Expand description
GPU kernel provider for XLOG
Re-exports§
pub use arrow_device::ArrowDeviceArray;pub use arrow_device::ArrowDeviceArrayOwned;pub use arrow_device::ARROW_DEVICE_CUDA;pub use cuda_compat::AsKernelParam;pub use cuda_compat::CudaFunction;pub use cuda_compat::IntoKernelParamStorage;pub use cuda_compat::KernelParamStorage;pub use cuda_compat::KernelScalar;pub use cuda_compat::LaunchAsync;pub use device::CudaDevice;pub use device_pool::GpuDevicePool;pub use dlpack::DLManagedTensor;pub use dlpack::DlpackManagedTensor;pub use dlpack::DlpackTable;pub use joint_constraint::CarrierBufferId;pub use joint_constraint::CarrierError;pub use joint_constraint::CarrierExport;pub use joint_constraint::JointConstraintCarrier;pub use joint_solver::Component;pub use joint_solver::ConstraintGraph;pub use joint_solver::Feasibility;pub use joint_solver::FuelMeter;pub use joint_solver::SolveStrategy;pub use joint_solver::SolverError;pub use joint_solver::SOLVER_ABI_IDENTITY;pub use memory::CudaBuffer;pub use memory::CudaColumn;pub use memory::GpuMemoryManager;pub use memory::RuntimeAllocBlock;pub use multi_gpu_memory::MultiGpuMemoryManager;pub use provider::circuit_kernels;pub use provider::dedup_kernels;pub use provider::filter_kernels;pub use provider::groupby_kernels;pub use provider::ilp_kernels;pub use provider::join_kernels;pub use provider::pack_kernels;pub use provider::pir_kernels;pub use provider::scan_kernels;pub use provider::set_ops_kernels;pub use provider::sort_kernels;pub use provider::CompareOp;pub use provider::CudaKernelProvider;pub use provider::CudaProviderBuilder;pub use provider::FjDeltaCols;pub use provider::FjNode;pub use provider::FjPlan;pub use provider::FjSubAtom;pub use provider::IlpExactNaryPatterns;pub use provider::IlpExactNaryRequest;pub use provider::JoinIndexV2;pub use provider::JoinType;pub use provider::CIRCUIT_MODULE;pub use provider::DEDUP_MODULE;pub use provider::GROUPBY_MODULE;pub use provider::ILP_MODULE;pub use provider::JOIN_MODULE;pub use provider::PACK_MODULE;pub use provider::PIR_MODULE;pub use provider::SCAN_MODULE;pub use provider::SET_OPS_MODULE;pub use provider::SORT_MODULE;pub use wcoj_metadata::HeatDist;pub use wcoj_metadata::LayoutSignature;pub use wcoj_metadata::RootMetadata;pub use wcoj_metadata::VertexId;pub use wcoj_metadata::WcojCycle4HgWorkPlanU32;pub use wcoj_metadata::WcojCycle4HgWorkPlanU64;pub use wcoj_metadata::WcojRelationMetadata;pub use wcoj_metadata::WcojTriangleHgWorkPlanU32;pub use wcoj_metadata::WcojTriangleHgWorkPlanU64;
Modules§
- arrow_
device - cuda_
compat - cuda_
graph - CUDA Graph RAII helpers for production graph capture/replay.
- device
- CUDA device management
- device_
pool - Multi-GPU device pool management
- device_
runtime - Stream-ordered device memory runtime.
- dlpack
- DLPack interop for zero-copy GPU exchange.
- joint_
constraint - Joint constraint carrier: buffer ownership, registration, and the device-resident label-feasibility solve stage.
- joint_
solver - Joint constraint solver — skeleton: deterministic component decomposition, pinned-width envelope, and typed fuel accounting.
- kernel_
manifest_ data - launch
- Owner-bearing access recorder for one CUDA launch.
- memory
- CUDA memory management
- multi_
gpu_ memory - Multi-GPU memory management
- provider
- CUDA kernel provider implementation
- sys
- type_
seam - GpuScalar — marker trait for Rust scalar types that round-trip through GPU column storage.
- wcoj_
metadata
Structs§
- Cuda
Slice Vec<T>on a cuda device. You can allocate and modify this with CudaStream.- Cuda
Stream - A wrapper around sys::CUstream that you can schedule work on.
- Cuda
View &[T]on a cuda device. An immutable sub-view into a CudaSlice created by CudaSlice::as_view()/CudaSlice::slice().- Cuda
View Mut &mut [T]on a cuda device. A mutable sub-view into a CudaSlice created by CudaSlice::as_view_mut()/CudaSlice::slice_mut().- Driver
Error - Wrapper around sys::CUresult. See nvidia’s CUresult docs
- Identity256
- Canonical 256-bit identity, preserving byte order without native-word conversion.
- Launch
Config - Configuration for [result::launch_kernel]
- Semantic
Action Catalogue - The sole built-in action law; both Rust and CUDA are generated from its declaration.
- Semantic
Active Row - Identity of one packed active-cache row in its original forward. Kinds are FILLED=1, MASK=2 and FEEDBACK=3; physical_row is not the packed coordinate.
- Semantic
Active Rows - Acquired bank-local active-cache state. A computed empty result has no rows but is distinct from the initial parent which has never consumed a forward.
- Semantic
Admission - Immutable typed content and identities held by one graph for one acquired base.
- Semantic
Admission Limits - Aggregate cold-input limits, checked before copying symbols or allocating views.
- Semantic
Admission Records - Cold input transferred into the semantic owner’s immutable admission.
- Semantic
Catalogue Binding - Semantic
Catalogue Field - Field codebook role and NULL-only cold cardinality in this catalogue.
- Semantic
Catalogue Signature - Declarative action/opcode signature, including dependent opcode selection.
- Semantic
Checkpoint Native Projection - Read-only identity projected from one fully validated original publication material.
- Semantic
Component - One row of the generated roster. Text fields are text positions; edit fields index the eighteen-field action signature. Offsets count input elements.
- Semantic
Component Input - Cold-bound FP32 logits and binary product support in original category order.
- Semantic
Continuation Input - Original model outputs for one acquired parent. The Session supplies their instance, publication word, task and generation binding from its real lease.
- Semantic
Extents - Exact reachable logical extents.
- Semantic
Feedback - Features, validity and current causal provenance from one acquired bank. Original outputs retain their step independently of later bank retirement.
- Semantic
Feedback Record Material - Original non-tensor feedback material from a completed acquired step.
identityis its native range seal, not SHA-256 ofbytesalone. Raw slots (role 15), statement bytes (16) and provenance records (17) stay separate; encoded feature tensors are not a replacement for any of these records. - Semantic
Feedback Schema - Value-independent flat coordinates emitted by the native feedback producer.
- Semantic
Fork Handle - Semantic
Gradient Delivery Binding - Retained original tensor producers for one prepared bank’s autograd delivery boundary. The binding owns the consumed DLPack records until the Python-side delivery owner is retired; it grants no publication authority.
- Semantic
Hypergraph - Device-resident immutable-root semantic owner retaining its provider.
- Semantic
Hypergraph Capacities - Checked fixed capacities for one resident semantic graph.
- Semantic
Hypergraph Execution Stats - Production-path execution counters.
- Semantic
Initial Prefill Content Witness - Exact content captured by one imported task’s original model prefill. This handle never creates a new baseline during verification.
- Semantic
Initial Prefill Lease - Opaque handle to the single imported-task prefill stage. This is not a model-forward receipt; the actual input and outputs must be joined before the parent can be published.
- Semantic
Initial Source Layout - Initial source geometry issued from the admitted observation mapping before the model computes its prefix cache. The rows retain zero provenance indices; parent binding constructs and verifies the private ledger after prefill.
- Semantic
Inserted Support - Complete result of inserting a new support event.
- Semantic
Learning Phase Record - Retained evidence for one actual native cold phase change. Ordinary Update and restore preserve this history; neither can reset the cumulative progress.
- Semantic
Learning Phase Transition - A transition request, not an assertion that a scientific criterion passed. The privileged Controller supplies its original signed, durably read-back Admission, separately from checking original data-use grants. This is permission for private work, never a claim that the scientific comparison passed.
- Semantic
Model Context - Model metadata projected from one acquired publication, not supplied by the model. Content identities grant no authority independently of that reader.
- Semantic
Model Contract Layout - Byte spans in the producer’s one ModelContract record. The native owner hashes the immutable schema bytes without interpreting their format, then binds them to the actual model generation and device numerical identity. Digest fields are raw 32 bytes; generation is an unaligned little-endian u64.
- Semantic
Model Forward Witness - Opaque proof that the native receipt node for one exact logits occurrence was inserted while its model producer’s prepared bank capture was active.
- Semantic
Model Memory - Full producer-owned backing buffers and their independent storage/view map. Allocation inputs are contiguous U8 vectors with role zero and ordinal index. Tensor and version-counter identities remain in the opaque model contract.
- Semantic
Model Storage - A producer-declared storage object within one complete backing allocation. Storage identity is its ordinal, not an address or a tensor version counter.
- Semantic
Model View - A typed publication coordinate within a producer-declared storage object.
- Semantic
Observed Source - An observed token payload admitted by the trusted application controller.
originretains the complete source/manifest/transformation record, not an independently supplied native digest or a model-generated provenance row. - Semantic
Output Intent - Read-only delivery material from a sealed publication. Reading an intent neither proves that the checkpoint was durably committed nor grants authority to perform its external effect.
- Semantic
Parent Binding - Initial canonical parent only. A model continuation cannot be supplied here: it is admitted after a real Acquire and a forward using that acquired parent. PrefixIdentity is native-owned: fresh publication derives its 32 bytes from the actual PrefixSource. Supplying a caller record for that role is an error.
- Semantic
Policy Field Layout - FP32 element ranges for one field in the native policy parameter snapshot. NULL is a category, but has no stored embedding row. The action-class field has no NULL: its category zero is the trainable no-edit action.
- Semantic
Policy Layout - Packing of the immutable policy inputs consumed by the serial component loop. These ranges describe storage, not authority to substitute a model generation. A session derives the layout from its retained admission’s actual codebooks.
- Semantic
Predicate Record - A predicate and its complete existing XLOG schema, not a caller-supplied digest.
- Semantic
Prepared Step - A fixed future step issued by its Session during cold segment construction. It owns no host publication identity; its actual parent is acquired on device.
- Semantic
Program Admission - Exact initial program and input facts prepared before a resident task starts.
- Semantic
Program Fact - One typed binary tuple. Predicate indices come from the admitted declaration order.
- Semantic
Program Rule - One range-restricted positive rule with one or two binary body atoms. Equal variable indices encode repeated-variable equalities.
- Semantic
Publication Input Projection - Read-only original inputs projected through the complete publication codec. These content records are not proof of historical execution or use authority.
- Semantic
Publication Tensor Projection - Original cache geometry and logical bytes under the native tensor seal. Logical bytes exclude stride gaps and unused capacity, exactly as the seal does.
- Semantic
Published Identity - Actual device-validated publication identity. This identifies content and lineage; it is not a use-authority token or an externally issuable lease.
- Semantic
Published Lease - An actual device-acquired reader. Only its originating Session can inspect, export or release it; it is neither cloneable nor externally constructible.
- Semantic
Replay Material - Owned cold replay inputs: complete predecessor material and the successor’s external native publication evidence. Hash consistency does not establish historical execution, current rights, or permission to use a model.
- Semantic
Resident Model Memory - Direct read-only aliases of the selected resident neural bank. Each backing allocation is exported exactly once; storages, views and layouts preserve the producer’s original alias geometry. The acquired reader remains the use owner and therefore prevents reuse of this physical bank through final use.
- Semantic
RngBinding - Semantic
Root Digest - Content digest of an immutable semantic root.
- Semantic
Root Handle - Semantic
Root Snapshot - Immutable device-derived root receipt.
- Semantic
Segment Cold Capacity - Frozen bounds for the content, work, and external allocations currently admitted during prepared-segment construction. Numerical-domain buffers require their own native-derived extension before the full peak is claimed.
- Semantic
Selected Training View - Retained native result of device selection. The views are fixed-capacity CUDA ports; the resident selection record supplies their logical extent.
- Semantic
Source Mapping - A token is selected only by offset into its admitted observed payload.
- Semantic
State Record - One actual canonical control record or queue payload retained by a parent. Mutable queue capacity is explicit; bytes contain the real encoded header and entries, including an empty queue’s header rather than an invented state.
- Semantic
Statement Handle - Semantic
Statement Identity - Content identity of a semantic statement.
- Semantic
Statement Key - Canonical, identity-bearing statement key.
- Semantic
Statement Ref - Statement identity paired with its opaque physical handle.
- Semantic
Structural Cost Descriptor - Task-independent unit and finite cap for native structural work. The unit names the versioned counting law and the resident program mode, not the authority, task identity, or outcome of a sampled candidate.
- Semantic
Support Event - One provenance-bearing support event. Presence is exactly one.
- Semantic
Support Handle - Semantic
Support Identity - Content identity of one exact support event.
- Semantic
Support Record - One unit of support and its complete, role-checked record references.
- Semantic
Support Ref - Support identity paired with its opaque physical handle.
- Semantic
Task Content Identity - Independently recomputable identities of the exact semantic task content. These values carry no use authority. They let retained symbolic training rows bind the ordered query, admitted theory plus executed observer program, and observed four-valued result to the task that the native owner actually executed during cold binding.
- Semantic
Task Evaluation - One native evaluation of the fixed base/learned roster. The selector runs only on device; these are the resulting facts, not caller-supplied scores.
- Semantic
Task Evaluation Spec - Three ordered query selections for the bounded transition ABI. Predicate, arguments, observer program, eligibility, and scoring are application inputs. This specification carries no authority; only the application controller may attach an authorized task to its retained semantic session.
- Semantic
Task Facts - Device-computed facts, using the three actual canonical truth-query results.
- Semantic
Task Goal Witness - Native seal of the independently validated goal authority carried by the controller. The semantic root is the exact XLOG root inspected for the task; the remaining roots bind the complete mandatory dependency and constraint closures. These values are derived by the trusted controller after its closure checks, never supplied by the model-update consumer.
- Semantic
Task Objective Law - Evaluator interval and objective coefficients frozen by the native task before actions, independently of any future training-group membership.
- Semantic
Task Observation - Actual native observer output and the exact source/input/result custody used to derive it. These bytes describe execution; they grant no application rights.
- Semantic
Task Observation Roots - Owner-validated cold read coverage for a task’s queried semantic heads. These records grant no use or mutation rights and are not current device query receipts. Runtime feedback must still establish its actual live view.
- Semantic
Task Priority Goal - One normalized truth-match objective within an ordered task priority level.
The native query result contributes one when it equals
target_truth, zero otherwise; candidate progress subtracts the acquired base’s contribution. - Semantic
Task Priority Level - Semantic
Task Scoring - Explicit nonnegative coefficients for query agreement and measured work. These coefficients determine the measured value and final return; ordered task priorities, agreement, and cost select the candidate in that order.
- Semantic
Tensor Content Witness - Private device content witness for the exact tensors of one retained parent. Transient witnesses prove stability after capture; model witnesses compare against the originally sealed model baseline. Neither proves computation. The trusted model owner binds the witness to its actual original execution. No digest, device pointer, status, or writable witness storage is exported.
- Semantic
Tensor Input - A genuine producer-owned model tensor and its exact cold model contract. No numeric conversion occurs. Logical intervals describe covered positions; non-positional weights and optimizer state use the explicit interval 0..0.
- Semantic
Tensor Layout - Semantic
Text Row - One original compact model row; source_slot is the acquired ring identity, never the packed model coordinate. The row’s array index identifies logits.
- Semantic
Text Slot - Original active-ring source slot. Kind and provenance are independent: a filled model-input row is not an assertion that the token was observed.
- Semantic
Token Provenance - Original token origin projected from the native append-only source ledger. A generated token retains its publication, draw and receipt coordinates; none of these fields independently grants authority or proves publication.
- Semantic
Training Canary - Frozen bounds and resource ceilings for one candidate-update canary.
- Semantic
Training Canary Record - Semantic
Training Canary Refusal Record - Device-authored evidence for the highest-precedence failed update canary.
- Semantic
Training Domain - Prospective training domain fixed before any action, not a selected roster or a numerical certificate. The application supplies the original sealed manifests and complete loader geometry; native admission retains them.
- Semantic
Training Manifest - Exact sealed input identity and its four admission caps, in source order: entries, total keys, records per input, bytes per record.
- Semantic
Training Objective - Complete frozen objective carried by the canonical replay-roster owner.
- Semantic
Training Objective Group - One nonempty reduction group in the frozen training objective.
- Semantic
Training Objective Group Record - Semantic
Training Objective Record - Semantic
Training Roster Row - Device-written metadata for every row in the selected objective roster.
- Semantic
Training View Layout - Validated geometry of the bytes consumed by native training-view selection.
- Semantic
Training View Origin - Authenticated native execution that produced one episode training view. Corpus anchors have no execution origin and use their independent graph.
- Semantic
Training View Origin Record - Semantic
Training View Row - One already-admitted training view retained for cold device selection.
- Semantic
Training View Selection - Device-written selection coordinates. Status zero is the only admissible result; downstream Update work must predicate on this resident word.
- Semantic
Transition Host IoStats - Provider counters and explicit session stream waits. Compare snapshots around launch only: cold binding/capture and terminal observation are outside it.
- Semantic
Transition Lane - Device terminal and both ordered edit results. A refused lane has no root; successful edit receipts preceding rollback are not published graph state.
- Semantic
Transition Observation - Semantic
Transition Receipt - Device-produced receipt. Integers use three little-endian 64-bit limbs. The exact probability is p/q; the exact importance factor is p/factor_denominator. Singleton rows canonically use 1/1 for both.
- Semantic
Transition Session - Single owner of one serial transition stream, graph, scratch, and receipt bank. Retains the provider independently of the caller’s handle.
- Semantic
Transition Work - Work performed by the two semantic edits of one learned candidate. These are device counters, not estimates from the sampled actions or final truth alone.
- Semantic
Typed Record - Original typed content. Qualifier indices address this admission’s records.
- Semantic
Version Handle - Semantic
Version Identity - Content identity of an immutable statement version.
- Semantic
Version Ref - Immutable statement-version identity and device-derived truth.
Enums§
- Model
Work Kind - Semantic
Action Descriptor - Immutable descriptor meanings derived from the graph’s admitted records. Category zero is NULL. Nonzero categories are sorted by logical typed bytes, independent of arena slots and digests. Equal-valued support records retain separate categories, ordered by their original admission index.
- Semantic
Argument - One typed argument in a canonical statement identity.
- Semantic
Handle Kind - Kind of opaque semantic handle used in typed diagnostics.
- Semantic
Hypergraph Error - Typed failures from semantic owner and device lifecycle validation.
- Semantic
Insert Outcome - Result of exact support insertion.
- Semantic
Learning Copy Reset - One operation on an explicitly identified typed view. Full backing allocations, including padding, are copied first. No operation may silently change another view, even when the producer deliberately shares its physical storage.
- Semantic
Learning Phase - Semantic
Polarity - Polarity of one exact support event.
- Semantic
Prepared Step Outcome - Results retain each actual device invocation. An early admission refusal owns no numerical invocation, selected receipts, or derivative authority. Intermediate root records are historical receipts; a later step may retire their slots. Acquire the canonical final reader for the current publication.
- Semantic
Record Role - Schema-declared role of a retained typed record.
- Semantic
State Role - Semantic role of a retained state range. Indices name actual layers, tensors, or records inside the cold-bound role roster, never arbitrary device pointers.
- Semantic
Task Refusal - A task-local refusal is not a semantic infrastructure failure.
- Semantic
Training Canary Kind - Mandatory acceptance gate evaluated from original full-vocabulary logits.
- Semantic
Training Canary Refusal Reason - Closed device reason for refusing a candidate model update canary.
- Semantic
Training Objective Group Kind - Frozen evaluator term attached to the complete device-resident training roster.
- Semantic
Training View Basis - Origin of one authentic replay training view.
- Semantic
Training View Port - Fixed device port of the native-selected complete training roster.
- Semantic
Transition Error - Semantic
Transition Kind - Admission cost of the next operation, checked against one acquired parent.
- Semantic
Transition Outcome - Completed device outcome. A refusal carries no successor, selected-score receipts, or derivative authority; it is not a successful observation.
- Semantic
Transition Refusal - Ordinary device refusal after authentic parent and cleanup reconciliation.
- Semantic
Truth - The four-valued semantic state derived from reachable support events.
- Semantic
View - A sealed root or one live fork used for device observation.
Constants§
- SEMANTIC_
FEEDBACK_ SUPPORT_ FIELDS - Named columns of each current feedback support row. These service values do not enter the learned feature schema. Rows follow newest-to-oldest ancestry.
- SEMANTIC_
TRAINING_ CANARY_ EVALUATOR_ ABI - SEMANTIC_
TRANSITION_ COMPONENT_ COUNT - SEMANTIC_
TRANSITION_ GENERATION
Traits§
- Device
Ptr - Abstraction over CudaSlice/CudaView
- Device
PtrMut - Abstraction over CudaSlice/CudaViewMut
- Device
Repr - Something that can be copied to device memory and turned into a parameter for [result::launch_kernel].
- Device
Slice - Base trait for abstracting over CudaSlice/CudaView/CudaViewMut.
- Semantic
Task Program - A native observer executed during cold task binding. Implementations belong to the program executor, not the CUDA substrate. Python accepts executable XLOG source through that executor, never caller-provided answers or receipts. A successful observation must complete all submitted work before returning host results. An error may follow device submission, so the calling session is poisoned and cannot reuse its previous task binding.
- Valid
AsZero Bits - Marker trait to indicate that the type is valid when all of its bits are set to 0.
Functions§
- initial_
source_ layout - Place the exact admitted tokens into an ordered computed prefix and original 32-row ring. No model cache or publication is claimed by this cold plan. The existing parent bind rechecks the same source mapping and cache geometry.
- output_
intents_ from_ publication_ material - pending_
output_ intents_ from_ publication_ material - Resolve the current undelivered set from both sealed native queues. The trusted receiver key authenticates every persisted acknowledgement before any intent is omitted from a restored delivery retry.
- semantic_
structural_ cost_ descriptor - The only structural-cost law accepted by a cold training objective.
- validate_
training_ view_ layout - Check the original v2 byte layout before either cold admission or device allocation.