1.0.0-rewrite

Release notes for the Deeplearning4j 1.0.0-rewrite: a transitional release introducing the DSP execution engine, LLM/VLM stack, GGUF import, PEFT/RL alignment, and massive hardware expansion. Namespac

This is a transitional release. The 1.0.0-rewrite introduces the full new feature set (DSP, LLM/VLM, PEFT, hardware backends) while maintaining backward compatibility with existing org.deeplearning4j and org.nd4j imports. A namespace consolidation is underway — the next release will complete the import cleanup, unifying packages under a consistent structure. See Namespace Consolidation below for what to expect.

Highlights

The 22 PRs (#10435–#10456) introduce capabilities that reposition DL4J as a first-class runtime for modern LLM and VLM workloads. Each feature area has its own full documentation page — the links below take you to comprehensive guides with API references, code examples, and configuration details.

Feature	Docs	Scale
DSP Execution Engine	Compiled graph runtime with CUDA graph capture/replay, Triton/NVRTC/PTX JIT, 26-pass graph optimizer, and multi-backend dispatch	~240 new files
LLM & VLM Stack	Generation pipelines, paged KV cache, speculative decoding, continuous batching, tokenizers, eval framework, model editing	456 new files
GGML/GGUF Import	Import quantized LLMs from .gguf files with 20+ architecture handlers, complete quantization codecs, round-trip export	121 new files
PEFT & RL Alignment	12 PEFT methods, 9 RL alignment trainers, FP8 mixed precision, 8-bit Adam, knowledge distillation, dataset curation	~80 new files
ONNX Import & Export	~120 new ONNX ops including Microsoft LLM contrib, bidirectional SameDiff-to-ONNX export	~130 files
New Operations (~130)	Fused attention, KV cache, PEFT linear layers, normalization, positional encoding, quantization, SSM/recurrent, MoE, audio/signal	~80 new files
Hardware Backends	TPU (PJRT), Hexagon DSP (QNN), ZLUDA (AMD/Intel), SDX, ARM ACL (124 ops), Apple Accelerate (28 ops), cuDNN expansion, llama.cpp, MPS, MLIR	382+ files
DSP Runtime SDK	Stable C ABI with Java, Kotlin, Python, Rust, Swift, C# bindings	See DSP docs
Build System & CI	CUDA 12.9, 21 CMake modules, backend namespace isolation, 18 test suites	~60 files

---

Namespace Consolidation

Current State (1.0.0-rewrite)

This release uses three package roots in parallel. All three work and are fully supported:

Root	Where Used	Example
org.nd4j.*	Core ND4J, SameDiff, PEFT, RL trainers, GGML, dataset curation, execution infrastructure	org.nd4j.autodiff.samediff.config.LoraConfig
org.eclipse.deeplearning4j.*	New high-level application modules (LLM, VLM, audio, OmniHub, pipelines, SafeTensors)	org.eclipse.deeplearning4j.llm.GenerationPipeline
org.deeplearning4j.*	Legacy DL4J neural network APIs (MultiLayerNetwork, ComputationGraph, Keras import)	org.deeplearning4j.nn.multilayer.MultiLayerNetwork

This split is intentional for this transitional release — existing code continues to work unchanged.

Next Release: Import Cleanup

The next release will consolidate these into a unified namespace structure. What to expect:

• org.eclipse.deeplearning4j.* becomes the canonical root for all application-level modules

• org.nd4j.* remains for core array/tensor/SameDiff APIs (these are stable)

• org.deeplearning4j.* legacy imports will be deprecated with re-export shims — existing code compiles with deprecation warnings, not errors

• PEFT, RL trainers, and training configs currently under org.nd4j.autodiff.samediff.config.* will move to a more discoverable location

• The org.nd4j.ggml.* package may be consolidated under the pipeline framework

Action for this release: Use whichever imports work. Do not refactor your imports to anticipate the cleanup — the next release will provide clear migration tooling.

---

Examples

The deeplearning4j-examples repository contains 34 example files demonstrating the new features. All examples are in the samediff-examples, onnx-import-examples, and dl4j-examples modules.

DSP Execution Engine (5 examples)

Example	What It Shows
DSPExecutionExample.java	DSP introduction — dynamic-shape inference, GraphExecutionMode, shape-keyed plan caching
DSPAdvancedExample.java	Full DSP API — DspHandle, slot introspection, Chrome trace export, tensor/pipeline parallelism
DSPBackendsAndKernelSelectionExample.java	All 19 GraphExecutionMode backends, KernelSelectionConfig, 24 optimization passes, TPU/Hexagon/ZLUDA/Metal/MLX/NNAPI
DSPDiskCacheAndTritonExample.java	DspPlanDiskCache, Triton kernel cache, TritonCacheTool CLI, plan binary format
DSPDiagnosticsAndDebuggingExample.java	20 diagnostic categories, DspDebugger, DspHandle live introspection, DspPlanAssertions

Path: samediff-examples/src/main/java/org/nd4j/examples/samediff/advanced/execution/

LLM Generation (3 examples)

Example	What It Shows
LLMGenerationPipelineExample.java	GenerationPipeline API, KV cache strategies, speculative decoding, tensor parallelism, VLM integration
QwenTextGenerationExample.java	End-to-end: GGUF download → GGMLModelImport → HuggingFaceTokenizer → GenerationPipeline → chat with ChatTemplate
GraphOptimizerExample.java	GraphOptimizer — algebraic simplification, peephole, strength reduction, CSE, attention fusion

Path: samediff-examples/src/main/java/org/nd4j/examples/samediff/quickstart/modeling/

VLM and Audio (3 examples)

Example	What It Shows
SmolDoclingVLMExample.java	Vision-language model — ONNX components, image tiling, vision encoding, embedding merging, text generation
VideoVLMExample.java	Video VLM — VideoFrameSampler strategies (UNIFORM/FIXED_FPS/KEYFRAME), VideoPreprocessor, SmolVLM2/Qwen3-VL
WhisperSpeechToTextExample.java	Whisper ASR — model download, mel spectrogram, transcription, timestamps

GGML/GGUF Import (4 examples)

Example	What It Shows
GGMLImportExportExample.java	Full GGML API — ConversionOptions, ExportOptions, GGUFReader/GGUFWriter, round-trip quantization
GGMLModelImportExample.java	Low-level import — architecture detection, quantization types Q2_K through IQ2_XS
HuggingFaceGGUFImport.java	HuggingFaceHubDownloader → AutoModel → format auto-detection (GGUF/SafeTensors/ONNX/PyTorch)
SafeTensorsImportExample.java	3-level SafeTensors API — AutoModel, SafeTensorsPipelineLoader, SafeTensorsReader

Path: onnx-import-examples/src/main/java/org/deeplearning4j/modelimportexamples/omnihub/

PEFT (3 examples)

Example	What It Shows
SFTLoRATrainingConfigExample.java	LoraConfig, QLoraConfig, SFTConfig, GRPOConfig, DPOConfig, BF16 mixed precision
AdvancedPEFTConfigExample.java	AdaLoraConfig, DoraConfig, IA3Config, PrefixTuningConfig, PromptTuningConfig, PPOConfig, KTOConfig, ORPOConfig, SimPOConfig
SpecializedPEFTConfigExample.java	LoftQConfig, LohaConfig, LokrConfig, VeraConfig, DyLoraConfig, AdapterConfig, LoraAdapterCache (multi-adapter serving)

Path: samediff-examples/src/main/java/org/nd4j/examples/samediff/quickstart/training/

RL Alignment (1 example)

Example	What It Shows
RLAlignmentConfigExample.java	All 10 RL methods: DPO (standard/IPO/RDPO), GRPO, PPO, KTO, ORPO, SimPO, DAPO (asymmetric clipping), GSPO, Dr.GRPO, RewardModelConfig, RLPipelineConfig

Training Infrastructure (5 examples)

Example	What It Shows
MixedPrecisionTrainingExample.java	FP16/BF16/FP8 mixed precision, LossScaleConfig, GradientAccumulator, FP8TrainingConfig
DataCurationPipelineExample.java	TextDeduplicator, TextQualityFilter, InstructionDataFormatter, StratifiedSplitter, LengthBucketingIterator, WeightedDataMixer
KnowledgeDistillationExample.java	DistillationTrainer — logit KD, feature KD, attention KD, combined
TransferLearningAndFreezingExample.java	PeftModel, variable freezing, GradientCheckpointConfig, ContinuedPretrainingConfig
NewOptimizersExample.java	Adam8bit (4x memory reduction), AdaBelief

New Operations (3 examples)

Example	What It Shows
TransformerOpsExample.java	FlashAttention, SlidingWindowAttention, FusedRoPE/FusedMRoPE, RmsNorm, SiLU, SelectiveScan (Mamba), TokenSample, FP8Matmul, ColumnParallelLinear/RowParallelLinear
SameDiffOpsExample.java	All op namespaces including new sd.audio(), sd.signal(), mixed precision types
AudioOpsExample.java	13 audio DSP ops — mel spectrogram, MFCC, Griffin-Lim, pitch detection, spectral features

Additional Examples

Example	What It Shows
TorchScriptImportExample.java	TorchScriptModelImport — import PyTorch .pt models via nd4j-torchscript
OmniHubPretrainedModels.java	OmniHubUtils, Pretrained, HuggingFaceHubDownloader — model zoo with format auto-detection
TtsTrainingPipelineExample.java	TTS fine-tuning with LoRA — TtsTrainingPipeline, AudioDataProcessor
LRScheduleConfigExample.java	CosineWarmupSchedule + 9 other schedules
KnowledgeDistillationConfigExample.java	DistillationConfig builder variants
TransferLearningConfigExample.java	TransferLearning, FineTuneConfiguration, VariableGroup

---

New Maven Modules

<!-- LLM generation pipeline -->
<dependency>
    <groupId>org.deeplearning4j</groupId>
    <artifactId>samediff-llm</artifactId>
    <version>1.0.0-rewrite</version>
</dependency>

<!-- Vision-language model support -->
<dependency>
    <groupId>org.deeplearning4j</groupId>
    <artifactId>samediff-vlm</artifactId>
    <version>1.0.0-rewrite</version>
</dependency>

<!-- Whisper ASR / audio -->
<dependency>
    <groupId>org.deeplearning4j</groupId>
    <artifactId>samediff-audio</artifactId>
    <version>1.0.0-rewrite</version>
</dependency>

<!-- GGUF model import -->
<dependency>
    <groupId>org.nd4j</groupId>
    <artifactId>nd4j-ggml</artifactId>
    <version>1.0.0-rewrite</version>
</dependency>

<!-- HuggingFace-compatible tokenizers -->
<dependency>
    <groupId>org.nd4j</groupId>
    <artifactId>nd4j-tokenizers</artifactId>
    <version>1.0.0-rewrite</version>
</dependency>

<!-- TorchScript/PyTorch model import -->
<dependency>
    <groupId>org.nd4j</groupId>
    <artifactId>nd4j-torchscript</artifactId>
    <version>1.0.0-rewrite</version>
</dependency>

<!-- Pipeline SPI modules -->
<dependency>
    <groupId>org.deeplearning4j</groupId>
    <artifactId>samediff-pipeline-core</artifactId>
    <version>1.0.0-rewrite</version>
</dependency>
<dependency>
    <groupId>org.deeplearning4j</groupId>
    <artifactId>samediff-pipeline-ggml</artifactId>
    <version>1.0.0-rewrite</version>
</dependency>
<dependency>
    <groupId>org.deeplearning4j</groupId>
    <artifactId>samediff-pipeline-safetensors</artifactId>
    <version>1.0.0-rewrite</version>
</dependency>
<dependency>
    <groupId>org.deeplearning4j</groupId>
    <artifactId>samediff-pipeline-onnx</artifactId>
    <version>1.0.0-rewrite</version>
</dependency>

<!-- DSP Runtime SDK (Java) -->
<dependency>
    <groupId>org.eclipse.deeplearning4j</groupId>
    <artifactId>nd4j-dsp-runtime-java</artifactId>
    <version>1.0.0-rewrite</version>
</dependency>

<!-- Hardware backends (add if hardware is present) -->
<dependency>
    <groupId>org.nd4j</groupId>
    <artifactId>nd4j-tpu</artifactId>          <!-- Google Cloud TPU -->
    <version>1.0.0-rewrite</version>
</dependency>
<dependency>
    <groupId>org.nd4j</groupId>
    <artifactId>nd4j-hexagon</artifactId>       <!-- Qualcomm Hexagon DSP -->
    <version>1.0.0-rewrite</version>
</dependency>

---

Bug Fixes & Improvements

The following issues were resolved in merges since M2.1:

Memory Leaks

• Fix CUDA lstmLayer permute/transpose memory leak (#10404)

• Fix lstmLayer.cu weight transformation memory leak (#10403)

• Fix MmulHelper::mmulNxN memory leak (#10394)

• Fix BaseNDArray.toFlatArray() memory leak for view arrays (#10410)

Correctness

• Fix COORDS2INDEX macro to use strides instead of shapes (#10393)

• Fix DataType inconsistency in float[] constant buffer handling (#10411)

• Fix inverted boolean logic in DeallocatorService listener delegation (#10412)

• Fix byte order handling in DataTypeConversions (#10401)

Security

• Fix command injection vulnerabilities in Windows bat scripts (#10409, #10407)

API & Infrastructure

• Simplify batched GEMM API (#10361)

• JavaCPP resource configuration for native image (GraalVM) support (#10287)

• Autodiff core improvements (#10280)

• CMake modernization (#10245)

• Maven version updates for Java 25 support (#10243)

• SameDiff file format scaling improvements (#10209)

---

Migration Guide

This Is a Transitional Release

1.0.0-rewrite is designed to be additive — existing M2.1 code continues to work. The new features are opt-in via new Maven modules and new API entry points. No existing imports or APIs are removed.

However, be aware that the next release will begin the namespace consolidation. To prepare:

• Avoid deep coupling to internal package paths (e.g., org.nd4j.autodiff.samediff.config.* may move)

• Prefer the high-level entry points (GenerationPipeline, GGMLModelImport, PeftModelFactory) over assembling components manually

• The examples in deeplearning4j-examples show the intended usage patterns

Opting In to DSP Optimization

Graph execution mode defaults to AUTO. To explicitly select a mode:

// Set via system property
-Dnd4j.execution.mode=TRITON

// Or programmatically (restricted to BenchmarkConfigApplier in production)
sd.setGraphExecutionMode(GraphExecutionMode.TRITON);

To skip specific optimizer passes during debugging:

-Dnd4j.optimizer.skip=AttentionFusion
-Dnd4j.optimizer.logApplied=true

New Backend Discovery

Hardware backends (TPU, Hexagon, ZLUDA) are discovered automatically if the corresponding Maven module is on the classpath and the hardware is present. No code changes are required.

Serialization Compatibility

• DspPlanDiskCache stores compiled plans in ~/.kompile/cache/dsp/ using a content-based key. Cache invalidates automatically on model changes.

• New model formats (SDNB, SDZ) are supported alongside the existing FlatBuffers serialization. Both SameDiff.load() and SameDiff.save() continue to work.

• GGUF models imported via GGMLModelImport can be converted to native SDZ format via convertToSDZ() for faster subsequent loads.

Build System

• CUDA 12.9 is now supported alongside 12.6

• Backend namespace isolation (SD_BACKEND_NAMESPACE) enables true multi-backend co-loading in a single JVM process

• 18 named test suites available via run-tests.yml: quick, sanity, nd4j, samediff, dl4jcore, keras, datavec, onnx, tensorflow, integration, libnd4j, llm, vlm, ggml, zoo, longrunning, all