Struct PythonTorchProject 

pub struct PythonTorchProject { /* private fields */ }

Python/Torch-backed workload bridge implemented on top of the generic p2p runtime.

Implementations

impl PythonTorchProject

pub fn new(config: PythonTorchWorkloadConfig) -> Result<Self>

Spawns the backing Python worker and probes its runtime capability.

pub fn new_with_data_pipeline( config: PythonTorchWorkloadConfig, data_pipeline: LeaseDataPipeline<String, PythonBatchRef>, ) -> Result<Self>

Spawns the backing Python worker with an explicit lease data pipeline.

pub fn sharded_data_pipeline( dataset: &PythonTorchDatasetConfig, ) -> LeaseDataPipeline<String, PythonBatchRef>

Returns the default shard-backed lease data pipeline for one dataset config.

pub fn micro_epoch_pipeline( descriptor: LeaseDataPipelineDescriptor, dataset_registration: impl Fn() -> Result<DatasetRegistration> + Send + Sync + 'static, microshard_plan: impl Fn(&DatasetRegistration) -> Result<MicroShardPlan> + Send + Sync + 'static, payload: impl Fn(&AssignmentLease, &[CachedMicroShard]) -> Result<Value> + Send + Sync + 'static, ) -> LeaseDataPipeline<String, PythonBatchRef>

Builds a generic Python micro-epoch pipeline backed by workload-defined payloads.

pub fn indexed_dataset_pipeline( pipeline_name: impl Into<String>, dataset_registration: impl Fn() -> Result<DatasetRegistration> + Send + Sync + 'static, microshard_plan: impl Fn(&DatasetRegistration) -> Result<MicroShardPlan> + Send + Sync + 'static, payload: impl Fn(&AssignmentLease, &[CachedMicroShard]) -> Result<Value> + Send + Sync + 'static, ) -> LeaseDataPipeline<String, PythonBatchRef>

Builds a Python pipeline for existing torch Dataset/Sampler-style data flows.

pub fn generated_dataset_pipeline( pipeline_name: impl Into<String>, dataset_registration: impl Fn() -> Result<DatasetRegistration> + Send + Sync + 'static, microshard_plan: impl Fn(&DatasetRegistration) -> Result<MicroShardPlan> + Send + Sync + 'static, payload: impl Fn(&AssignmentLease, &[CachedMicroShard]) -> Result<Value> + Send + Sync + 'static, ) -> LeaseDataPipeline<String, PythonBatchRef>

Builds a Python pipeline for deterministic synthetic or recipe-driven data generation.

pub fn probe_capability(&self) -> &CapabilityEstimate

Returns the worker-advertised capability estimate.

pub fn runtime_device_name(&self) -> &str

Returns the resolved runtime device tag.

pub fn workload_name(&self) -> &str

Returns the Python-side workload name advertised by the worker.

pub fn data_pipeline_descriptor(&self) -> &LeaseDataPipelineDescriptor

Returns the static lease/micro-epoch data pipeline descriptor.

pub fn data_pipeline_kind(&self) -> LeaseDataPipelineKind

Returns the configured lease/micro-epoch pipeline kind.

pub fn data_pipeline_registration(&self) -> Result<DatasetRegistration>

Returns the dataset registration backing the current pipeline.

pub fn local_upstream_root(&self) -> Result<Option<PathBuf>>

Returns the local upstream root when the current pipeline is backed by a Local dataset registration.

pub fn configured_shard_root(&self) -> &Path

Returns the configured shard root for the default sharded Python dataset config. This is configuration data only and may be unrelated to the active pipeline when new_with_data_pipeline(...) is used.

Trait Implementations

impl Clone for PythonTorchProject

fn clone(&self) -> PythonTorchProject

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for PythonTorchProject

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl P2pWorkload for PythonTorchProject

type Device = String

Defines the device alias.

type Model = PythonModelHandle

Defines the model alias.

type Batch = PythonBatchRef

Defines the batch alias.

type WindowStats = BTreeMap<String, MetricValue>

Defines the window stats alias.

fn init_model(&self, device: &Self::Device) -> Self::Model

Initializes a model instance for the provided backend device.

fn benchmark( &self, _model: &Self::Model, _device: &Self::Device, ) -> CapabilityEstimate

Benchmarks the workload and reports the runtime capability estimate.

fn train_window( &self, ctx: &mut WindowCtx<Self::Device, Self::Model, Self::Batch>, ) -> Result<WindowReport<Self::WindowStats>, TrainError>

Runs one training window for the leased batches.

fn evaluate(&self, model: &Self::Model, split: EvalSplit) -> MetricReport

Evaluates the model on the requested dataset split.

fn apply_patch(&mut self, patch: &RuntimePatch) -> PatchOutcome

Applies a runtime patch to the workload implementation.

fn supported_patch_classes(&self) -> PatchSupport

Returns the patch classes accepted by the workload implementation.

fn runtime_device(&self) -> Self::Device

Returns the runtime device used by the workload.

fn dataset_registration(&self) -> Result<DatasetRegistration>

Returns the dataset registration used to plan microshards.

fn microshard_plan( &self, _registration: &DatasetRegistration, ) -> Result<MicroShardPlan>

Builds the microshard plan for the registered dataset.

fn load_batches( &self, lease: &AssignmentLease, cached_microshards: &[CachedMicroShard], ) -> Result<Vec<Self::Batch>>

Loads training batches for the lease from cached microshards.

fn load_model_artifact( &self, model: Self::Model, descriptor: &ArtifactDescriptor, store: &FsArtifactStore, _device: &Self::Device, ) -> Result<Self::Model>

Loads a model artifact from the artifact store into the runtime model representation.

fn materialize_model_artifact( &self, model: &Self::Model, artifact_kind: ArtifactKind, head_id: HeadId, base_head_id: Option<HeadId>, store: &FsArtifactStore, ) -> Result<ArtifactDescriptor>

Materializes a model artifact into the checkpoint store.

fn contribution_metrics( &self, report: &WindowReport<Self::WindowStats>, ) -> BTreeMap<String, MetricValue>

Returns receipt metrics for a completed training window.

fn contribution_weight(&self, report: &WindowReport<Self::WindowStats>) -> f64

Returns the contribution weight used for receipt scoring.

fn reconcile_canonical_model( &self, local_model: &Self::Model, canonical_model: Self::Model, strategy: TrainerCanonicalReconcileStrategy, ) -> Result<Self::Model>

Reconciles a speculative local trainer model with a newly visible canonical model.

fn merge_candidate_models( &self, base_model: &Self::Model, candidates: &[MergeModelCandidate<'_, Self::Model>], policy: MergePolicy, ) -> Result<Option<Self::Model>>

Optionally merges candidate models into one merged model.

fn apply_single_root_ema( &self, base_model: &Self::Model, merged_model: Self::Model, policy: MergePolicy, ) -> Result<Self::Model>

Optionally applies single-root EMA after merge selection.

fn supported_workload(&self) -> SupportedWorkload

Returns the supported workload manifest.

fn model_schema_hash(&self) -> ContentId

Returns the stable model schema hash.

fn workload_id(&self) -> WorkloadId

Returns the workload id.

fn switch_runtime_workload( &mut self, workload_id: &WorkloadId, ) -> Result<(), Error>

Switches the live runtime to one other compiled workload when supported.

fn checkpoint_format_hash(&self) -> ContentId

Returns the checkpoint format hash.

fn verify_revision(&self, revision: &RevisionManifest) -> Result<(), Error>

Verifies that the runtime workload matches one revision manifest.