Drop AWS P3 support and use DLAMI for all AWS GPU instances

src/dstack/_internal/core/backends/aws/compute.py

@@ -1244,7 +1244,6 @@ def _supported_instances(offer: InstanceOffer) -> bool:
         "p5e.",
         "p4d.",
         "p4de.",
-        "p3.",
         "g7e.",
         "g6.",
         "g6e.",

src/dstack/_internal/core/backends/aws/resources.py

@@ -6,8 +6,6 @@
 
 import dstack.version as version
 from dstack._internal.core.backends.aws.models import AWSOSImageConfig
-from dstack._internal.core.backends.base.compute import requires_nvidia_proprietary_kernel_modules
-from dstack._internal.core.consts import DSTACK_OS_IMAGE_WITH_PROPRIETARY_NVIDIA_KERNEL_MODULES
 from dstack._internal.core.errors import BackendError, ComputeError, ComputeResourceNotFoundError
 from dstack._internal.utils.logging import get_logger
 
@@ -31,17 +29,15 @@ def get_image_id_and_username(
         image_name = image.name
         image_owner = image.owner
         username = image.user
-    elif _supported_by_dlami(instance_type):
+    elif gpu_name is not None:
+        # AWS Deep Learning AMIs (DLAMI) support all GPU instance types currently supported by dstack.
+        # dstack's cuda AMI is still built but not used.
+        # It may be used again in case some instance types are not supported by DLAMI.
         image_name = "Deep Learning Base OSS Nvidia Driver GPU AMI (Ubuntu 22.04) *"
         image_owner = DLAMI_OWNER_ACCOUNT_ID
         username = "ubuntu"
     else:
-        if gpu_name is None:
-            image_name = f"dstack-{version.base_image}"
-        elif 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}"
+        image_name = f"dstack-{version.base_image}"
         image_owner = DSTACK_ACCOUNT_ID
         username = "ubuntu"
     response = ec2_client.describe_images(
@@ -636,25 +632,6 @@ def _is_private_subnet_with_internet_egress(
     return False
 
 
-def _supported_by_dlami(instance_type: str) -> bool:
-    # Currently only p3. instances are not supported by DLAMI among GPU instances.
-    return any(
-        instance_type.startswith(family)
-        for family in [
-            "g4dn.",
-            "g5.",
-            "g6.",
-            "gr6.",
-            "g6e.",
-            "p4d.",
-            "p4de.",
-            "p5.",
-            "p5e.",
-            "p6-b200.",
-        ]
-    )
-
-
 def get_reservation(
     ec2_client: botocore.client.BaseClient,
     reservation_id: str,

src/dstack/_internal/core/models/fleets.py

@@ -261,7 +261,7 @@ class BackendFleetConfiguraionProps(CoreModel):
     instance_types: Annotated[
         Optional[List[str]],
         Field(
-            description="The cloud-specific instance types to consider for provisioning (e.g., `[p3.8xlarge, n1-standard-4]`)"
+            description="The cloud-specific instance types to consider for provisioning (e.g., `[g6e.24xlarge, n1-standard-4]`)"
         ),
     ] = None
     spot_policy: Annotated[

src/dstack/_internal/core/models/profiles.py

@@ -257,7 +257,7 @@ class ProfileParams(CoreModel):
     instance_types: Annotated[
         Optional[List[str]],
         Field(
-            description="The cloud-specific instance types to consider for provisioning (e.g., `[p3.8xlarge, n1-standard-4]`)"
+            description="The cloud-specific instance types to consider for provisioning (e.g., `[g6e.24xlarge, n1-standard-4]`)"
         ),
     ] = None
     reservation: Annotated[

src/tests/_internal/core/backends/aws/test_resources.py

@@ -150,14 +150,23 @@ def test_raises_resource_not_found_if_none_available(
         assert "image 'dstack-0.0' not found" in caplog.text
 
     @pytest.mark.parametrize(
-        ["cuda", "expected"],
+        ["cuda", "expected_name", "expected_owner"],
         [
-            [False, "dstack-0.0"],
-            [True, "dstack-cuda-0.0"],
+            [False, "dstack-0.0", "142421590066"],
+            [
+                True,
+                "Deep Learning Base OSS Nvidia Driver GPU AMI (Ubuntu 22.04) *",
+                "898082745236",
+            ],
         ],
     )
-    def test_uses_dstack_image_name_and_account_id_if_image_config_not_provided(
-        self, monkeypatch: pytest.MonkeyPatch, ec2_client_mock: Mock, cuda: bool, expected: str
+    def test_uses_default_image_name_and_account_id_if_image_config_not_provided(
+        self,
+        monkeypatch: pytest.MonkeyPatch,
+        ec2_client_mock: Mock,
+        cuda: bool,
+        expected_name: str,
+        expected_owner: str,
     ):
         monkeypatch.setattr("dstack.version.base_image", "0.0")
         _, username = get_image_id_and_username(
@@ -167,7 +176,7 @@ def test_uses_dstack_image_name_and_account_id_if_image_config_not_provided(
         )
         assert username == "ubuntu"
         ec2_client_mock.describe_images.assert_called_once_with(
-            Filters=[{"Name": "name", "Values": [expected]}], Owners=["142421590066"]
+            Filters=[{"Name": "name", "Values": [expected_name]}], Owners=[expected_owner]
         )
 
     @pytest.mark.parametrize(

src/tests/_internal/server/routers/test_fleets.py

@@ -1534,7 +1534,7 @@ async def test_errors_if_ssh_key_is_bad(
         [
             pytest.param("backends", [BackendType.AWS], id="backends"),
             pytest.param("regions", ["eu-west-1"], id="regions"),
-            pytest.param("instance_types", ["p3.8xlarge"], id="instance_types"),
+            pytest.param("instance_types", ["g6e.24xlarge"], id="instance_types"),
             pytest.param("idle_duration", 60, id="idle_duration"),
             pytest.param("tags", {}, id="tags"),  # falsy value
         ],