vllm_omni.host_weight_runtime ¶
Neutral contracts for immutable runtime-ready host weight artifacts.
Modules:
| Name | Description |
|---|---|
config | Typed policy for the loader-adjacent Host Weight Runtime. |
errors | Stable Host Weight Runtime failures and actions. |
filesystem | Local filesystem Host Weight Store implementation. |
identity | Deterministic semantic identity for runtime-ready host weights. |
inspection | Typed, read-only snapshots for local Host Weight Store inspection. |
lease | Process-local ownership for immutable mmap-backed host tensors. |
manifest | Versioned immutable manifests for runtime-ready host weights. |
outcomes | Typed store, resolution, and post-load publication outcomes. |
protocols | Provider and consumer protocols for Host Weight Runtime extensions. |
runtime | Policy-driven exact resolution over HostWeightStore contracts. |
AdaptationIdentity dataclass ¶
metadata class-attribute instance-attribute ¶
metadata: CanonicalJson = field(
default_factory=CanonicalJson.empty
)
ArtifactInspection dataclass ¶
ArtifactInventoryEntry dataclass ¶
ArtifactInventoryState ¶
ArtifactWriter ¶
Bases: Protocol
open_tensor_file ¶
open_tensor_file(
relative_name: str,
specs: tuple[TensorWriteSpec, ...],
*,
ordered: bool = False,
) -> TensorFileWriter
AttemptResult ¶
BuildRequest dataclass ¶
CanonicalJson dataclass ¶
Immutable snapshot of deterministic JSON metadata.
CapacityPolicy dataclass ¶
ComponentIdentity dataclass ¶
metadata class-attribute instance-attribute ¶
metadata: CanonicalJson = field(
default_factory=CanonicalJson.empty
)
CoordinationScope ¶
DomainInspection dataclass ¶
FailureCode ¶
ACTIVE_BUILD_TIMEOUT class-attribute instance-attribute ¶
CANONICAL_SOURCE_FAILED class-attribute instance-attribute ¶
COHORT_POLICY_MISMATCH class-attribute instance-attribute ¶
DOMAIN_POLICY_MISMATCH class-attribute instance-attribute ¶
INSUFFICIENT_CAPACITY class-attribute instance-attribute ¶
MANIFEST_INCOMPATIBLE class-attribute instance-attribute ¶
NUMA_PLACEMENT_INVALID class-attribute instance-attribute ¶
PRODUCER_UNAVAILABLE class-attribute instance-attribute ¶
PRODUCER_UNSUPPORTED class-attribute instance-attribute ¶
REMOTE_ARTIFACT_INVALID class-attribute instance-attribute ¶
RESTORER_UNAVAILABLE class-attribute instance-attribute ¶
RESTORE_COMMIT_FAILED class-attribute instance-attribute ¶
RESTORE_PREFLIGHT_FAILED class-attribute instance-attribute ¶
STORAGE_CLASS_MISMATCH class-attribute instance-attribute ¶
STORE_LIMIT_EXCEEDED class-attribute instance-attribute ¶
FileManifestEntry dataclass ¶
HostWeightError ¶
HostWeightFailure dataclass ¶
details class-attribute instance-attribute ¶
details: CanonicalJson = field(
default_factory=CanonicalJson.empty
)
HostWeightInspector ¶
Bases: Protocol
inspect_artifact ¶
inspect_artifact(
identity: WeightArtifactIdentity,
*,
validation: ValidationLevel,
) -> ArtifactInspection
HostWeightLease ¶
Stable tensor views and resources retained until transport teardown.
HostWeightLeaseCarrier ¶
Process-local, single-take ownership carrier for a host-weight lease.
The loader creates the carrier after a restore transaction has committed. The runner/backend takes the lease exactly once before starting any asynchronous transport work. If setup fails before that handoff, closing the carrier releases the loader-owned lease.
HostWeightResolution dataclass ¶
HostWeightRuntime ¶
Resolve or publish one exact representation without owning restoration.
from_config classmethod ¶
from_config(
config: HostWeightRuntimeConfig,
*,
observer: Callable[[ResolutionReport], None]
| None = None,
publication_observer: Callable[
[PostLoadPublicationReport], None
]
| None = None,
) -> HostWeightRuntime
Construct lazily so disabled mode performs no filesystem probe.
publish_after_load ¶
publish_after_load(
identity: WeightArtifactIdentity | None = None,
*,
producer: WeightProducer | None = None,
) -> PostLoadPublicationReport
Synchronously warm a future startup without revising resolution.
A successful store operation returns a validated lease. The runtime closes that lease before returning because POST_LOAD_ONLY production must not mutate or rebind the model serving the current startup.
resolve ¶
resolve(
identity: WeightArtifactIdentity | None = None,
*,
producer: WeightProducer | None = None,
) -> HostWeightResolution
HostWeightRuntimeConfig dataclass ¶
capacity class-attribute instance-attribute ¶
capacity: CapacityPolicy = field(
default_factory=CapacityPolicy
)
integrity class-attribute instance-attribute ¶
integrity: IntegrityPolicy = field(
default_factory=IntegrityPolicy
)
production class-attribute instance-attribute ¶
production: ProductionPolicy = field(
default_factory=ProductionPolicy
)
remote class-attribute instance-attribute ¶
remote: RemoteImportPolicy = field(
default_factory=RemoteImportPolicy
)
HostWeightStore ¶
Bases: Protocol
Exact artifact access whose deadlines bound lock acquisition only.
get_or_build ¶
get_or_build(
request: BuildRequest,
producer: WeightProducer,
*,
validation: ValidationLevel,
deadline: float,
) -> StoreResult
lookup ¶
lookup(
identity: WeightArtifactIdentity,
*,
validation: ValidationLevel,
deadline: float | None = None,
) -> StoreResult
IdentityValidationError ¶
Bases: ValueError
Raised when an artifact identity is not deterministic JSON.
IntegrityPolicy dataclass ¶
local_lookup class-attribute instance-attribute ¶
local_lookup: ValidationLevel = (
ValidationLevel.MANIFEST_AND_METADATA
)
LeaseProvenance dataclass ¶
LookupPhase ¶
ManifestValidationError ¶
Bases: ValueError
Raised when a published manifest violates its schema.
MappedHostRegion dataclass ¶
PostLoadPublicationOutcome ¶
Terminal outcome for one explicit post-load publication attempt.
PostLoadPublicationReport dataclass ¶
ProducerIdentity dataclass ¶
ProductionMetadata dataclass ¶
format_metadata class-attribute instance-attribute ¶
format_metadata: CanonicalJson = field(
default_factory=CanonicalJson.empty
)
ProductionPolicy dataclass ¶
ProductionSourceMode ¶
PublicationMarker dataclass ¶
RemoteArtifactDescriptor dataclass ¶
RemoteArtifactProvider ¶
Bases: Protocol
fetch ¶
fetch(
descriptor: RemoteArtifactDescriptor,
writer: ArtifactWriter,
) -> ProductionMetadata
RemoteImportPolicy dataclass ¶
on_local_miss class-attribute instance-attribute ¶
on_local_miss: RemoteOnMiss = RemoteOnMiss.DISABLED
RemoteOnMiss ¶
ResolutionAttempt dataclass ¶
ResolutionOutcome ¶
ResolutionReport dataclass ¶
ResolutionStage ¶
RuntimeMode ¶
RuntimeWeightLayout dataclass ¶
metadata class-attribute instance-attribute ¶
metadata: CanonicalJson = field(
default_factory=CanonicalJson.empty
)
StorageClass ¶
StorageDomainPolicy dataclass ¶
StorageScope ¶
StoreResult dataclass ¶
StoreStatus ¶
TensorFileWriter ¶
TensorKind ¶
TensorManifestEntry dataclass ¶
semantic_dict ¶
Return content identity independent of physical file sharding.
TensorWriteSpec dataclass ¶
ValidationLevel ¶
WaitPolicy dataclass ¶
WeightArtifactIdentity dataclass ¶
WeightManifest dataclass ¶
format_metadata class-attribute instance-attribute ¶
format_metadata: CanonicalJson = field(
default_factory=CanonicalJson.empty
)
WeightProducer ¶
Bases: Protocol
Synchronous V1 producer; callers do not preempt an active build.
produce ¶
produce(writer: ArtifactWriter) -> ProductionMetadata
Produce one exact artifact synchronously or raise.
WeightProductionSpec dataclass ¶
coordination_scope class-attribute instance-attribute ¶
coordination_scope: CoordinationScope = (
CoordinationScope.SINGLE_PROCESS
)
lookup_phase class-attribute instance-attribute ¶
lookup_phase: LookupPhase = LookupPhase.PRE_LOAD_SAFE
WeightRepresentation dataclass ¶
metadata class-attribute instance-attribute ¶
metadata: CanonicalJson = field(
default_factory=CanonicalJson.empty
)
WeightRestorePlan ¶
WeightRestorer ¶
Bases: Protocol
Validate a lease-to-model restore without mutating either input.
plan_restore ¶
plan_restore(
model: object, lease: HostWeightLease
) -> WeightRestorePlan
Return a validated plan; all model mutation is deferred to commit().
WeightSourceIdentity dataclass ¶
metadata class-attribute instance-attribute ¶
metadata: CanonicalJson = field(
default_factory=CanonicalJson.empty
)