GPULlama3.java 0.3.3

---

📦 Installation

Maven

<dependency>
    <groupId>io.github.beehive-lab</groupId>
    <artifactId>gpu-llama3</artifactId>
    <version>0.3.3</version>
</dependency>

Gradle

implementation 'io.github.beehive-lab:gpu-llama3:0.3.3'

---

📖 Documentation | 🔗 Maven Central

GPULlama3.java 0.3.2

Model Support

• [models] Support for IBM Granite Models 3.2, 3.3 & 4.0 with FP16 and Q8 (#92)

Other Changes

• [docs] Update docs to use SDKMAN! and point to TornadoVM 2.2.0 (#93)
• Add JBang catalog and local usage examples to README.md (#91)
• Add jbang script and configuration to make easy to run (#90)

---

📦 Installation

Maven

<dependency>
    <groupId>io.github.beehive-lab</groupId>
    <artifactId>gpu-llama3</artifactId>
    <version>0.3.2</version>
</dependency>

Gradle

implementation 'io.github.beehive-lab:gpu-llama3:0.3.2'

---

📖 Documentation | 🔗 Maven Central

GPULlama3.java 0.3.1

Model Support

• Add compatibility method for langchain4j and quarkus in ModelLoader (#87)

---

📦 Installation

Maven

<dependency>
    <groupId>io.github.beehive-lab</groupId>
    <artifactId>gpu-llama3</artifactId>
    <version>0.3.1</version>
</dependency>

Gradle

implementation 'io.github.beehive-lab:gpu-llama3:0.3.1'

---

📖 Documentation | 🔗 Maven Central

GPULlama3.java 0.3.0

Model Support

• [refactor] Generalize the design of tornadovm package to support multiple new models and types for GPU exec (#62)
• Refactor/cleanup model loaders (#58)
• Add Support for Q8_0 Models (#59)

Bug Fixes

• [fix] Normalization compute step for non-nvidia hardware (#84)

Other Changes

• Update README to enhance TornadoVM performance section and clarify GP… (#85)
• Simplify installation by replacing TornadoVM submodule with pre-built SDK (#82)
• [FP16] Improved performance by fusing dequantize with compute in kernels: 20-30% Inference Speedup (#78)
• [cicd] Prevent workflows from running on forks (#83)
• [CI][packaging] Automate process of deploying a new release with Github actions (#81)
• [Opt] Manipulation of Q8_0 tensors with Tornado ByteArrays (#79)
• Optimization in Q8_0 loading (#74)
• [opt] GGUF Load Optimization for tensors in TornadoVM layout (#71)
• Add SchedulerType support to all TornadoVM layer planners and layer… (#66)
• Weight Abstractions (#65)
• Bug fixes in sizes and names of GridScheduler (#64)
• Add Maven wrapper support (#56)
• Add changes used in Devoxx Demo (#54)

---

📦 Installation

Maven

<dependency>
    <groupId>io.github.beehive-lab</groupId>
    <artifactId>gpu-llama3</artifactId>
    <version>0.3.0</version>
</dependency>

Gradle

implementation 'io.github.beehive-lab:gpu-llama3:0.3.0'

---

📖 Documentation | 🔗 Maven Central

v0.2.2

What's Changed

• Fully working support for LangChain4j
• LlamaApp cleanup by @orionpapadakis in #51
• Fix execution path control by @orionpapadakis in #52
• Add support for encoding ordinary text in Qwen3Tokenizer and update Q… by @mikepapadim in #53

Full Changelog: v0.2.1...v0.2.2

v0.2.1

What's Changed

• Minor cleanup by @mikepapadim in #47
• Add useTornadovm flag to model loader to handle Builder option in Langchain4j by @mikepapadim in #50

Full Changelog: v0.2.0...v0.2.1

v0.2.0

Model Support

• Mistral – support for GGUF-format Mistral models with optimized GPU execution.
• Qwen2.5 – GGUF-format Qwen2.5 models supported, including performance improvements for attention layers.
• Qwen3 – compatible with GGUF-format Qwen3 models and updated integration.
• DeepSeek-R1-Distill-Qwen-1.5B – GGUF-format DeepSeek distilled models supported for efficient inference.
• Phi-3 – full support for GGUF-format Microsoft Phi-3 models for high-performance workloads.

What's Changed

• [refactor] Renamed aux package to resolve Windows issue by @stratika in #11
• Windows support for GPULlama3.java by @stratika in #12
• [API] Update TornadoVM API to use latest warmup features by @mikepapadim in #13
• [model] Add support for Mistral models by @orionpapadakis in #17
• Cleanups post Mistral Integration by @mikepapadim in #27
• Add a Docker section to README with available images and usage examples by @mikepapadim in #28
• Refactor TornadoVMMasterPlan to simplify scheduling decision for non-Nvidia HW and Mistral Models by @mikepapadim in #32
• File not found error handling in loadModel method in GGUF.java by @dhruvarayasam in #34
• Update README for clarity by @mikepapadim in #36
• [models] Support for Qwen3 models by @orionpapadakis in #37
• [models][phi-3] Support for Microsoft's Phi-3 models by @mikepapadim in #38
• Reorganize package structure and update imports to use `org.beehive.g… by @mikepapadim in #42
• Update README.md by @kotselidis in #44
• [models][deepseek][qwen2.5] Add support for Qwen2.5 and Deepseek-Distilled-Qwen models by @orionpapadakis in #40
• Improve attention performance for qwen2.5 & deepseek by @orionpapadakis in #46

New Contributors

• @orionpapadakis made their first contribution in #17
• @dhruvarayasam made their first contribution in #34

Full Changelog: v0.1.0-beta...v0.2.0

v0.1.0-beta

• Llama 3 model compatibility - Full support for Llama 3.0, 3.1, and 3.2 models
• GGUF format support - Native handling of GGUF model files
• Support for FP16 models for reduced memory usage and faster computation
• GPU Acceleration on NVIDIA GPUs using both OpenCL and PTX backends
• [Experimental] Support for Apple Silicon (M1/M2/M3) via OpenCL (subject to hardware/compiler limitations)
• [Experimental] Initial support for Q8 and Q4 quantized models, using runtime dequantization to FP16