands,
 )
@@ -17,13 +18,14 @@ from dstack._internal.core.models.instances import (
     InstanceOfferWithAvailability,
 )
 from dstack._internal.core.models.placement import PlacementGroup
-from dstack._internal.core.models.runs import JobProvisioningData, Requirements
+from dstack._internal.core.models.runs import JobProvisioningData
 from dstack._internal.utils.logging import get_logger
 
 logger = get_logger(__name__)
 
 
 class CloudRiftCompute(
+    ComputeWithAllOffersCached,
     ComputeWithCreateInstanceSupport,
     Compute,
 ):
@@ -32,15 +34,11 @@ class CloudRiftCompute(
         self.config = config
         self.client = RiftClient(self.config.creds.api_key)
 
-    def get_offers(
-        self, requirements: Optional[Requirements] = None
-    ) -> List[InstanceOfferWithAvailability]:
+    def get_all_offers_with_availability(self) -> List[InstanceOfferWithAvailability]:
         offers = get_catalog_offers(
             backend=BackendType.CLOUDRIFT,
             locations=self.config.regions or None,
-            requirements=requirements,
         )
-
         offers_with_availabilities = self._get_offers_with_availability(offers)
         return offers_with_availabilities
 

dstack/_internal/core/backends/cudo/compute.py

@@ -5,6 +5,7 @@ import requests
 from dstack._internal.core.backends.base.backend import Compute
 from dstack._internal.core.backends.base.compute import (
     ComputeWithCreateInstanceSupport,
+    ComputeWithFilteredOffersCached,
     generate_unique_instance_name,
     get_shim_commands,
 )
@@ -29,6 +30,7 @@ MAX_RESOURCE_NAME_LEN = 30
 
 
 class CudoCompute(
+    ComputeWithFilteredOffersCached,
     ComputeWithCreateInstanceSupport,
     Compute,
 ):
@@ -37,8 +39,8 @@ class CudoCompute(
         self.config = config
         self.api_client = CudoApiClient(config.creds.api_key)
 
-    def get_offers(
-        self, requirements: Optional[Requirements] = None
+    def get_offers_by_requirements(
+        self, requirements: Requirements
     ) -> List[InstanceOfferWithAvailability]:
         offers = get_catalog_offers(
             backend=BackendType.CUDO,

dstack/_internal/core/backends/datacrunch/compute.py

@@ -1,4 +1,4 @@
-from typing import Dict, List, Optional
+from typing import Callable, Dict, List, Optional
 
 from datacrunch import DataCrunchClient
 from datacrunch.exceptions import APIException
@@ -6,11 +6,12 @@ from datacrunch.instances.instances import Instance
 
 from dstack._internal.core.backends.base.backend import Compute
 from dstack._internal.core.backends.base.compute import (
+    ComputeWithAllOffersCached,
     ComputeWithCreateInstanceSupport,
     generate_unique_instance_name,
     get_shim_commands,
 )
-from dstack._internal.core.backends.base.offers import get_catalog_offers
+from dstack._internal.core.backends.base.offers import get_catalog_offers, get_offers_disk_modifier
 from dstack._internal.core.backends.datacrunch.models import DataCrunchConfig
 from dstack._internal.core.errors import NoCapacityError
 from dstack._internal.core.models.backends.base import BackendType
@@ -36,6 +37,7 @@ CONFIGURABLE_DISK_SIZE = Range[Memory](min=IMAGE_SIZE, max=None)
 
 
 class DataCrunchCompute(
+    ComputeWithAllOffersCached,
     ComputeWithCreateInstanceSupport,
     Compute,
 ):
@@ -47,18 +49,19 @@ class DataCrunchCompute(
             client_secret=self.config.creds.client_secret,
         )
 
-    def get_offers(
-        self, requirements: Optional[Requirements] = None
-    ) -> List[InstanceOfferWithAvailability]:
+    def get_all_offers_with_availability(self) -> List[InstanceOfferWithAvailability]:
         offers = get_catalog_offers(
             backend=BackendType.DATACRUNCH,
             locations=self.config.regions,
-            requirements=requirements,
-            configurable_disk_size=CONFIGURABLE_DISK_SIZE,
         )
         offers_with_availability = self._get_offers_with_availability(offers)
         return offers_with_availability
 
+    def get_offers_modifier(
+        self, requirements: Requirements
+    ) -> Callable[[InstanceOfferWithAvailability], Optional[InstanceOfferWithAvailability]]:
+        return get_offers_disk_modifier(CONFIGURABLE_DISK_SIZE, requirements)
+
     def _get_offers_with_availability(
         self, offers: List[InstanceOffer]
     ) -> List[InstanceOfferWithAvailability]:
@@ -182,10 +185,9 @@ class DataCrunchCompute(
 
 def _get_vm_image_id(instance_offer: InstanceOfferWithAvailability) -> str:
     # https://api.datacrunch.io/v1/images
-    if (
-        len(instance_offer.instance.resources.gpus) > 0
-        and instance_offer.instance.resources.gpus[0].name == "V100"
-    ):
+    if len(instance_offer.instance.resources.gpus) > 0 and instance_offer.instance.resources.gpus[
+        0
+    ].name in ["V100", "A6000"]:
         # Ubuntu 22.04 + CUDA 12.0 + Docker
         return "2088da25-bb0d-41cc-a191-dccae45d96fd"
     # Ubuntu 24.04 + CUDA 12.8 Open + Docker

dstack/_internal/core/backends/digitalocean_base/compute.py

@@ -5,6 +5,7 @@ from gpuhunt.providers.digitalocean import DigitalOceanProvider
 
 from dstack._internal.core.backends.base.backend import Compute
 from dstack._internal.core.backends.base.compute import (
+    ComputeWithAllOffersCached,
     ComputeWithCreateInstanceSupport,
     generate_unique_instance_name,
     get_user_data,
@@ -20,7 +21,7 @@ from dstack._internal.core.models.instances import (
     InstanceOfferWithAvailability,
 )
 from dstack._internal.core.models.placement import PlacementGroup
-from dstack._internal.core.models.runs import JobProvisioningData, Requirements
+from dstack._internal.core.models.runs import JobProvisioningData
 from dstack._internal.utils.logging import get_logger
 
 logger = get_logger(__name__)
@@ -37,6 +38,7 @@ DOCKER_INSTALL_COMMANDS = [
 
 
 class BaseDigitalOceanCompute(
+    ComputeWithAllOffersCached,
     ComputeWithCreateInstanceSupport,
     Compute,
 ):
@@ -50,13 +52,10 @@ class BaseDigitalOceanCompute(
             DigitalOceanProvider(api_key=config.creds.api_key, api_url=api_url)
         )
 
-    def get_offers(
-        self, requirements: Optional[Requirements] = None
-    ) -> List[InstanceOfferWithAvailability]:
+    def get_all_offers_with_availability(self) -> List[InstanceOfferWithAvailability]:
         offers = get_catalog_offers(
             backend=self.BACKEND_TYPE,
             locations=self.config.regions,
-            requirements=requirements,
             catalog=self.catalog,
         )
         return [

dstack/_internal/core/backends/gcp/compute.py

@@ -17,6 +17,7 @@ import dstack._internal.core.backends.gcp.resources as gcp_resources
 from dstack import version
 from dstack._internal.core.backends.base.compute import (
     Compute,
+    ComputeWithAllOffersCached,
     ComputeWithCreateInstanceSupport,
     ComputeWithGatewaySupport,
     ComputeWithMultinodeSupport,
@@ -30,10 +31,15 @@ from dstack._internal.core.backends.base.compute import (
     get_shim_commands,
     get_user_data,
     merge_tags,
+    requires_nvidia_proprietary_kernel_modules,
+)
+from dstack._internal.core.backends.base.offers import (
+    get_catalog_offers,
+    get_offers_disk_modifier,
 )
-from dstack._internal.core.backends.base.offers import get_catalog_offers
 from dstack._internal.core.backends.gcp.features import tcpx as tcpx_features
 from dstack._internal.core.backends.gcp.models import GCPConfig
+from dstack._internal.core.consts import DSTACK_OS_IMAGE_WITH_PROPRIETARY_NVIDIA_KERNEL_MODULES
 from dstack._internal.core.errors import (
     ComputeError,
     ComputeResourceNotFoundError,
@@ -82,6 +88,7 @@ class GCPVolumeDiskBackendData(CoreModel):
 
 
 class GCPCompute(
+    ComputeWithAllOffersCached,
     ComputeWithCreateInstanceSupport,
     ComputeWithMultinodeSupport,
     ComputeWithPlacementGroupSupport,
@@ -107,14 +114,10 @@ class GCPCompute(
         self._extra_subnets_cache_lock = threading.Lock()
         self._extra_subnets_cache = TTLCache(maxsize=30, ttl=60)
 
-    def get_offers(
-        self, requirements: Optional[Requirements] = None
-    ) -> List[InstanceOfferWithAvailability]:
+    def get_all_offers_with_availability(self) -> List[InstanceOfferWithAvailability]:
         regions = get_or_error(self.config.regions)
         offers = get_catalog_offers(
             backend=BackendType.GCP,
-            requirements=requirements,
-            configurable_disk_size=CONFIGURABLE_DISK_SIZE,
             extra_filter=_supported_instances_and_zones(regions),
         )
         quotas: Dict[str, Dict[str, float]] = defaultdict(dict)
@@ -142,9 +145,13 @@ class GCPCompute(
             offer_keys_to_offers[key] = offer_with_availability
             offers_with_availability.append(offer_with_availability)
             offers_with_availability[-1].region = region
-
         return offers_with_availability
 
+    def get_offers_modifier(
+        self, requirements: Requirements
+    ) -> Callable[[InstanceOfferWithAvailability], Optional[InstanceOfferWithAvailability]]:
+        return get_offers_disk_modifier(CONFIGURABLE_DISK_SIZE, requirements)
+
     def terminate_instance(
         self, instance_id: str, region: str, backend_data: Optional[str] = None
     ) -> None:
@@ -288,7 +295,11 @@ class GCPCompute(
 
         image = _get_image(
             instance_type_name=instance_offer.instance.name,
-            cuda=len(instance_offer.instance.resources.gpus) > 0,
+            gpu_name=(
+                instance_offer.instance.resources.gpus[0].name
+                if len(instance_offer.instance.resources.gpus) > 0
+                else None
+            ),
         )
 
         for zone in zones:
@@ -899,7 +910,7 @@ class GCPImage:
     is_ufw_installed: bool
 
 
-def _get_image(instance_type_name: str, cuda: bool) -> GCPImage:
+def _get_image(instance_type_name: str, gpu_name: Optional[str]) -> GCPImage:
     if instance_type_name == "a3-megagpu-8g":
         image_name = "dstack-a3mega-5"
         is_ufw_installed = False
@@ -908,8 +919,11 @@ def _get_image(instance_type_name: str, cuda: bool) -> GCPImage:
             id="projects/cos-cloud/global/images/cos-105-17412-535-78",
             is_ufw_installed=False,
         )
-    elif cuda:
-        image_name = f"dstack-cuda-{version.base_image}"
+    elif gpu_name is not None:
+        if not requires_nvidia_proprietary_kernel_modules(gpu_name):
+            image_name = f"dstack-cuda-{version.base_image}"
+        else:
+            image_name = f"dstack-cuda-{DSTACK_OS_IMAGE_WITH_PROPRIETARY_NVIDIA_KERNEL_MODULES}"
         is_ufw_installed = True
     else:
         image_name = f"dstack-{version.base_image}"

dstack/_internal/core/backends/hotaisle/compute.py

@@ -9,6 +9,7 @@ from gpuhunt.providers.hotaisle import HotAisleProvider
 
 from dstack._internal.core.backends.base.compute import (
     Compute,
+    ComputeWithAllOffersCached,
     ComputeWithCreateInstanceSupport,
     get_shim_commands,
 )
@@ -23,7 +24,7 @@ from dstack._internal.core.models.instances import (
     InstanceOfferWithAvailability,
 )
 from dstack._internal.core.models.placement import PlacementGroup
-from dstack._internal.core.models.runs import JobProvisioningData, Requirements
+from dstack._internal.core.models.runs import JobProvisioningData
 from dstack._internal.utils.logging import get_logger
 
 logger = get_logger(__name__)
@@ -44,6 +45,7 @@ INSTANCE_TYPE_SPECS = {
 
 
 class HotAisleCompute(
+    ComputeWithAllOffersCached,
     ComputeWithCreateInstanceSupport,
     Compute,
 ):
@@ -56,16 +58,12 @@ class HotAisleCompute(
             HotAisleProvider(api_key=config.creds.api_key, team_handle=config.team_handle)
         )
 
-    def get_offers(
-        self, requirements: Optional[Requirements] = None
-    ) -> List[InstanceOfferWithAvailability]:
+    def get_all_offers_with_availability(self) -> List[InstanceOfferWithAvailability]:
         offers = get_catalog_offers(
             backend=BackendType.HOTAISLE,
             locations=self.config.regions or None,
-            requirements=requirements,
             catalog=self.catalog,
         )
-
         supported_offers = []
         for offer in offers:
             if offer.instance.name in INSTANCE_TYPE_SPECS:
@@ -78,7 +76,6 @@ class HotAisleCompute(
                 logger.warning(
                     f"Skipping unsupported Hot Aisle instance type: {offer.instance.name}"
                 )
-
         return supported_offers
 
     def get_payload_from_offer(self, instance_type) -> dict:

dstack/_internal/core/backends/kubernetes/compute.py

@@ -9,13 +9,14 @@ from kubernetes import client
 
 from dstack._internal.core.backends.base.compute import (
     Compute,
+    ComputeWithFilteredOffersCached,
     ComputeWithGatewaySupport,
     generate_unique_gateway_instance_name,
     generate_unique_instance_name_for_job,
     get_docker_commands,
     get_dstack_gateway_commands,
 )
-from dstack._internal.core.backends.base.offers import match_requirements
+from dstack._internal.core.backends.base.offers import filter_offers_by_requirements
 from dstack._internal.core.backends.kubernetes.models import (
     KubernetesConfig,
     KubernetesNetworkingConfig,
@@ -58,6 +59,7 @@ NVIDIA_GPU_NAMES = NVIDIA_GPU_NAME_TO_GPU_INFO.keys()
 
 
 class KubernetesCompute(
+    ComputeWithFilteredOffersCached,
     ComputeWithGatewaySupport,
     Compute,
 ):
@@ -70,8 +72,8 @@ class KubernetesCompute(
         self.networking_config = networking_config
         self.api = get_api_from_config_data(config.kubeconfig.data)
 
-    def get_offers(
-        self, requirements: Optional[Requirements] = None
+    def get_offers_by_requirements(
+        self, requirements: Requirements
     ) -> List[InstanceOfferWithAvailability]:
         nodes = self.api.list_node()
         instance_offers = []
@@ -99,7 +101,7 @@ class KubernetesCompute(
                 availability=InstanceAvailability.AVAILABLE,
                 instance_runtime=InstanceRuntime.RUNNER,
             )
-            instance_offers.extend(match_requirements([instance_offer], requirements))
+            instance_offers.extend(filter_offers_by_requirements([instance_offer], requirements))
         return instance_offers
 
     def run_job(

dstack/_internal/core/backends/lambdalabs/compute.py

@@ -7,6 +7,7 @@ from typing import Dict, List, Optional
 
 from dstack._internal.core.backends.base.compute import (
     Compute,
+    ComputeWithAllOffersCached,
     ComputeWithCreateInstanceSupport,
     generate_unique_instance_name,
     get_shim_commands,
@@ -22,12 +23,13 @@ from dstack._internal.core.models.instances import (
     InstanceOfferWithAvailability,
 )
 from dstack._internal.core.models.placement import PlacementGroup
-from dstack._internal.core.models.runs import JobProvisioningData, Requirements
+from dstack._internal.core.models.runs import JobProvisioningData
 
 MAX_INSTANCE_NAME_LEN = 60
 
 
 class LambdaCompute(
+    ComputeWithAllOffersCached,
     ComputeWithCreateInstanceSupport,
     Compute,
 ):
@@ -36,13 +38,10 @@ class LambdaCompute(
         self.config = config
         self.api_client = LambdaAPIClient(config.creds.api_key)
 
-    def get_offers(
-        self, requirements: Optional[Requirements] = None
-    ) -> List[InstanceOfferWithAvailability]:
+    def get_all_offers_with_availability(self) -> List[InstanceOfferWithAvailability]:
         offers = get_catalog_offers(
             backend=BackendType.LAMBDA,
             locations=self.config.regions or None,
-            requirements=requirements,
         )
         offers_with_availability = self._get_offers_with_availability(offers)
         return offers_with_availability

dstack/_internal/core/backends/local/compute.py

@@ -28,9 +28,7 @@ class LocalCompute(
     ComputeWithVolumeSupport,
     Compute,
 ):
-    def get_offers(
-        self, requirements: Optional[Requirements] = None
-    ) -> List[InstanceOfferWithAvailability]:
+    def get_offers(self, requirements: Requirements) -> List[InstanceOfferWithAvailability]:
         return [
             InstanceOfferWithAvailability(
                 backend=BackendType.LOCAL,

dstack/_internal/core/backends/nebius/compute.py

@@ -3,7 +3,7 @@ import random
 import shlex
 import time
 from functools import cached_property
-from typing import List, Optional
+from typing import Callable, List, Optional
 
 from nebius.aio.operation import Operation as SDKOperation
 from nebius.aio.service_error import RequestError, StatusCode
@@ -12,13 +12,14 @@ from nebius.sdk import SDK
 
 from dstack._internal.core.backends.base.backend import Compute
 from dstack._internal.core.backends.base.compute import (
+    ComputeWithAllOffersCached,
     ComputeWithCreateInstanceSupport,
     ComputeWithMultinodeSupport,
     ComputeWithPlacementGroupSupport,
     generate_unique_instance_name,
     get_user_data,
 )
-from dstack._internal.core.backends.base.offers import get_catalog_offers
+from dstack._internal.core.backends.base.offers import get_catalog_offers, get_offers_disk_modifier
 from dstack._internal.core.backends.nebius import resources
 from dstack._internal.core.backends.nebius.fabrics import get_suitable_infiniband_fabrics
 from dstack._internal.core.backends.nebius.models import NebiusConfig, NebiusServiceAccountCreds
@@ -76,6 +77,7 @@ SUPPORTED_PLATFORMS = [
 
 
 class NebiusCompute(
+    ComputeWithAllOffersCached,
     ComputeWithCreateInstanceSupport,
     ComputeWithMultinodeSupport,
     ComputeWithPlacementGroupSupport,
@@ -106,15 +108,11 @@ class NebiusCompute(
             ).metadata.id
         return self._subnet_id_cache[region]
 
-    def get_offers(
-        self, requirements: Optional[Requirements] = None
-    ) -> List[InstanceOfferWithAvailability]:
+    def get_all_offers_with_availability(self) -> List[InstanceOfferWithAvailability]:
         offers = get_catalog_offers(
             backend=BackendType.NEBIUS,
             locations=list(self._region_to_project_id),
-            requirements=requirements,
             extra_filter=_supported_instances,
-            configurable_disk_size=CONFIGURABLE_DISK_SIZE,
         )
         return [
             InstanceOfferWithAvailability(
@@ -124,6 +122,11 @@ class NebiusCompute(
             for offer in offers
         ]
 
+    def get_offers_modifier(
+        self, requirements: Requirements
+    ) -> Callable[[InstanceOfferWithAvailability], Optional[InstanceOfferWithAvailability]]:
+        return get_offers_disk_modifier(CONFIGURABLE_DISK_SIZE, requirements)
+
     def create_instance(
         self,
         instance_offer: InstanceOfferWithAvailability,

dstack/_internal/core/backends/oci/compute.py

@@ -1,17 +1,18 @@
 from concurrent.futures import ThreadPoolExecutor
 from functools import cached_property
-from typing import List, Optional
+from typing import Callable, List, Optional
 
 import oci
 
 from dstack._internal.core.backends.base.compute import (
     Compute,
+    ComputeWithAllOffersCached,
     ComputeWithCreateInstanceSupport,
     ComputeWithMultinodeSupport,
     generate_unique_instance_name,
     get_user_data,
 )
-from dstack._internal.core.backends.base.offers import get_catalog_offers
+from dstack._internal.core.backends.base.offers import get_catalog_offers, get_offers_disk_modifier
 from dstack._internal.core.backends.oci import resources
 from dstack._internal.core.backends.oci.models import OCIConfig
 from dstack._internal.core.backends.oci.region import make_region_clients_map
@@ -47,6 +48,7 @@ CONFIGURABLE_DISK_SIZE = Range[Memory](min=Memory.parse("50GB"), max=Memory.pars
 
 
 class OCICompute(
+    ComputeWithAllOffersCached,
     ComputeWithCreateInstanceSupport,
     ComputeWithMultinodeSupport,
     Compute,
@@ -60,14 +62,10 @@ class OCICompute(
     def shapes_quota(self) -> resources.ShapesQuota:
         return resources.ShapesQuota.load(self.regions, self.config.compartment_id)
 
-    def get_offers(
-        self, requirements: Optional[Requirements] = None
-    ) -> List[InstanceOfferWithAvailability]:
+    def get_all_offers_with_availability(self) -> List[InstanceOfferWithAvailability]:
         offers = get_catalog_offers(
             backend=BackendType.OCI,
             locations=self.config.regions,
-            requirements=requirements,
-            configurable_disk_size=CONFIGURABLE_DISK_SIZE,
             extra_filter=_supported_instances,
         )
 
@@ -96,6 +94,11 @@ class OCICompute(
 
         return offers_with_availability
 
+    def get_offers_modifier(
+        self, requirements: Requirements
+    ) -> Callable[[InstanceOfferWithAvailability], Optional[InstanceOfferWithAvailability]]:
+        return get_offers_disk_modifier(CONFIGURABLE_DISK_SIZE, requirements)
+
     def terminate_instance(
         self, instance_id: str, region: str, backend_data: Optional[str] = None
     ) -> None:
@@ -115,7 +118,11 @@ class OCICompute(
         availability_domain = instance_offer.availability_zones[0]
 
         listing, package = resources.get_marketplace_listing_and_package(
-            cuda=len(instance_offer.instance.resources.gpus) > 0,
+            gpu_name=(
+                instance_offer.instance.resources.gpus[0].name
+                if len(instance_offer.instance.resources.gpus) > 0
+                else None
+            ),
             client=region.marketplace_client,
         )
         resources.accept_marketplace_listing_agreements(

dstack/_internal/core/backends/oci/resources.py

@@ -23,7 +23,9 @@ import oci
 from oci.object_storage.models import CreatePreauthenticatedRequestDetails
 
 from dstack import version
+from dstack._internal.core.backends.base.compute import requires_nvidia_proprietary_kernel_modules
 from dstack._internal.core.backends.oci.region import OCIRegionClient
+from dstack._internal.core.consts import DSTACK_OS_IMAGE_WITH_PROPRIETARY_NVIDIA_KERNEL_MODULES
 from dstack._internal.core.errors import BackendError
 from dstack._internal.core.models.instances import InstanceOffer
 from dstack._internal.utils.common import batched
@@ -352,11 +354,14 @@ def terminate_instance_if_exists(client: oci.core.ComputeClient, instance_id: st
 
 
 def get_marketplace_listing_and_package(
-    cuda: bool, client: oci.marketplace.MarketplaceClient
+    gpu_name: Optional[str], client: oci.marketplace.MarketplaceClient
 ) -> Tuple[oci.marketplace.models.Listing, oci.marketplace.models.ImageListingPackage]:
     listing_name = f"dstack-{version.base_image}"
-    if cuda:
-        listing_name = f"dstack-cuda-{version.base_image}"
+    if gpu_name is not None:
+        if not requires_nvidia_proprietary_kernel_modules(gpu_name):
+            listing_name = f"dstack-cuda-{version.base_image}"
+        else:
+            listing_name = f"dstack-cuda-{DSTACK_OS_IMAGE_WITH_PROPRIETARY_NVIDIA_KERNEL_MODULES}"
 
     listing_summaries = list_marketplace_listings(listing_name, client)
     if len(listing_summaries) != 1:

dstack/_internal/core/backends/runpod/compute.py

@@ -1,17 +1,18 @@
 import json
 import uuid
 from datetime import timedelta
-from typing import List, Optional
+from typing import Callable, List, Optional
 
 from dstack._internal.core.backends.base.backend import Compute
 from dstack._internal.core.backends.base.compute import (
+    ComputeWithAllOffersCached,
     ComputeWithVolumeSupport,
     generate_unique_instance_name,
     generate_unique_volume_name,
     get_docker_commands,
     get_job_instance_name,
 )
-from dstack._internal.core.backends.base.offers import get_catalog_offers
+from dstack._internal.core.backends.base.offers import get_catalog_offers, get_offers_disk_modifier
 from dstack._internal.core.backends.runpod.api_client import RunpodApiClient
 from dstack._internal.core.backends.runpod.models import RunpodConfig
 from dstack._internal.core.consts import DSTACK_RUNNER_SSH_PORT
@@ -27,6 +28,7 @@ from dstack._internal.core.models.instances import (
     InstanceOfferWithAvailability,
     SSHKey,
 )
+from dstack._internal.core.models.resources import Memory, Range
 from dstack._internal.core.models.runs import Job, JobProvisioningData, Requirements, Run
 from dstack._internal.core.models.volumes import Volume, VolumeProvisioningData
 from dstack._internal.utils.common import get_current_datetime
@@ -39,8 +41,12 @@ MAX_RESOURCE_NAME_LEN = 60
 
 CONTAINER_REGISTRY_AUTH_CLEANUP_INTERVAL = 60 * 60 * 24  # 24 hour
 
+# RunPod does not seem to have any limits on the disk size.
+CONFIGURABLE_DISK_SIZE = Range[Memory](min=Memory.parse("1GB"), max=None)
+
 
 class RunpodCompute(
+    ComputeWithAllOffersCached,
     ComputeWithVolumeSupport,
     Compute,
 ):
@@ -51,13 +57,11 @@ class RunpodCompute(
         self.config = config
         self.api_client = RunpodApiClient(config.creds.api_key)
 
-    def get_offers(
-        self, requirements: Optional[Requirements] = None
-    ) -> List[InstanceOfferWithAvailability]:
+    def get_all_offers_with_availability(self) -> List[InstanceOfferWithAvailability]:
         offers = get_catalog_offers(
             backend=BackendType.RUNPOD,
             locations=self.config.regions or None,
-            requirements=requirements,
+            requirements=None,
             extra_filter=lambda o: _is_secure_cloud(o.region) or self.config.allow_community_cloud,
         )
         offers = [
@@ -68,6 +72,11 @@ class RunpodCompute(
         ]
         return offers
 
+    def get_offers_modifier(
+        self, requirements: Requirements
+    ) -> Callable[[InstanceOfferWithAvailability], Optional[InstanceOfferWithAvailability]]:
+        return get_offers_disk_modifier(CONFIGURABLE_DISK_SIZE, requirements)
+
     def run_job(
         self,
         run: Run,

dstack/_internal/core/backends/template/compute.py.jinja

@@ -2,6 +2,7 @@ from typing import List, Optional
 
 from dstack._internal.core.backends.base.backend import Compute
 from dstack._internal.core.backends.base.compute import (
+    ComputeWithAllOffersCached,
     ComputeWithCreateInstanceSupport,
     ComputeWithGatewaySupport,
     ComputeWithMultinodeSupport,
@@ -28,6 +29,7 @@ logger = get_logger(__name__)
 
 class {{ backend_name }}Compute(
     # TODO: Choose ComputeWith* classes to extend and implement
+    # ComputeWithAllOffersCached,
     # ComputeWithCreateInstanceSupport,
     # ComputeWithMultinodeSupport,
     # ComputeWithReservationSupport,
@@ -42,7 +44,7 @@ class {{ backend_name }}Compute(
         self.config = config
 
     def get_offers(
-        self, requirements: Optional[Requirements] = None
+        self, requirements: Requirements
     ) -> List[InstanceOfferWithAvailability]:
         # If the provider is added to gpuhunt, you'd typically get offers
         # using `get_catalog_offers()` and extend them with availability info.