Summary

PR 14 of 22 PRs in the ag_new_release_updates_2 branch split. Merge after Layer 3 (native platform backends + DSP engine).

• CudaDeviceContextProvider: New single canonical path for device-switching; replaces 15+ duplicated sequences (update affinity map → nativeOps.setDevice() → reset CudaZeroHandler ThreadLocal → fetch stream pointers) across CudaAffinityManager, AtomicAllocator, CudaZeroHandler, and multiple executioner paths
• Coherence-aware data buffers: BaseCudaDataBuffer tracks CoherenceState per buffer; syncToHost()/syncToDevice() are no-ops when state is already valid; all 14 typed CUDA buffer subclasses updated
• AtomicAllocator integration: allocateMemory() now calls DeviceMemoryManager.trackAllocation() before native allocation; enables cap enforcement and eviction callbacks without polling
• 4 new backends scaffolded: TPU (PJRT/Cloud TPU v4/v5, BF16 default, HLO compilation caching, priority=50), Hexagon (SNPE/QNN dispatch), ZLUDA (AMD GPUs via HIP→CUDA translation, reuses existing JCublas backend), SDX (Snapdragon X Elite NPU, early-stage scaffold)
• GraalVM Native Image support: jni-config.json, reflect-config.json, resource-config.json, native-image.properties added to CPU backend under META-INF/native-image/
• CPU backend: Updated for DeviceContextProvider delegation, DeviceMemoryManager registration, OpaqueDataBufferDeallocator, MultiBackendWorkspace interface
• AllocationPoint: Added getCoherenceState()/setCoherenceState() methods; deallocation unified via OpaqueNDArrayDeallocator

What Changed

CPU Backend — nd4j-cpu-backend-common (18 files)

• CpuAffinityManager.java — delegates thread affinity to CpuDeviceContextProvider
• CpuMemoryManager.java — registers with DeviceMemoryManager; honors memory caps
• CpuNDArrayFactory.java — routes through DeviceAwareNDArrayFactory
• BaseCpuDataBuffer.java — coherence state tracking; OpaqueDataBufferDeallocator
• NativeOpExecutioner.java — integrates DeviceAwareOpExecutioner, BackendRoutingStrategy, HelperRouter; DSP plan path updated
• CpuWorkspace.java — implements MultiBackendWorkspace interface
• CpuTADManager.java, DirectShapeInfoProvider.java, CpuLapack.java, CpuLevel1.java, DefaultDataBufferFactory.java, CpuDeallocator.java, ConstantBuffersCache.java, CpuOpContext.java, CpuOpContextDeallocator.java, CpuNativeRandom.java — API alignment and coherence/deallocator updates
• module-info.java — Java 9 module descriptor

CUDA Backend — nd4j-cuda (40+ files)

• CudaDeviceContextProvider.java — single canonical device-switch path: updates affinity map, traces via MultiGpuTracer, calls nativeOps.setDevice() (triggers C++ ContextBuffers::release() + lazy reinit), resets CudaZeroHandler ThreadLocal, fetches stream pointers via buildContext(); replaces 15+ duplicated call sites
• AllocationPoint.java — getCoherenceState()/setCoherenceState() methods
• AtomicAllocator.java — allocateMemory() and free() update DeviceMemoryManager atomics for cap enforcement
• CudaAffinityManager.java — now delegates to CudaDeviceContextProvider; retains only high-level API
• BaseCudaDataBuffer.java — coherence state per buffer; syncToHost()/syncToDevice() guarded by CoherenceState; OpaqueDataBufferDeallocator
• CudaFloatDataBuffer.java, CudaHalfDataBuffer.java, CudaBfloat16DataBuffer.java, CudaDoubleDataBuffer.java, CudaIntDataBuffer.java, CudaLongDataBuffer.java, CudaByteDataBuffer.java, CudaBoolDataBuffer.java, CudaShortDataBuffer.java, CudaUByteDataBuffer.java, CudaUInt16DataBuffer.java, CudaUInt32DataBuffer.java, CudaUInt64DataBuffer.java, CudaUtf8Buffer.java — all updated for coherence tracking and new deallocator chain
• CudaDataBufferFactory.java — produces coherence-aware buffers
• MemoryHandler.java, CudaZeroHandler.java — reset context via CudaDeviceContextProvider on device switch
• CudaMemoryManager.java — registers allocations in DeviceMemoryManager; checks device caps before allocating
• SynchronousFlowController.java — uses tl_dspGapStream for gap op stream synchronization
• CudaWorkspace.java — implements MultiBackendWorkspace with coherence-aware buffer management
• CudaExecutioner.java — integrates DeviceAwareOpExecutioner, MultiBackendExecutioner; DSP execution path updated
• CudaEnvironment.java, JCublasBackend.java — updated for Environment subsystem config classes
• JCublasNDArray.java, JCublasNDArrayFactory.java — coherence tracking; device-routed factory
• JcublasLapack.java, JcublasLevel1.java — routed through DeviceAwareOpExecutioner
• CudaDeallocator.java, CudaContext.java, CudaOpContext.java, CudaOpContextDeallocator.java — updated deallocator chain and device context propagation
• Configuration.java — extended with multi-device routing, FP8 support, Triton section fusion properties
• BasicTADManager.java, DeviceTADManager.java — updated for device descriptor abstraction
• Nd4jCudaPresets.java — updated preset annotations for FP8/Triton/paged attention entry points
• module-info.java, nd4j-jcublas.properties — module exports and properties updated

CUDA Platform and Presets

• nd4j-cuda-platform/pom.xml, nd4j-cuda-preset/pom.xml — updated platform BOM and preset build
• Nd4jMinimalPresets.java, Nd4jMinimal.java, CpuStatisticsProvider.java — updated minimizer backend

New: TPU Backend — nd4j-tpu (7 files)

• JTpuBackend.java — PJRT (Portable Runtime) for Cloud TPU v4/v5; auto-detects TPU via JNI probe; priority=50
• JTpuNDArray.java — NDArray backed by XLA buffer handles
• JTpuNDArrayFactory.java — produces JTpuNDArray instances
• TpuEnvironment.java — HLO compilation caching, BF16 defaults
• TpuExecutioner.java — routes ops to PJRT XLA execution
• TpuOpContext.java, nd4j-jtpu.properties

New: TPU Preset — nd4j-tpu-preset (2 files)

• Nd4jTpuPresets.java / Nd4jTpuHelper.java — JavaCPP preset and helper for PJRT bindings

New: Hexagon Backend — nd4j-hexagon (6 files)

• HexagonBackend.java, HexagonEnvironment.java, HexagonExecutioner.java, HexagonOpContext.java — Qualcomm Hexagon DSP via SNPE/QNN
• META-INF/services/org.nd4j.linalg.factory.Nd4jBackend — service registration
• nd4j-hexagon-preset/Nd4jHexagonPresets.java — JavaCPP preset for SNPE/QNN bindings

New: ZLUDA Backend — nd4j-zluda (4 files)

• JZludaBackend.java — CUDA-compatible backend via ZLUDA HIP→CUDA translation for AMD GPUs; reuses JCublas backend without Java-layer changes
• ZludaEnvironment.java, module-info.java, nd4j-jzluda.properties

New: SDX Backend — nd4j-sdx (1 file)

• SdxRuntime.java — Snapdragon X Elite NPU integration (early-stage scaffold)

GraalVM Native Image — nd4j-native/META-INF/native-image/ (4 files)

• jni-config.json, reflect-config.json, resource-config.json, native-image.properties — GraalVM Native Image configuration for CPU backend AOT compilation

Build Infrastructure

• nd4j-backend-impls/pom.xml — adds tpu, hexagon, zluda, sdx modules
• nd4j-presets-common/pom.xml, nd4j-native-platform/pom.xml — updated dependencies

Dependencies

• Depends on: PR12 (provides DeviceContextProvider, DeviceMemoryManager, DeviceDescriptor, CoherenceState, MultiBackendWorkspace interfaces); PR13 (new op impls tested through updated executioners)
• Required by: PR15 (SameDiff session execution depends on updated CudaExecutioner and NativeOpExecutioner), PR16 (DSP runtime uses CudaDeviceContextProvider for stream management)

Merge Order

This PR is in Layer 4 (Java backend implementations — parallel with PR12/PR13/PR15, all needed before PR16).