Set explicit GPU defaults in ResourcesSpec and improve default GPU vendor selection (#3573)

* Set explicit GPU default (`0..`) in `ResourcesSpec` and minor improvements in resource pretty-printing

* Change how GPU vendor default is set to make it more explicit

- Default to NVIDIA only if user has no image
- Keep backward compatibility with old/new server/CLI
- Make `dstack offer` consistent with `dstack apply`

Author: peterschmidt85

src/dstack/_internal/cli/commands/offer.py

@@ -74,7 +74,10 @@ def _register(self):
 
     def _command(self, args: argparse.Namespace):
         super()._command(args)
-        conf = TaskConfiguration(commands=[":"])
+        # Set image and user so that the server (a) does not default gpu.vendor
+        # to nvidia — `dstack offer` should show all vendors, and (b) does not
+        # attempt to pull image config from the Docker registry.
+        conf = TaskConfiguration(commands=[":"], image="scratch", user="root")
 
         configurator = OfferConfigurator(api_client=self.api)
         configurator.apply_args(conf, args)

src/dstack/_internal/cli/services/configurators/run.py

@@ -383,7 +383,13 @@ def interpolate_env(self, conf: RunConfigurationT):
 
     def validate_gpu_vendor_and_image(self, conf: RunConfigurationT) -> None:
         """
-        Infers and sets `resources.gpu.vendor` if not set, requires `image` if the vendor is AMD.
+        Infers GPU vendor if not set. Defaults to Nvidia when using the default
+        CUDA image. Requires explicit `image` if the vendor is AMD or Tenstorrent.
+
+        NOTE: We don't set the inferred vendor on gpu_spec for compatibility with
+        older servers. Servers set the vendor using the same logic in
+        set_resources_defaults(). The inferred vendor is used here only for
+        validation and display (see _infer_gpu_vendor).
         """
         gpu_spec = conf.resources.gpu
         if gpu_spec is None:
@@ -425,12 +431,18 @@ def validate_gpu_vendor_and_image(self, conf: RunConfigurationT) -> None:
                 # CUDA image, not a big deal.
                 has_amd_gpu = gpuhunt.AcceleratorVendor.AMD in vendors
                 has_tt_gpu = gpuhunt.AcceleratorVendor.TENSTORRENT in vendors
+                # Set vendor inferred from name on the spec (server needs it for filtering).
+                gpu_spec.vendor = vendor
             else:
-                # If neither gpu.vendor nor gpu.name is set, assume Nvidia.
-                vendor = gpuhunt.AcceleratorVendor.NVIDIA
+                # No vendor or name specified. Default to Nvidia if using the default
+                # CUDA image, since it's only compatible with Nvidia GPUs.
+                # We don't set the inferred vendor on the spec — the server does the
+                # same inference in set_resources_defaults() for compatibility with
+                # older servers that don't handle vendor + count.min=0 correctly.
+                if conf.image is None and conf.docker is not True:
+                    vendor = gpuhunt.AcceleratorVendor.NVIDIA
                 has_amd_gpu = False
                 has_tt_gpu = False
-            gpu_spec.vendor = vendor
         else:
             has_amd_gpu = vendor == gpuhunt.AcceleratorVendor.AMD
             has_tt_gpu = vendor == gpuhunt.AcceleratorVendor.TENSTORRENT

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

@@ -319,6 +319,9 @@ def _vendor_from_string(cls, v: str) -> gpuhunt.AcceleratorVendor:
         return gpuhunt.AcceleratorVendor.cast(v)
 
 
+DEFAULT_GPU_SPEC = GPUSpec(count=Range[int](min=0, max=None))
+
+
 class DiskSpecConfig(CoreConfig):
     @staticmethod
     def schema_extra(schema: Dict[str, Any]):
@@ -387,7 +390,8 @@ class ResourcesSpec(generate_dual_core_model(ResourcesSpecConfig)):
             "you may need to configure this"
         ),
     ] = None
-    gpu: Annotated[Optional[GPUSpec], Field(description="The GPU requirements")] = None
+    # Optional for backward compatibility
+    gpu: Annotated[Optional[GPUSpec], Field(description="The GPU requirements")] = DEFAULT_GPU_SPEC
     disk: Annotated[Optional[DiskSpec], Field(description="The disk resources")] = DEFAULT_DISK
 
     def pretty_format(self) -> str:
@@ -397,6 +401,7 @@ def pretty_format(self) -> str:
         if self.gpu:
             gpu = self.gpu
             resources.update(
+                gpu_vendor=gpu.vendor,
                 gpu_name=",".join(gpu.name) if gpu.name else None,
                 gpu_count=gpu.count,
                 gpu_memory=gpu.memory,

src/dstack/_internal/server/services/resources.py

@@ -1,3 +1,5 @@
+from typing import Optional
+
 import gpuhunt
 from pydantic import parse_obj_as
 
@@ -19,3 +21,24 @@ def set_resources_defaults(resources: ResourcesSpec) -> None:
         else:
             cpu.arch = gpuhunt.CPUArchitecture.X86
         resources.cpu = cpu
+
+
+def set_gpu_vendor_default(
+    resources: ResourcesSpec,
+    image: Optional[str],
+    docker: Optional[bool],
+) -> None:
+    """Default GPU vendor to Nvidia when using the default CUDA image,
+    since it's only compatible with Nvidia GPUs.
+    Mirrors the client-side logic in validate_gpu_vendor_and_image().
+    Should only be called for runs (not fleets) since fleets don't have image context."""
+    gpu = resources.gpu
+    if (
+        gpu is not None
+        and gpu.vendor is None
+        and gpu.name is None
+        and gpu.count.max != 0
+        and image is None
+        and docker is not True
+    ):
+        gpu.vendor = gpuhunt.AcceleratorVendor.NVIDIA

src/dstack/_internal/server/services/runs/__init__.py

@@ -65,7 +65,10 @@
 from dstack._internal.server.services.plugins import apply_plugin_policies
 from dstack._internal.server.services.probes import is_probe_ready
 from dstack._internal.server.services.projects import list_user_project_models
-from dstack._internal.server.services.resources import set_resources_defaults
+from dstack._internal.server.services.resources import (
+    set_gpu_vendor_default,
+    set_resources_defaults,
+)
 from dstack._internal.server.services.runs.plan import get_job_plans
 from dstack._internal.server.services.runs.spec import (
     can_update_run_spec,
@@ -343,8 +346,8 @@ async def get_plan(
         )
         if current_resource is not None:
             # For backward compatibility (current_resource may has been submitted before
-            # some fields, e.g., CPUSpec.arch, were added)
-            set_resources_defaults(current_resource.run_spec.configuration.resources)
+            # some fields, e.g., CPUSpec.arch, gpu.vendor were added)
+            _set_run_resources_defaults(current_resource.run_spec)
             if not current_resource.status.is_finished() and can_update_run_spec(
                 current_resource.run_spec, effective_run_spec
             ):
@@ -354,7 +357,7 @@ async def get_plan(
         session=session,
         project=project,
         profile=profile,
-        run_spec=run_spec,
+        run_spec=effective_run_spec,
         max_offers=max_offers,
     )
     run_plan = RunPlan(
@@ -410,8 +413,8 @@ async def apply_plan(
     current_resource = run_model_to_run(current_resource_model, return_in_api=True)
 
     # For backward compatibility (current_resource may has been submitted before
-    # some fields, e.g., CPUSpec.arch, were added)
-    set_resources_defaults(current_resource.run_spec.configuration.resources)
+    # some fields, e.g., CPUSpec.arch, gpu.vendor were added)
+    _set_run_resources_defaults(current_resource.run_spec)
     try:
         spec_diff = check_can_update_run_spec(current_resource.run_spec, run_spec)
     except ServerClientError:
@@ -421,7 +424,7 @@ async def apply_plan(
         raise
     if not force:
         if plan.current_resource is not None:
-            set_resources_defaults(plan.current_resource.run_spec.configuration.resources)
+            _set_run_resources_defaults(plan.current_resource.run_spec)
         if (
             plan.current_resource is None
             or plan.current_resource.id != current_resource.id
@@ -782,6 +785,16 @@ def run_model_to_run(
     return run
 
 
+def _set_run_resources_defaults(run_spec: RunSpec) -> None:
+    """Apply resource defaults to a run spec, including GPU vendor inference."""
+    set_resources_defaults(run_spec.configuration.resources)
+    set_gpu_vendor_default(
+        run_spec.configuration.resources,
+        image=run_spec.configuration.image,
+        docker=getattr(run_spec.configuration, "docker", None),
+    )
+
+
 def _get_run_jobs_with_submissions(
     run_model: RunModel,
     job_submissions_limit: Optional[int],

src/dstack/_internal/server/services/runs/spec.py

@@ -8,7 +8,10 @@
 from dstack._internal.server import settings
 from dstack._internal.server.models import UserModel
 from dstack._internal.server.services.docker import is_valid_docker_volume_target
-from dstack._internal.server.services.resources import set_resources_defaults
+from dstack._internal.server.services.resources import (
+    set_gpu_vendor_default,
+    set_resources_defaults,
+)
 from dstack._internal.utils.logging import get_logger
 
 logger = get_logger(__name__)
@@ -108,6 +111,11 @@ def validate_run_spec_and_set_defaults(
     if run_spec.configuration.priority is None:
         run_spec.configuration.priority = RUN_PRIORITY_DEFAULT
     set_resources_defaults(run_spec.configuration.resources)
+    set_gpu_vendor_default(
+        run_spec.configuration.resources,
+        image=run_spec.configuration.image,
+        docker=getattr(run_spec.configuration, "docker", None),
+    )
     if run_spec.ssh_key_pub is None:
         if user.ssh_public_key:
             run_spec.ssh_key_pub = user.ssh_public_key

src/dstack/_internal/utils/common.py

@@ -91,28 +91,14 @@ def pretty_resources(
     cpus: Optional[Any] = None,
     memory: Optional[Any] = None,
     gpu_count: Optional[Any] = None,
+    gpu_vendor: Optional[Any] = None,
     gpu_name: Optional[Any] = None,
     gpu_memory: Optional[Any] = None,
     total_gpu_memory: Optional[Any] = None,
     compute_capability: Optional[Any] = None,
     disk_size: Optional[Any] = None,
 ) -> str:
-    """
-    >>> pretty_resources(cpus=4, memory="16GB")
-    '4xCPU, 16GB'
-    >>> pretty_resources(cpus=4, memory="16GB", gpu_count=1)
-    '4xCPU, 16GB, 1xGPU'
-    >>> pretty_resources(cpus=4, memory="16GB", gpu_count=1, gpu_name='A100')
-    '4xCPU, 16GB, 1xA100'
-    >>> pretty_resources(cpus=4, memory="16GB", gpu_count=1, gpu_name='A100', gpu_memory="40GB")
-    '4xCPU, 16GB, 1xA100 (40GB)'
-    >>> pretty_resources(cpus=4, memory="16GB", gpu_count=1, total_gpu_memory="80GB")
-    '4xCPU, 16GB, 1xGPU (total 80GB)'
-    >>> pretty_resources(cpus=4, memory="16GB", gpu_count=2, gpu_name='A100', gpu_memory="40GB", total_gpu_memory="80GB")
-    '4xCPU, 16GB, 2xA100 (40GB, total 80GB)'
-    >>> pretty_resources(gpu_count=1, compute_capability="8.0")
-    '1xGPU (8.0)'
-    """
+    """Format resource requirements as a human-readable string."""
     parts = []
     if cpus is not None:
         cpu_arch_lower: Optional[str] = None
@@ -131,7 +117,6 @@ def pretty_resources(
         parts.append(f"disk={disk_size}")
     if gpu_count:
         gpu_parts = []
-        gpu_parts.append(f"{gpu_name or 'gpu'}")
         if gpu_memory is not None:
             gpu_parts.append(f"{gpu_memory}")
         if gpu_count is not None:
@@ -141,8 +126,13 @@ def pretty_resources(
         if compute_capability is not None:
             gpu_parts.append(f"{compute_capability}")
 
-        gpu = ":".join(gpu_parts)
-        parts.append(gpu)
+        if gpu_name:
+            parts.append("gpu=" + ":".join([f"{gpu_name}"] + gpu_parts))
+        elif gpu_vendor:
+            vendor_str = gpu_vendor.value if isinstance(gpu_vendor, enum.Enum) else str(gpu_vendor)
+            parts.append("gpu=" + ":".join([vendor_str] + gpu_parts))
+        else:
+            parts.append("gpu=" + ":".join(gpu_parts))
     return " ".join(parts)
 
 

src/tests/_internal/cli/services/configurators/test_run.py

@@ -132,13 +132,34 @@ def validate(self, conf: BaseRunConfiguration) -> None:
     def test_no_gpu(self):
         conf = self.prepare_conf()
         self.validate(conf)
-        assert conf.resources.gpu is None
+        assert conf.resources.gpu is not None
+        # Vendor is not written to spec for compatibility with older servers.
+        # The server infers nvidia in set_resources_defaults().
+        assert conf.resources.gpu.vendor is None
+        assert conf.resources.gpu.name is None
+        assert conf.resources.gpu.count.min == 0
 
     def test_zero_gpu(self):
         conf = self.prepare_conf(gpu_spec="0")
         self.validate(conf)
         assert conf.resources.gpu.vendor is None
 
+    def test_gpu_no_vendor_no_image_defaults_to_nvidia(self):
+        """Vendor is inferred as nvidia for validation but NOT written to spec."""
+        conf = self.prepare_conf(gpu_spec="1")
+        self.validate(conf)
+        assert conf.resources.gpu.vendor is None
+
+    def test_gpu_no_vendor_with_image_no_default(self):
+        conf = self.prepare_conf(gpu_spec="1", image="my-custom-image")
+        self.validate(conf)
+        assert conf.resources.gpu.vendor is None
+
+    def test_gpu_no_vendor_docker_true_no_default(self):
+        conf = self.prepare_conf(gpu_spec="1", docker=True)
+        self.validate(conf)
+        assert conf.resources.gpu.vendor is None
+
     @pytest.mark.parametrize(
         ["gpu_spec", "expected_vendor"],
         [

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

@@ -344,7 +344,14 @@ async def test_creates_fleet(self, test_db, session: AsyncSession, client: Async
                         "cpu": {"min": 2, "max": None},
                         "memory": {"min": 8.0, "max": None},
                         "shm_size": None,
-                        "gpu": None,
+                        "gpu": {
+                            "vendor": None,
+                            "name": None,
+                            "count": {"min": 0, "max": None},
+                            "memory": None,
+                            "total_memory": None,
+                            "compute_capability": None,
+                        },
                         "disk": {"size": {"min": 100.0, "max": None}},
                     },
                     "backends": None,
@@ -467,7 +474,14 @@ async def test_creates_ssh_fleet(self, test_db, session: AsyncSession, client: A
                         "cpu": {"min": 2, "max": None},
                         "memory": {"min": 8.0, "max": None},
                         "shm_size": None,
-                        "gpu": None,
+                        "gpu": {
+                            "vendor": None,
+                            "name": None,
+                            "count": {"min": 0, "max": None},
+                            "memory": None,
+                            "total_memory": None,
+                            "compute_capability": None,
+                        },
                         "disk": {"size": {"min": 100.0, "max": None}},
                     },
                     "backends": None,
@@ -639,7 +653,14 @@ async def test_updates_ssh_fleet(self, test_db, session: AsyncSession, client: A
                         "cpu": {"min": 2, "max": None},
                         "memory": {"min": 8.0, "max": None},
                         "shm_size": None,
-                        "gpu": None,
+                        "gpu": {
+                            "vendor": None,
+                            "name": None,
+                            "count": {"min": 0, "max": None},
+                            "memory": None,
+                            "total_memory": None,
+                            "compute_capability": None,
+                        },
                         "disk": {"size": {"min": 100.0, "max": None}},
                     },
                     "backends": None,

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

@@ -47,6 +47,10 @@
 from dstack._internal.server.models import JobModel, RunModel
 from dstack._internal.server.schemas.runs import ApplyRunPlanRequest
 from dstack._internal.server.services.projects import add_project_member
+from dstack._internal.server.services.resources import (
+    set_gpu_vendor_default,
+    set_resources_defaults,
+)
 from dstack._internal.server.services.runs import run_model_to_run
 from dstack._internal.server.services.runs.spec import validate_run_spec_and_set_defaults
 from dstack._internal.server.testing.common import (
@@ -1535,6 +1539,13 @@ async def test_returns_update_or_create_action_on_conf_change(
             run_spec=run_spec,
         )
         run = run_model_to_run(run_model)
+        # Apply the same defaults the server applies to current_resource
+        set_resources_defaults(run.run_spec.configuration.resources)
+        set_gpu_vendor_default(
+            run.run_spec.configuration.resources,
+            image=run.run_spec.configuration.image,
+            docker=getattr(run.run_spec.configuration, "docker", None),
+        )
         run_spec.configuration = new_conf
         response = await client.post(
             f"/api/project/{project.name}/runs/get_plan",