[models] Support for Qwen3 models

[orionpapadakis]

On going work for #19

• add model support
• refactor model loaders and inference engines to be modular

Check list:

[x] CPU inference path in a working state
[x] GPU inference path in a working state

[orionpapadakis]

ready for review

[Copilot]

Pull Request Overview

This PR adds support for Qwen3 models to the codebase, implementing a modular architecture that refactors model loading and inference engines to support multiple model types. The implementation includes both CPU and GPU inference paths through TornadoVM for Qwen3 models, alongside architectural improvements to the existing LLaMA and Mistral model support.

Key changes include:

• Adding Qwen3 model support with specialized tokenization, configuration, and inference logic
• Refactoring the model loading system to use a modular pattern with abstract base classes
• Implementing separate state management and weight handling for different model architectures

Reviewed Changes

Copilot reviewed 44 out of 44 changed files in this pull request and generated 5 comments.

Show a summary per file

File	Description
TornadoVMLayerPlanner.java	Refactored to support generic model types with parameterized base class
Qwen3TornadoVMLayerPlanner.java	New Qwen3-specific GPU execution planner with custom kernel configurations
Qwen3Kernels.java	Qwen3-specific GPU kernels including RMSNorm and RoPE rotation implementations
Qwen3Tokenizer.java	Complete Qwen3 tokenizer implementation with BPE encoding/decoding
Model architecture files	New Qwen3Configuration, Qwen3 model class, and supporting infrastructure
Weight/State refactoring	Separated standard and TornadoVM weight classes, model-specific state classes
Model loader refactoring	Abstract ModelLoader base with concrete implementations for each model type

Comments suppressed due to low confidence (3)

src/main/java/com/example/tornadovm/TransformerComputeKernelsLayered.java:441

• [nitpick] The variable name 'shared_tile_max_holder' is verbose and the comment suggests it's a workaround. Consider renaming to 'tileMaxBuffer' for clarity and consistency with other buffer variables.

        float[] shared_tile_max_holder = context.allocateFloatLocalArray(1); // FIX: For broadcasting tile max

src/main/java/com/example/tornadovm/TransformerComputeKernelsLayered.java:623

• [nitpick] The parameter name 'hb' is not descriptive. Consider renaming to 'output' or 'outputBuffer' to match the comment and improve readability.

            FloatArray hb,                  // output

src/main/java/com/example/inference/state/Qwen3State.java:25

• The variable 'nEmbdHead' is assigned 'numberOfHeads()' but based on context, it should likely be 'numberOfHeadsValue()' or a calculated embedding head size. This naming suggests a mismatch between the variable name and its actual value.

        int nEmbdHead = qwen3config.numberOfHeads();

[mikepapadim]

@orionpapadakis also, update the readmen with Qwen models instructiosn etc