LogicalTensor
What does this tensor mean?
For researchers
Identity, class, shape, logical operation, orientation, data type, and graph boundary: the semantic contract.
VOLUME I / ARCHITECTURE DEEP DIVE
How Prism turns
model bytes into execution.
Start with the plain-language path. Open the technical layers when you want the compiler contracts, data structures, and runtime boundaries.
● Start here● Researcher detail
The short version
Prism is a compiler and runtime for local AI deployment. A training system produces a model artifact. Prism analyzes it once, makes deployment decisions, seals those decisions into a ComputeImage, and executes that image on available hardware.
MODELGGUF / SafeTensors
→COMPILERadmission · layout · planning
→COMPUTEIMAGEsealed artifact
→RUNTIMEMetal · CPU · ANE
One tensor, three questions.
Prism separates concerns that model formats and runtimes often blend together.
PhysicalTileLayout
How are its bits stored?
For researchers
Codec, tile family, group axis, metadata, padding, alignment, and interleave: the storage and kernel ABI.
ExecutionView
How does a lane consume it?
For researchers
Lane, offsets, metadata ranges, repacking, and residency. Multiple views can derive from one physical layout.
The compiler makes policy explicit.
Prism moves representation admission, layout planning, execution-view selection, residency, scheduling metadata, validation, and artifact sealing into compilation.
01IngestModel identity, metadata, weights, graph.
02AnalyzeTensor classes and candidate behavior.
03AdmitQuality and resource gates.
04PlanLayouts, views, residency, placement.
05SealComputeImage plus evidence.
Compiler researcher detail
GGUF → ModelGraph → TensorAnalysis → CandidateCodecs → Admission → PhysicalLayout → ExecutionViews → PrismIR → PhysicalPlanning → ComputeImage
A deployment artifact,
not a renamed checkpoint.
A ComputeImage is the boundary between compilation and execution. The runtime loads it without reconstructing deployment policy from the original model.
METADATAidentity · version · manifest
LOGICAL TENSORSsemantic tensor table
PHYSICAL LAYOUTScodec and tile contracts
EXECUTION VIEWSlane-specific materializations
PLAN + RECEIPTSdecisions and validation
PAYLOADSmapped tensor bytes
ABI detail
The CImage layout ABI makes logical identity, physical storage, and execution views independently inspectable. A target profile can select resident or mutually exclusive views without changing the logical graph.
Read the CImage ABI ↗The runtime executes decisions.
It does not invent them.
Prism resolves an admitted view, establishes residency, schedules work, dispatches a backend, and records the outcome.
SESSION
↓SCHEDULER
↙ ↘METAL GPU
CPU / ANE
↘ ↙MEMORY + RESIDENCY
↓EXECUTION RECEIPT
Runtime researcher detail
Canonical state owns accepted work, model identity, ownership, leases, committed state, and terminal outcomes. Caches, queues, backend handles, and projections remain derived or ephemeral.
Why not extend an inference runtime?
Existing engines can be excellent execution backends. Prism makes a different architectural choice: representation admission, layout planning, execution views, residency, scheduling metadata, validation, and artifact sealing happen once during compilation.
TYPICAL RUNTIME
Load model → derive strategy dynamically → dispatch
PRISM
Compile policy → seal artifact → execute explicit plan
What you can run today.
Apple Silicon compilation and Metal execution are the primary supported path. Linux CPU is under active hardening. ANE/Core ML and other targets have integration or planning surfaces at differing maturity; AMD, Intel, NVIDIA, Tenstorrent, and distributed execution remain development surfaces.
See current state ↗ Open the field guide ↗