How bhangun/alkhawarizm works
Al-Khawarizm is a Java foundation for tensor math, model topology, weight handling, and hardware execution—not a text-generation or training-loop application. The principal workflow is that a downstream engine selects a model family and device, loads model weights, then uses neural-network primitives and a compute backend for forward-pass operations. The README also documents autograd for downstream training and model conversion, quantization, and RAG-related capabilities. Several links below reflect README-level architecture rather than inspected runtime wiring; unsampled implementation details are represented at their existing module directories.
Main entry points: ComputeBackend.java Compute backend SPI · models Model families · NNModule.java NN primitives · SafetensorEngine.java SafeTensor engine SPI
Components 22
Model topology
- Model families
models - Model SPI
core/alkhawarizm-spi-model/src/main/java/tech/kayys/alkhawarizm/spi/model/ModelConfig.java - NN primitives
core/alkhawarizm-nn/src/main/java/tech/kayys/alkhawarizm/core/nn/NNModule.java
Weights and formats
- SafeTensor loading
core/alkhawarizm-safetensor-loader - SafeTensor engine SPI
core/alkhawarizm-safetensor-spi/src/main/java/tech/kayys/alkhawarizm/safetensor/spi/SafetensorEngine.java - Weight quantization
core/alkhawarizm-safetensor-quantization/src/main/java/tech/kayys/alkhawarizm/safetensor/quantization/QuantizationEngine.java - GGUF loading
core/alkhawarizm-gguf-core/src/main/java/tech/kayys/alkhawarizm/gguf/loader/GGUFLoader.java - GGUF conversion
core/alkhawarizm-gguf-converter/src/main/java/tech/kayys/alkhawarizm/converter/GGUFConverter.java - Model storage SPI
core/alkhawarizm-spi-model/src/main/java/tech/kayys/alkhawarizm/spi/model/ModelStorageService.java
Tensor compute
- Tensor API
core/alkhawarizm-tensor/src/main/java/tech/kayys/alkhawarizm/core/tensor/Tensor.java - Compute backend SPI
core/alkhawarizm-tensor/src/main/java/tech/kayys/alkhawarizm/core/backend/ComputeBackend.java - Inference memory
core/alkhawarizm-tensor/src/main/java/tech/kayys/alkhawarizm/core/memory/InferenceSession.java
Hardware backends
- CPU backend
backend/cpu/alkhawarizm-backend-cpu/src/main/java/tech/kayys/alkhawarizm/backend/cpu/CpuBackend.java - CUDA backend
backend/cuda/alkhawarizm-backend-cuda/src/main/java/tech/kayys/alkhawarizm/backend/cuda/CUDABackend.java - Metal backend
backend/metal/alkhawarizm-backend-metal/src/main/java/tech/kayys/alkhawarizm/backend/metal/MetalComputeBackend.java - ROCm backend
backend/rocm/alkhawarizm-kernel-rocm/src/main/java/tech/kayys/alkhawarizm/rocm/binding/RocmHipBinding.java - Blackwell backend
backend/blackwell/alkhawarizm-kernel-blackwell/src/main/java/tech/kayys/alkhawarizm/blackwell/binding/BlackwellBinding.java
Supporting capabilities
- Autograd
core/autograd/src/main/java/tech/kayys/alkhawarizm/autograd/AutogradEngine.java - RAG API
core/alkhawarizm-safetensor-api/src/main/java/tech/kayys/alkhawarizm/safetensor/api/RagPipeline.java - Diffusion pipeline
models/alkhawarizm-model-common/src/main/java/tech/kayys/alkhawarizm/models/StableDiffusionPipeline.java - Speech pipeline
models/alkhawarizm-model-common/broken_models/WhisperPipeline.java
Other components
- Downstream engine
Connections 22
- Downstream engine → Model SPI: uses contracts
README.md - Downstream engine → Model families: uses topology
README.md - Model families → Model SPI: implements SPI
README.md - Downstream engine → SafeTensor engine SPI: loads models
core/alkhawarizm-safetensor-spi/src/main/java/tech/kayys/alkhawarizm/safetensor/spi/SafetensorEngine.java - SafeTensor engine SPI → SafeTensor loading: loads weights
- Downstream engine → GGUF loading: loads GGUF
- Downstream engine → Model storage SPI: stores artifacts
- Downstream engine → Weight quantization: quantizes weights
- Downstream engine → NN primitives: executes topology
README.md - NN primitives → Tensor API: operates on
core/alkhawarizm-nn/src/main/java/tech/kayys/alkhawarizm/core/nn/NNModule.java - Tensor API → Compute backend SPI: dispatches compute
core/alkhawarizm-tensor/src/main/java/tech/kayys/alkhawarizm/core/tensor/Tensor.java - Downstream engine → Inference memory: uses memory
- Compute backend SPI → CPU backend: implemented by
README.md - Compute backend SPI → CUDA backend: implemented by
README.md - Compute backend SPI → Metal backend: implemented by
README.md - Compute backend SPI → ROCm backend: implemented by
README.md - Compute backend SPI → Blackwell backend: implemented by
- GGUF conversion → GGUF loading: produces format
- Downstream engine → Autograd: uses gradients
README.md - Downstream engine → RAG API: uses RAG API
- Downstream engine → Diffusion pipeline: uses model pipeline
- Downstream engine → Speech pipeline: uses speech pipeline