Drop support for ResourcesSpec.cpu as Range[int] (#4091)

Breaks compatibility with clients < 0.19.8 (2025-05-07).

CPUSpec.parse still supports and must continue to support infinitely
parsing min/max ranges for old ResourceSpecs stored in the DB.

Author: un-def

frontend/src/pages/Offers/List/index.tsx

@@ -48,7 +48,7 @@ const getRequestParams = ({
                 env: {},
                 resources: {
                     // cpu/memory/disk should match ResourcesSpec.unconstrained() used by `dstack offer` CLI command
-                    cpu: { min: 1 },
+                    cpu: { count: { min: 1 } },
                     memory: { min: 0.0 },
                     disk: null,
                     gpu: {

frontend/src/types/gpu.d.ts

@@ -22,8 +22,13 @@ declare interface IGPUSpecRequest {
     compute_capability?: any[];
 }
 
+declare interface ICPUSpecRequest {
+    arch?: 'x86' | 'arm' | null;
+    count?: TRange | number | string;
+}
+
 declare interface IResourcesSpecRequest {
-    cpu?: TRange | number | string;
+    cpu?: ICPUSpecRequest | number | string;
     memory?: TRange | number | string;
     shm_size?: number | string;
     gpu?: IGPUSpecRequest | number | string;

frontend/src/types/run.d.ts

@@ -49,7 +49,7 @@ declare type TRange = { min?: number; max?: number };
 
 declare type TResourceRequest = {
     gpu?: TGPUResources | string | number;
-    cpu?: string | number | TRange;
+    cpu?: string | number | ICPUSpecRequest;
     memory?: string | number | TRange;
     shm_size?: string | number;
     disk?:

src/dstack/_internal/cli/models/presets.py

@@ -14,7 +14,7 @@
 from dstack._internal.core.models.common import CoreModel
 from dstack._internal.core.models.configurations import ServiceConfiguration
 from dstack._internal.core.models.profiles import ProfileParams
-from dstack._internal.core.models.resources import CPUSpec, ResourcesSpec
+from dstack._internal.core.models.resources import ResourcesSpec
 
 
 class PresetBenchmarkWorkload(CoreModel):
@@ -159,7 +159,7 @@ class PresetListOutput(CoreModel):
 
 
 def _validate_exact_resources(resources: ResourcesSpec) -> None:
-    cpu = CPUSpec.model_validate(resources.cpu)
+    cpu = resources.cpu
     if not _is_exact(cpu.count) or not _is_exact(resources.memory):
         raise ValueError("preset validation resources must be exact")
     if resources.disk is None or not _is_exact(resources.disk.size):

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

@@ -59,7 +59,6 @@
 from dstack._internal.core.models.repos import RepoHeadWithCreds
 from dstack._internal.core.models.repos.base import Repo
 from dstack._internal.core.models.repos.remote import RemoteRepo, RemoteRepoCreds
-from dstack._internal.core.models.resources import CPUSpec
 from dstack._internal.core.models.runs import JobStatus, JobSubmission, RunPlan, RunSpec, RunStatus
 from dstack._internal.core.services.diff import diff_models
 from dstack._internal.core.services.repos import get_repo_creds_and_default_branch
@@ -525,8 +524,7 @@ def validate_cpu_arch_and_image(self, conf: RunConfigurationT) -> None:
         """
         Infers `resources.cpu.arch` if not set, requires `image` if the architecture is ARM.
         """
-        # TODO: Remove in 0.20. Use conf.resources.cpu directly
-        cpu_spec = CPUSpec.model_validate(conf.resources.cpu)
+        cpu_spec = conf.resources.cpu
         arch = cpu_spec.arch
         if arch is None:
             gpu_spec = conf.resources.gpu

src/dstack/_internal/core/backends/base/offers.py

@@ -16,7 +16,7 @@
     InstanceType,
     Resources,
 )
-from dstack._internal.core.models.resources import DEFAULT_DISK, CPUSpec, GPUSpec, Memory, Range
+from dstack._internal.core.models.resources import DEFAULT_DISK, GPUSpec, Memory, Range
 from dstack._internal.core.models.runs import Job, Requirements, Run
 from dstack._internal.utils.common import get_or_error
 
@@ -170,8 +170,7 @@ def requirements_to_query_filter(req: Optional[Requirements]) -> gpuhunt.QueryFi
 
     res = req.resources
     if res.cpu:
-        # TODO: Remove in 0.20. Use res.cpu directly
-        cpu = CPUSpec.model_validate(res.cpu)
+        cpu = res.cpu
         q.cpu_arch = cpu.arch
         q.min_cpu = cpu.count.min
         q.max_cpu = cpu.count.max

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

@@ -26,7 +26,7 @@
     InstanceType,
     Resources,
 )
-from dstack._internal.core.models.resources import CPUSpec, Memory, ResourcesSpec
+from dstack._internal.core.models.resources import Memory, ResourcesSpec
 from dstack._internal.utils import docker as docker_utils
 from dstack._internal.utils.common import get_or_error
 from dstack._internal.utils.logging import get_logger
@@ -179,7 +179,6 @@ class ResourceRequests(ResourceRequestsLimits):
 
     @classmethod
     def from_resources_spec(cls, spec: ResourcesSpec) -> Self:
-        assert isinstance(spec.cpu, CPUSpec)
         cpu = spec.cpu.count.min or 0
         memory_mib: int = 0
         if spec.memory.min is not None:
@@ -223,7 +222,6 @@ def from_kubernetes_map(cls, map_: Mapping[str, str]) -> Self:
 class ResourceLimits(ResourceRequestsLimits):
     @classmethod
     def from_resources_spec(cls, spec: ResourcesSpec) -> Self:
-        assert isinstance(spec.cpu, CPUSpec)
         cpu = spec.cpu.count.max
         memory_mib: Optional[int] = None
         if spec.memory.max is not None:
@@ -236,7 +234,6 @@ def from_resources_spec(cls, spec: ResourcesSpec) -> Self:
         if spec.gpu is not None:
             # GPU resources cannot be overcommitted, limit must be equal to request
             gpu = spec.gpu.count.min or 0
-        assert isinstance(spec.cpu, CPUSpec)
         return cls(
             cpu=cpu,
             memory_mib=memory_mib,

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

@@ -7,7 +7,6 @@
 from dstack._internal.core.models.instances import Gpu
 from dstack._internal.core.models.resources import (
     DEFAULT_MEMORY_SIZE,
-    CPUSpec,
     Memory,
     ResourcesSpec,
 )
@@ -118,7 +117,6 @@ class RequestedResources:
 
 
 def get_requested_resources_from_resources_spec(spec: ResourcesSpec) -> RequestedResources:
-    assert isinstance(spec.cpu, CPUSpec)
     # 1 is the default value of --cpus-per-task
     cpu_count = spec.cpu.count.min or 1
 

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

@@ -9,10 +9,7 @@
     Field,
     GetCoreSchemaHandler,
     GetJsonSchemaHandler,
-    SerializerFunctionWrapHandler,
-    Tag,
     field_validator,
-    model_serializer,
     model_validator,
 )
 from pydantic.json_schema import JsonSchemaValue
@@ -227,11 +224,9 @@ def parse(cls, v: Any) -> Any:
         # Range and min/max dict - for backward compatibility
         if isinstance(v, Range):
             return {"arch": None, "count": v}
-        # A subset rather than exactly {"min", "max"}: `ResourcesSpec` serializes `cpu` down to its
-        # count for old clients, and under `exclude_none=True` that leaves just `{"min": ...}`.
-        # Requiring both keys made the round trip land on the `Range[int]` arm of `ResourcesSpec.cpu`
-        # instead of coming back as a `CPUSpec`. `arch`/`count` are the only `CPUSpec` fields, so a
-        # mapping of min/max is unambiguously a range.
+        # `arch` and `count` are the only `CPUSpec` fields, so a mapping of `min`/`max` is
+        # unambiguously a count range. A subset rather than exactly `{"min", "max"}`, because a
+        # half-open range may omit the other key.
         if isinstance(v, Mapping) and v and v.keys() <= {"min", "max"}:
             return {"arch": None, "count": v}
         return v
@@ -395,20 +390,7 @@ def _parse(cls, v: Any) -> Any:
 
 
 class ResourcesSpec(CoreModel):
-    # TODO: remove `Range[int]` in 0.20. It is kept only for backward compatibility.
-    cpu: Annotated[
-        Union[
-            # `Tag` only names the arm in validation errors. Without it the `loc` of a bad `cpu`
-            # spells out the whole wrapped schema —
-            # `cpu.function-before[parse(), function-before[parse(), ... CPUSpec]].count` — which
-            # is what `dstack apply` shows the user.
-            Annotated[CPUSpec, Tag("CPUSpec")],
-            Annotated[Range[int], Tag("Range[int]")],
-        ],
-        # `CPUSpec` and `Range[int]` both accept a bare int/str, so the arm has to be picked by
-        # declaration order rather than by pydantic v2's "smart" union resolution.
-        Field(description="The CPU requirements", union_mode="left_to_right"),
-    ] = CPUSpec()
+    cpu: Annotated[CPUSpec, Field(description="The CPU requirements")] = CPUSpec()
     memory: Annotated[Range[Memory], Field(description="The RAM size (e.g., `8GB`)")] = (
         DEFAULT_MEMORY_SIZE
     )
@@ -435,8 +417,7 @@ def unconstrained(cls) -> "ResourcesSpec":
         )
 
     def pretty_format(self) -> str:
-        # TODO: Remove in 0.20. Use self.cpu directly
-        cpu = CPUSpec.model_validate(self.cpu)
+        cpu = self.cpu
         resources: Dict[str, Any] = dict(cpu_arch=cpu.arch, cpus=cpu.count, memory=self.memory)
         if self.gpu:
             gpu = self.gpu
@@ -452,18 +433,3 @@ def pretty_format(self) -> str:
             resources.update(disk_size=self.disk.size)
         res = pretty_resources(**resources)
         return res
-
-    @model_serializer(mode="wrap")
-    def _serialize(self, handler: SerializerFunctionWrapHandler) -> Dict[str, Any]:
-        res = handler(self)
-        self._update_serialized_cpu(res)
-        return res
-
-    # TODO: Remove in 0.20. Added for backward compatibility.
-    def _update_serialized_cpu(self, values: Dict):
-        cpu = values.get("cpu")
-        if cpu:
-            arch = cpu.get("arch")
-            count = cpu.get("count")
-            if count and arch in [None, gpuhunt.CPUArchitecture.X86.value]:
-                values["cpu"] = count

src/dstack/_internal/server/services/requirements/combine.py

@@ -171,7 +171,7 @@ def _combine_tags(value1: dict[str, str], value2: dict[str, str]) -> dict[str, s
 
 def _combine_resources(value1: ResourcesSpec, value2: ResourcesSpec) -> ResourcesSpec:
     return ResourcesSpec(
-        cpu=_combine_cpu(value1.cpu, value2.cpu),  # type: ignore[attr-defined]
+        cpu=_combine_cpu(value1.cpu, value2.cpu),
         memory=_combine_memory(value1.memory, value2.memory),
         shm_size=_combine_shm_size_optional(value1.shm_size, value2.shm_size),
         gpu=_combine_gpu_optional(value1.gpu, value2.gpu),