llama3.java

Llama3.java

Practical Llama 3, 3.1 and 3.2 inference implemented in a single Java file.

This project is the successor of llama2.java based on llama2.c by Andrej Karpathy and his excellent educational videos.

Besides the educational value, this project will be used to test and tune compiler optimizations and features on the JVM, particularly for the Graal compiler.

Features

• Single file, no dependencies
• GGUF format parser
• Llama 3+ tokenizer based on minbpe
• Llama 3+ inference with Grouped-Query Attention
• Support Llama 3.1 (ad-hoc RoPE scaling) and 3.2 (tie word embeddings)
• Support F16, BF16, F32 weights + Q4_0, Q4_1, Q4_K, Q5_K ,Q6_K, Q8_0 quantizations
• Fast matrix-vector multiplication routines using Java's Vector API
• Simple CLI with --chat and --instruct modes.
• GraalVM's Native Image support
• AOT model pre-loading for instant time-to-first-token

Interactive --chat mode in action:

Practical LLM inference in modern Java

Presented at Devoxx Belgium, 2024

Setup

Download pure Q4_0 and (optionally) Q8_0 quantized .gguf files from:

• https://huggingface.co/mukel/Llama-3.2-1B-Instruct-GGUF
• https://huggingface.co/mukel/Llama-3.2-3B-Instruct-GGUF
• https://huggingface.co/mukel/Meta-Llama-3.1-8B-Instruct-GGUF
• https://huggingface.co/mukel/Meta-Llama-3-8B-Instruct-GGUF

Or from unsloth https://huggingface.co/unsloth/Llama-3.2-1B-Instruct-GGUF

Optional: Pure quantizations

In the wild, Q8_0 quantizations are fine, but Q4_0 quantizations are rarely pure e.g. the token_embd.weights/output.weights tensor are quantized with Q6_K, instead of Q4_0.
A pure quantization can be generated from a high precision (F32, F16, BFLOAT16) .gguf source with the llama-quantize utility from llama.cpp as follows:

./llama-quantize --pure ./Meta-Llama-3-8B-Instruct-BF16.gguf ./Meta-Llama-3-8B-Instruct-Q4_0.gguf Q4_0

Pick any of the supported quantizations: Q4_0, Q4_1, Q4_K, Q5_K, Q6_K, Q8_0.

Build and run

Java 21+ is required, in particular for the MemorySegment mmap-ing feature.

jbang is a perfect fit for this use case, just:

jbang Llama3.java --help

Or execute directly, also via jbang:

chmod +x Llama3.java
./Llama3.java --help

Optional: Makefile

A simple Makefile is provided, run make jar to produce llama3.jar.

Run the resulting llama3.jar as follows:

java --enable-preview --add-modules jdk.incubator.vector -jar llama3.jar --help

GraalVM Native Image

Compile with make native to produce a llama3 executable, then:

./llama3 --model Llama-3.2-1B-Instruct-Q8_0.gguf --chat

AOT model preloading

Llama3.java supports AOT model preloading, enabling 0-overhead, instant inference, with minimal TTFT (time-to-first-token).

To AOT pre-load a GGUF model:

PRELOAD_GGUF=/path/to/model.gguf make native

A specialized, larger binary will be generated, with no parsing overhead for that particular model. It can still run other models, although incurring the usual parsing overhead.

Performance

GraalVM 25+ is recommended for the absolute best performance, it provides partial but good support for the Vector API.

By default, the "preferred" vector size is used, it can be force-set with -Dllama.VectorBitSize=0|128|256|512, 0 means disabled.

llama.cpp

Vanilla llama.cpp built with make.

./llama-cli --version
version: 3862 (3f1ae2e3)
built with cc (GCC) 14.2.1 20240805 for x86_64-pc-linux-gnu

Executed as follows:

./llama-bench -m Llama-3.2-1B-Instruct-Q4_0.gguf -p 0 -n 128

Llama3.java

taskset -c 0-15 ./llama3 \
  --model ./Llama-3-1B-Instruct-Q4_0.gguf \
  --max-tokens 128 \
  --seed 42 \
  --stream false \
  --prompt "Why is the sky blue?"

Hardware specs: 2019 AMD Ryzen 3950X 16C/32T 64GB (3800) Linux 6.6.47.

**Notes
Running on a single CCD e.g. taskset -c 0-15 ./llama3 ... since inference is constrained by memory bandwidth.

Results

Related Repositories

• llama3.java
• gemma4.java
• gptoss.java
• qwen35.java

License

MIT