Support GCP backend

.github/workflows/manual_gcp_image_stgn.yml

@@ -0,0 +1,46 @@
+name: Manual GCP image build for STGN
+
+on:
+  workflow_dispatch:
+
+jobs:
+  gcp-images-build:
+    defaults:
+      run:
+        working-directory: runner
+    strategy:
+      matrix:
+        variant: [ gcp-image, gcp-image-cuda ]
+        include:
+          - variant: gcp-image
+            image_prefix: stgn-dstack-nocuda
+          - variant: gcp-image-cuda
+            image_prefix: stgn-dstack-cuda
+    runs-on: ubuntu-latest
+    permissions:
+      contents: 'read'
+      id-token: 'write'
+    steps:
+      - uses: actions/checkout@v3
+      - name: 'Authenticate to Google Cloud'
+        uses: 'google-github-actions/auth@v1'
+        with:
+          workload_identity_provider: 'projects/531508670106/locations/global/workloadIdentityPools/github-identity-pool/providers/github-id-provider'
+          service_account: 'github-actions@dstack.iam.gserviceaccount.com'
+          create_credentials_file: true
+      - name: 'Set up Cloud SDK'
+        uses: 'google-github-actions/setup-gcloud@v1'
+      - name: Download packer
+        run: |
+          wget https://releases.hashicorp.com/packer/1.8.0/packer_1.8.0_linux_amd64.zip
+          unzip packer_1.8.0_linux_amd64.zip
+          chmod +x packer
+          cp -R ami/packer/* .
+      - name: Run packer
+        run: |
+          VERSION=${{ github.run_number }}
+          ./packer build -var build_id=$VERSION ${{ matrix.variant }}.json
+      - name: Publish images
+        run: |
+          VERSION=${{ github.run_number }}
+          gcloud compute images add-iam-policy-binding ${{ matrix.image_prefix }}-$VERSION --member='allAuthenticatedUsers' --role='roles/compute.imageUser'

.github/workflows/release.yml

@@ -195,6 +195,49 @@ jobs:
         run: |
           VERSION=${GITHUB_REF#refs/tags/}
           ./packer build -var build_id=$VERSION -var build_prefix='' -var-file=prod-var.json ${{ matrix.variants }}.json
+
+  gcp-images-build:
+    defaults:
+      run:
+        working-directory: runner
+    strategy:
+      matrix:
+        variant: [ gcp-image, gcp-image-cuda ]
+        include:
+          - variant: gcp-image
+            image_prefix: dstack-nocuda
+          - variant: gcp-image-cuda
+            image_prefix: dstack-cuda
+    runs-on: ubuntu-latest
+    permissions:
+      contents: 'read'
+      id-token: 'write'
+    steps:
+      - uses: actions/checkout@v3
+      - name: 'Authenticate to Google Cloud'
+        uses: 'google-github-actions/auth@v1'
+        with:
+          workload_identity_provider: 'projects/531508670106/locations/global/workloadIdentityPools/github-identity-pool/providers/github-id-provider'
+          service_account: 'github-actions@dstack.iam.gserviceaccount.com'
+          create_credentials_file: true
+      - name: 'Set up Cloud SDK'
+        uses: 'google-github-actions/setup-gcloud@v1'
+      - name: Download packer
+        run: |
+          wget https://releases.hashicorp.com/packer/1.8.0/packer_1.8.0_linux_amd64.zip
+          unzip packer_1.8.0_linux_amd64.zip
+          chmod +x packer
+          cp -R ami/packer/* .
+      - name: Run packer
+        run: |
+          VERSION=${GITHUB_REF#refs/tags/}
+          VERSION=${VERSION//./-}
+          ./packer build -var build_id=$VERSION -var build_prefix='' -var dstack_stage=PROD ${{ matrix.variant }}.json
+      - name: Publish image
+        run: |
+          VERSION=${GITHUB_REF#refs/tags/}
+          VERSION=${VERSION//./-}
+          gcloud compute images add-iam-policy-binding ${{ matrix.image_prefix }}-$VERSION --member='allAuthenticatedUsers' --role='roles/compute.imageUser'
 
   update-get-dstack-tag:
     runs-on: ubuntu-latest

cli/dstack/api/backend.py

@@ -2,14 +2,12 @@
 
 from dstack import version
 from dstack.backend.aws import AwsBackend
-from dstack.backend.base import Backend, RemoteBackend
+from dstack.backend.base import Backend, BackendType, RemoteBackend
+from dstack.backend.gcp import GCPBackend
 from dstack.backend.hub import HubBackend
 from dstack.backend.local import LocalBackend
 
-DEFAULT_REMOTE = "aws"
-DEFAULT = "local"
-
-backends_classes = [AwsBackend, LocalBackend]
+backends_classes = [AwsBackend, GCPBackend, LocalBackend]
 if not version.__is_release__:
     backends_classes.append(HubBackend)
 
@@ -28,6 +26,10 @@ def list_backends(all_backend: bool = False) -> List[Backend]:
     return l
 
 
+def list_remote_backends() -> List[RemoteBackend]:
+    return [b for b in list_backends() if b.type is BackendType.REMOTE]
+
+
 def dict_backends(all_backend: bool = False) -> Dict[str, Backend]:
     return {backend.name: backend for backend in list_backends(all_backend=all_backend)}
 
@@ -37,7 +39,10 @@ def get_backend_by_name(name: str) -> Optional[Backend]:
 
 
 def get_current_remote_backend() -> Optional[RemoteBackend]:
-    return get_backend_by_name(DEFAULT_REMOTE)
+    remote_backends = list_remote_backends()
+    if len(remote_backends) == 0:
+        return None
+    return remote_backends[0]
 
 
 def get_local_backend() -> LocalBackend:

cli/dstack/api/config.py

@@ -2,10 +2,11 @@
 
 from dstack import version
 from dstack.backend.aws.config import AWSConfig
+from dstack.backend.gcp.config import GCPConfigurator
 from dstack.backend.hub.config import HUBConfig
 from dstack.core.config import BackendConfig
 
-configs_classes = [AWSConfig]
+configs_classes = [AWSConfig, GCPConfigurator]
 if not version.__is_release__:
     configs_classes.append(HUBConfig)
 
@@ -19,7 +20,6 @@ def dict_config() -> Dict[str, BackendConfig]:
     configs = [cls() for cls in configs_classes]
     names = {}
     for config in configs:
-        if config.configured:
-            names[config.name] = config
+        names[config.name] = config
 
     return names

cli/dstack/backend/aws/compute.py

@@ -32,13 +32,13 @@ def get_request_head(self, job: Job, request_id: Optional[str]) -> RequestHead:
         )
 
     def get_instance_type(self, job: Job) -> Optional[InstanceType]:
-        return runners._get_instance_type(
+        return runners.get_instance_type(
             ec2_client=self.ec2_client,
             requirements=job.requirements,
         )
 
     def run_instance(self, job: Job, instance_type: InstanceType) -> str:
-        return runners._run_instance_retry(
+        return runners.run_instance_retry(
             ec2_client=self.ec2_client,
             iam_client=self.iam_client,
             bucket_name=self.bucket_name,
@@ -52,13 +52,13 @@ def run_instance(self, job: Job, instance_type: InstanceType) -> str:
         )
 
     def terminate_instance(self, request_id: str):
-        runners._terminate_instance(
+        runners.terminate_instance(
             ec2_client=self.ec2_client,
             request_id=request_id,
         )
 
     def cancel_spot_request(self, request_id: str):
-        runners._cancel_spot_request(
+        runners.cancel_spot_request(
             ec2_client=self.ec2_client,
             request_id=request_id,
         )

cli/dstack/backend/aws/config.py

@@ -31,8 +31,6 @@
 class AWSConfig(BackendConfig):
     NAME = "aws"
 
-    _configured = True
-
     bucket_name = None
     region_name = None
     profile_name = None

cli/dstack/backend/aws/runners.py

@@ -1,21 +1,28 @@
 import json
 import time
-from functools import cmp_to_key, reduce
+from functools import reduce
 from typing import List, Optional, Tuple
 
-import yaml
 from botocore.client import BaseClient
 
 from dstack import version
+from dstack.backend.base.compute import choose_instance_type
 from dstack.core.instance import InstanceType
 from dstack.core.job import Job, JobStatus, Requirements
 from dstack.core.repo import RepoAddress
 from dstack.core.request import RequestHead, RequestStatus
-from dstack.core.runners import Gpu, Resources, Runner
+from dstack.core.runners import Gpu, Resources
 
 CREATE_INSTANCE_RETRY_RATE_SECS = 3
 
 
+def get_instance_type(
+    ec2_client: BaseClient, requirements: Optional[Requirements]
+) -> Optional[InstanceType]:
+    instance_types = _get_instance_types(ec2_client)
+    return choose_instance_type(instance_types, requirements)
+
+
 def _get_instance_types(ec2_client: BaseClient) -> List[InstanceType]:
     response = None
     instance_types = []
@@ -54,86 +61,7 @@ def _get_instance_types(ec2_client: BaseClient) -> List[InstanceType]:
                     ),
                 )
             )
-
-    def compare(i1, i2):
-        r1_gpu_total_memory_mib = sum(map(lambda g: g.memory_mib, i1.resources.gpus or []))
-        r2_gpu_total_memory_mib = sum(map(lambda g: g.memory_mib, i2.resources.gpus or []))
-        if r1_gpu_total_memory_mib < r2_gpu_total_memory_mib:
-            return -1
-        elif r1_gpu_total_memory_mib > r2_gpu_total_memory_mib:
-            return 1
-        if i1.resources.cpus < i2.resources.cpus:
-            return -1
-        elif i1.resources.cpus > i2.resources.cpus:
-            return 1
-        if i1.resources.memory_mib < i2.resources.memory_mib:
-            return -1
-        elif i1.resources.memory_mib > i2.resources.memory_mib:
-            return 1
-        return 0
-
-    return sorted(instance_types, key=cmp_to_key(compare))
-
-
-def _matches(resources: Resources, requirements: Optional[Requirements]) -> bool:
-    if not requirements:
-        return True
-    if requirements.cpus and requirements.cpus > resources.cpus:
-        return False
-    if requirements.memory_mib and requirements.memory_mib > resources.memory_mib:
-        return False
-    if requirements.gpus:
-        gpu_count = requirements.gpus.count or 1
-        if gpu_count > len(resources.gpus or []):
-            return False
-        if requirements.gpus.name and gpu_count > len(
-            list(filter(lambda gpu: gpu.name == requirements.gpus.name, resources.gpus or []))
-        ):
-            return False
-        if requirements.gpus.memory_mib and gpu_count > len(
-            list(
-                filter(
-                    lambda gpu: gpu.memory_mib >= requirements.gpus.memory_mib,
-                    resources.gpus or [],
-                )
-            )
-        ):
-            return False
-        if requirements.interruptible and not resources.interruptible:
-            return False
-    return True
-
-
-def _get_instance_type(
-    ec2_client: BaseClient, requirements: Optional[Requirements]
-) -> Optional[InstanceType]:
-    instance_types = _get_instance_types(ec2_client)
-
-    instance_type = next(
-        (
-            instance_type
-            for instance_type in instance_types
-            if _matches(instance_type.resources, requirements)
-        ),
-        None,
-    )
-    interruptible = False
-    if requirements and requirements.interruptible:
-        interruptible = True
-    return (
-        InstanceType(
-            instance_name=instance_type.instance_name,
-            resources=Resources(
-                cpus=instance_type.resources.cpus,
-                memory_mib=instance_type.resources.memory_mib,
-                gpus=instance_type.resources.gpus,
-                interruptible=interruptible,
-                local=False,
-            ),
-        )
-        if instance_type
-        else None
-    )
+    return instance_types
 
 
 def _get_security_group_id(ec2_client: BaseClient, bucket_name: str, subnet_id: Optional[str]):
@@ -491,7 +419,7 @@ def _run_instance(
     return request_id
 
 
-def _run_instance_retry(
+def run_instance_retry(
     ec2_client: BaseClient,
     iam_client: BaseClient,
     bucket_name: str,
@@ -525,7 +453,7 @@ def _run_instance_retry(
         ):
             if attempts > 0:
                 time.sleep(CREATE_INSTANCE_RETRY_RATE_SECS)
-                return _run_instance_retry(
+                return run_instance_retry(
                     ec2_client,
                     iam_client,
                     bucket_name,
@@ -544,7 +472,7 @@ def _run_instance_retry(
             raise e
 
 
-def _cancel_spot_request(ec2_client: BaseClient, request_id: str):
+def cancel_spot_request(ec2_client: BaseClient, request_id: str):
     ec2_client.cancel_spot_instance_requests(SpotInstanceRequestIds=[request_id])
     response = ec2_client.describe_instances(
         Filters=[
@@ -557,7 +485,7 @@ def _cancel_spot_request(ec2_client: BaseClient, request_id: str):
         )
 
 
-def _terminate_instance(ec2_client: BaseClient, request_id: str):
+def terminate_instance(ec2_client: BaseClient, request_id: str):
     try:
         ec2_client.terminate_instances(InstanceIds=[request_id])
     except Exception as e: