Repository navigation
Prepare SKaiNET for TurboQuant with refactoring #451
Description
SKaiNET is close to TurboQuant-readiness, but core refactoring is needed first.
- separate logical dtype from physical encoding/layout
- introduce buffer/storage abstractions
- add explicit placement and memory-domain modeling
- reduce copy-heavy materialization/loading paths
- unify packed quantized tensor storage
Activity
Use borrowed arrays in SafeTensors loader, add wrap methods to Execut…
Add loadTensorStorage() to streaming GGUF and SafeTensors readers
Add file-backed tensor storage loading to streaming readers
Expose MemoryPlanner and MemoryTracker through ExecutionContext
Migrate Ternary2BitTensorData to PackedBlockStorage interface
Add FallbackMappedMemoryChunk for non-JVM platforms
Add explicit transfer operations to TensorStorage
Auto-instrument copy paths with ActiveMemoryTracker
Add @place and @Weights DSL annotations for placement intent
Update README and docs to recommend StreamingGGUFReader
Add BufferAccessor and JvmFileBackedResolver for mmap end-to-end
Add end-to-end storage integration tests with synthetic GGUF
Add TensorStorage → TensorData bridge for backend compatibility
Add unit tests for transfer ops, Q4_K and Ternary dequantization
Add tests for ActiveMemoryTracker, FallbackMappedMemoryChunk, non-con…
michalharakal commentedon Apr 8, 2026
Implemented with #458