It seems like it’s become a tradition that I announce I have joined a new company for the Java Advent of Code. At least this time it’s actually an old friend: I am excited to be back at Red Hat, in the llm-d team! Does that mean I forgot about Java? Of course not. If anything, this is an opportunity to learn more about GPU programming, and since Java is my comfort-zone language, what better occasion than looking into TornadoVM?

Recently, the TornadoVM team released gpullama3, a proof-of-concept demonstrating LLM inference on GPUs using pure Java. Let’s explore this together!

What is TornadoVM?

TornadoVM is a plugin for the OpenJDK that enables Java programs to automatically run on heterogeneous hardware (GPUs, FPGAs, and multi-core CPUs) using standard Java code annotated for parallel compute.

Under the hood, TornadoVM:

1. takes your Java bytecode
2. compiles it to GPU-specific kernels (via OpenCL, PTX, or SPIR-V)
3. manages data transfers between CPU and GPU memory
4. executes the computation on the GPU
5. returns the results back to your Java program

Transformer-based language models are computationally expensive but highly-parallelizable. At inference time, generating each token requires matrix multiplications across billions of parameters. GPUs excel at these operations because they can perform thousands of computations in parallel. Thus, TornadoVM is the perfect tool for this kind of workload.

Installing TornadoVM

Installing GPULlama3.java was surprisingly straightforward. Make sure you have your favorite flavor of JDK 21 installed. I use sdkman, so I made sure I had Temurin 21 installed:

sdk install java 21.0.9-tem

Then you’ll want to make sure you have installed cmake, a C/C++ toolchain, Python and pip. Now you can clone the repo:

git clone https://github.com/beehive-lab/GPULlama3.java

and follow the instructions on the README; for instance, on macOS/Linux:

# Enter the TornadoVM submodule directory
cd external/tornadovm

# Optional: Create and activate a Python virtual environment if needed
python3 -m venv venv
source ./venv/bin/activate

# Install TornadoVM with a supported JDK 21 and select the backends (--backend opencl,ptx).
# To see the compatible JDKs run: ./bin/tornadovm-installer --listJDKs
# For example, to install with OpenJDK 21 and build the OpenCL backend, run:
./bin/tornadovm-installer --jdk jdk21 --backend opencl

# Source the TornadoVM environment variables
source setvars.sh

You can verify the installation was successful by running one example:

cd tornado-examples
mvn package
cd ..
tornado -cp tornado-examples/target/tornado-examples-1.1.2-dev-e1d2d12.jar uk.ac.manchester.tornado.examples.compute.MatrixVectorRowMajor

Of course, make sure you replace tornado-examples-1.1.2-dev-e1d2d12.jar with the right jar name! Your output should look something like this:

WARNING: Using incubator modules: jdk.incubator.vector
Matrix-Vector Multiplication Benchmark
======================================
Configuration:
- Input dimension (columns): 8192
- Output dimension (rows): 2048
- Local work group size: 32
- Backend: OPENCL
- DP4A benchmarks enabled: false
- Warmup iterations: 140
- Benchmark iterations: 120

Initializing data...
Setting up TornadoVM execution...
Warming up sequential implementation...
Benchmarking sequential implementation...
Warming up parallel implementation...
Benchmarking parallel implementation...
Validating results...
Validation PASSED ✓

Performance Results:
====================
Matrix size: 2048 x 8192
Sequential Implementation:
Average time: 15.892 ms
Min time: 15.673 ms
Max time: 16.795 ms
Performance: 2.11 GFLOP/s
Parallel Implementation (TornadoVM):
Average time: 2.405 ms
Min time: 2.030 ms
Max time: 5.069 ms
Performance: 13.95 GFLOP/s
Pure TornadoVM @Parallel Implementation (TornadoVM):
Average time: 4.931 ms
Min time: 3.745 ms
Max time: 8.357 ms
Performance: 6.81 GFLOP/s
Parallel Implementation FP16 (TornadoVM):
Average time: 1.840 ms
Min time: 1.575 ms
Max time: 3.090 ms
Performance: 18.24 GFLOP/s
Q8 Vectorized:
Average time: 1.746 ms
Min time: 1.459 ms
Max time: 6.097 ms
Performance: 19.22 GFLOP/s

Speedup: KernelContext vs Java 6.61x
Speedup: @Parallel vs Java 3.22x
Speedup: KernelContext vs @Parallel 2.05x
Speedup: Q8 Vectorized vs KernelContext 1.38x
Speedup: Q8 Vectorized vs KernelContext FP16 1.05x

Baby’s First GPU Kernel

Here’s a simple Example.java:

import uk.ac.manchester.tornado.api.annotations.*;
import uk.ac.manchester.tornado.api.*;
import uk.ac.manchester.tornado.api.enums.DataTransferMode;
import uk.ac.manchester.tornado.api.types.arrays.FloatArray;

public class Example {

public static void vectorMul(FloatArray a, FloatArray b, FloatArray result) {
for (@Parallel int i = 0; i < result.getSize(); i++) {
result.set(i, a.get(i) * b.get(i));
}
}

public static void main(String... args) {
int size = 1024;

var a = new FloatArray(size);
var b = new FloatArray(size);
var result = new FloatArray(size);

for (int i = 0; i < size; i++) {
a.set(i, i * 2.0f);
b.set(i, i + 1.0f);
}

var taskGraph = new TaskGraph("multiply")
.transferToDevice(DataTransferMode.FIRST_EXECUTION, a, b)
.task("vectorMul", Example::vectorMul, a, b, result)
.transferToHost(DataTransferMode.EVERY_EXECUTION, result);

var snapshot = taskGraph.snapshot();
new TornadoExecutionPlan(snapshot).execute();

System.out.println("\nVector Multiplication Results (first 10):");
for (int i = 0; i < 10; i++) {
System.out.printf("result[%d] = %.2f (a[%d]=%.2f * b[%d]=%.2f)\n",
i, result.get(i), i, a.get(i), i, b.get(i));
}
}
}

The @Parallel annotation tells TornadoVM this loop can be parallelized. The TaskGraph API manages data movement and execution scheduling. You can compile it with the following (if you followed the installation guide correctly $TORNADO_SDK will point to the right path):

javac -g --enable-preview -source 21 -cp "$TORNADO_SDK/share/java/tornado/*" Example.java

Notice that -g is required for this to work correctly. Now you can run it with:

tornado Example

It will print the first 10 items in the resulting vector.

Playing with GPULlama3.java

The gpullama3 project demonstrates running a Llama 3 model entirely in Java with GPU acceleration. Assuming you are back at the root of the repo, continue with the setup procedure.

# Navigate back to the project root directory
cd ../../

# Source the project-specific environment paths -> this will ensure the correct paths are set for the project and the TornadoVM SDK
# Expect to see: [INFO] Environment configured for Llama3 with TornadoVM at: /home/YOUR_PATH_TO_TORNADOVM
source set_paths

# Build the project using Maven (skip tests for faster build)
# mvn clean package -DskipTests or just make
make

Now let’s download a compatible model using the HuggingFace CLI:

# download and install the hugging face CLI
pip install -U huggingface_hub

# download a model to ./models/
hf download beehive-lab/Llama-3.2-1B-Instruct-GGUF-FP16 --include '*.gguf' --local-dir models

Try it! Even on my poor MacBook Air with 8 GB RAM (provided I don’t have too many applications open) this returns:

❯ python llama-tornado --gpu --verbose-init --opencl --model models/Llama-3.2-1B-Instruct-FP16.gguf --prompt "tell me a joke"
WARNING: Using incubator modules: jdk.incubator.vector
Loading model weights in TornadoVM format (loading F16)

Starting TornadoVM initialization...
TornadoVM GPU execution plan creation: 1011.56 ms
Java to GPU JIT compiler warmup: 4994.25 ms
Transfer read-only weights to GPU: 13958.97 ms
Finished TornadoVM initialization...

Here's one:

What do you call a fake noodle?

(wait for it...)

An impasta!

Hope that made you laugh!


achieved tok/s: 3.00. Tokens: 42, seconds: 13.98

Disclaimer: even if you have better CPU/GPUs at your disposal, they are unlikely to affect the quality of the joke.

What just happened?

GPULlama3.java currently supports a few FP16 (16-bit floating point) and 8-bit quantized models:

• Llama 3.2 (1B) – FP16
• Llama 3.2 (3B) – FP16
• Llama 3 (8B) – FP16
• Mistral (7B) – FP16
• Qwen3 (0.6B) – FP16
• Qwen3 (1.7B) – FP16
• Qwen3 (4B) – FP16
• Qwen3 (8B) – FP16
• Phi-3-mini-4k – FP16
• Qwen2.5 (0.5B)
• Qwen2.5 (1.5B)
• DeepSeek-R1-Distill-Qwen (1.5B)

Depending on the model being selected, a different execution plan will be built. The execution plan corresponds to the model architecture. In our case, we picked the unquantized Llama 3.2 1B FP16. Let’s take a look at the setupTornadoForwardPlan() method in FP16LayerPlanner, used by LLama 3.2:

abstract class FP16LayerPlanner ... {
...
protected final void setupTornadoForwardPlan() {
List<ImmutableTaskGraph> allTaskGraphs = new ArrayList<>();
GridScheduler masterScheduler = new GridScheduler();

// 1. Activation layer (common to all models)
allTaskGraphs.add(activationLayer.getImmutableTaskGraph());
activationLayer.updateGridScheduler(masterScheduler);

// 2. FFN layers (N transformer layers - model-specific)
allTaskGraphs.addAll(ffnLayers.getFfnLayerTaskGraphs());
ffnLayers.updateGridScheduler(masterScheduler);

// 3. Logits layer (common to all models)
allTaskGraphs.add(logitsLayer.getTaskGraph().snapshot());
logitsLayer.updateGridScheduler(masterScheduler);

// Cache for future retrievals
this.immutableTaskGraphs = allTaskGraphs;
this.gridScheduler = masterScheduler;
}

}

In the Activation layer we mostly look up token embeddings and apply an initial normalization step, while the Logit layer is where we convert the model’s internal representation into token predictions. So let’s concentrate a bit more on the Feed-Forward Network layer (FFN), and in particular on the Attention implementation. The LlamaFP16FFNLayers#setupSingleFFNLayer method is a bit cryptic at a first glance; let’st start from its signature:

TaskGraph setupSingleFFNLayer(LlamaTornadoWeights weights, Configuration config, int layerIndex)

The method is building a TaskGraph, essentially describing the data flow of our GPU kernels. Let’s focus on QKV and attention, using Sebastian Raschka1‘s excellent Python Notebook as a reference. The following is the architecture diagram of the Llama 3.2 1B model:

For obvious reasons of brevity, we aren’t going to explore this in detail, but we do want to take a look at the implementation of the attention heads. In particular, let’s take a look at how we compute the Query, Key, Value matrices (Q,K,V = project(x) in the Python version):

.task("qmatmul",
TransformerComputeKernelsLayered::matrixVectorGeneric, context,
state.wrapXb, state.wrapQ,
weights.wqLayered[layerIndex].asHalfFloatArray(),
config.dim(), config.dim(),
LOCAL_WORK_GROUP_SIZE_ALLOC)
.task("kmatmul",
TransformerComputeKernelsLayered::matrixVectorGeneric, context,
state.wrapXb, state.wrapK,
weights.wkLayered[layerIndex].asHalfFloatArray(),
config.dim(), config.kvDim(),
LOCAL_WORK_GROUP_SIZE_ALLOC)
.task("vmatmul",
TransformerComputeKernelsLayered::matrixVectorGeneric, context,
state.wrapXb, state.wrapV,
weights.wvLayered[layerIndex].asHalfFloatArray(),
config.dim(), config.kvDim(),
LOCAL_WORK_GROUP_SIZE_ALLOC)

This is followed by the RoPE rescaling to encode token positions:

.task("rope", TransformerComputeKernelsLayered::ropeRotation, context,
state.positionHolder,
state.wrapQ, state.wrapK,
config.kvDim(),
config.headSize())

Now we are ready to compute **attention. **The generic version (there is also an NVidia-specific implementation) is:

unifiedLayer.task("parallel-attention", TransformerComputeKernelsLayered::processHeadsParallel,
state.wrapQ, state.wrapKeyCache, state.wrapValueCache, state.wrapXb,
config.numberOfHeads(), config.headSize(), config.kvDim(), config.kvMul(),
config.contextLength(),
state.positionHolder, state.wrapAtt,
layerIndex, config.contextLength());

Let’s drill down into TransformerComputeKernelsLayered::processHeadsParallel to see how that is performed. The following is one of the GPU kernels. It essentially computes:

You will notice that the method:

1. takes the Q,K (Query, Key) vectors that we computed earlier and it computes the attention score up to the current position pos (scaled by the square root of the head size), filling the wrapAtt vector
2. next, it applies softmax to the wrapAtt, turning the scores into attention weights (steps 2-4), accumulating partial results onto the same wrapAtt vector
3. finally (step 5) it computes a weighted sum of the **Value vectors **(value_cache) up to pos, using the calculated **Attention Weights **(wrapAtt).

/**
* Computes attention for a single head. Implements scaled dot-product attention with softmax normalization.
*
* Steps: 1. Compute attention scores: Q·K / sqrt(head_size) 2. Apply softmax (with max subtraction for numerical stability) 3. Compute weighted sum of values
*
* @param allQ
*         All query vectors
* @param key_cache
*         Cached keys
* @param value_cache
*         Cached values
* @param allXb
*         Output buffer
* @param h
*         Head index to process
* @param headSize
*         Dimension per head
* @param kvDim
*         Key/value dimension
* @param kvMul
*         Key multiplier for grouped attention
* @param loff
*         Layer offset in cache
* @param pos
*         Current position
* @param wrapAtt
*         Attention weights buffer
*/
private static void processHeadTornado(FloatArray allQ, FloatArray key_cache, FloatArray value_cache, FloatArray allXb, int h, int headSize, int kvDim, int kvMul, long loff, int pos,
FloatArray wrapAtt) {

// Base index for this head's attention weights
int headOffset = h * (pos + 1);

// STEP 1: Calculate attention scores for all timesteps
for (int t = 0; t <= pos; t++) {
int kvHeadIdx = h / kvMul;
int keyOffset = (int) (loff + t * kvDim + kvHeadIdx * headSize);

float score = 0.0f;
for (int i = 0; i < headSize; i++) {
score += allQ.get(h * headSize + i) * key_cache.get(keyOffset + i);
}
score = score / TornadoMath.sqrt(headSize);

// Store in attention buffer
wrapAtt.set(headOffset + t, score);
}

// STEP 2: Find max score for softmax stability
float maxScore = wrapAtt.get(headOffset);
for (int t = 1; t <= pos; t++) {
float val = wrapAtt.get(headOffset + t);
if (val > maxScore) {
maxScore = val;
}
}

// STEP 3: Compute exponentials and sum
float sum = 0.0f;
for (int t = 0; t <= pos; t++) {
int idx = headOffset + t;
float expScore = TornadoMath.exp(wrapAtt.get(idx) - maxScore);
wrapAtt.set(idx, expScore);
sum += expScore;
}

// STEP 4: Normalize
float normFactor = (sum > 0.0f) ? (1.0f / sum) : (1.0f / (pos + 1));
for (int t = 0; t <= pos; t++) {
int idx = headOffset + t;
wrapAtt.set(idx, wrapAtt.get(idx) * normFactor);
}

// STEP 5: Compute weighted sum of values for each dimension
for (int i = 0; i < headSize; i++) {
float weightedSum = 0.0f;
for (int t = 0; t <= pos; t++) {
int kvHeadIdx = h / kvMul;
int valueOffset = (int) (loff + t * kvDim + kvHeadIdx * headSize);
weightedSum += wrapAtt.get(headOffset + t) * value_cache.get(valueOffset + i);
}
allXb.set(h * headSize + i, weightedSum);
}
}

After the attention mechanism computes relationships between tokens, the result is added to the original input, normalized, and passed through a feed-forward network. This process repeats across multiple layers before finally producing the next-token prediction (the logit layer).

Because it’s an autoregressive model, this entire process repeats for each token, using the previously generated sequence as input.

In short, TornadoVM handled GPU compilation and execution transparently, allowing a pure Java program to perform LLM inference!

Conclusions

We’ve completed our whirlwind tour of Llama3GPU.java and TornadoVM. If your head is still spinning, don’t worry, you’re not alone! It’s a lot to take in, but I hope this post has sparked your interest and inspired you to dig deeper: I know I will!

1. Sebastian Raschka is the author of Build a Large Language Model from Scratch ↩︎

Author: Edoardo Vacchi

After my PhD at University of Milan on programming language design and implementation, I worked for three years at UniCredit Bank’s R&D department. Later, I have joined Red Hat where I worked on the Drools rule engine, the jBPM workflow engine and the Kogito cloud-native business automation platform. I joined Tetrate to work on the wazero WebAssembly runtime for Go. At Dylibso I contributed to wazero and Chicory Wasm runtime for the JVM, and other runtimes! I am currently back at Red Hat AI. I sometimes write on my own personal blog.