kalebbroo 

Core tensor types, backend abstractions, and shared interfaces for the HartsyInference AI inference engine.

Meta-package for HartsyInference — a pure C#/.NET AI inference engine for text, diffusion images, speech, music, vision, video, 3D, and interactive world models. Adds all backends (CPU, CUDA, Vulkan) and... More information

Image generation pipelines for HartsyInference — SD1.5, SDXL, Flux, SD3, LoRA, ControlNet, and schedulers.

Everything you load off disk to make a model, for HartsyInference — weight containers (SafeTensors, GGUF, PyTorch pickle, ONNX), quantization codecs (NF4, MXFP4, MXFP8, NVFP4, BlockScale), checkpoint... More information

CPU inference backend for HartsyInference with AVX2/AVX-512/NEON SIMD kernels.

Native LLM text generation for HartsyInference — a config-driven generic decoder transformer (Qwen2/Qwen3 and beyond), device-resident KV cache, sampler chain, chat templates, and a text-generation pipeline.... More information

Video generation for HartsyInference — LTX-Video, Wan, temporal attention, and video VAE. (Phase 3 — stub)

Vision inference for HartsyInference — CLIP embeddings, YOLO detection (planned), SAM segmentation (planned), and face detection (planned).

Audio inference for HartsyInference — Whisper STT, Kokoro TTS, voice conversion, mel-spectrogram preprocessing, audio codecs, and HuggingFace model caching.

Vulkan compute backend for HartsyInference. Cross-vendor GPU inference (NVIDIA, AMD, Intel, Apple via MoltenVK) via SPIR-V compute shaders and P/Invoke into the Vulkan loader.

CUDA GPU inference backend for HartsyInference with PTX kernels and cuBLAS integration.

3D asset generation for HartsyInference — image/text → mesh and Gaussian-splat. Representation-agnostic foundation (marching cubes, glTF/OBJ/PLY export, 3D sampling) plus model pipelines (Hunyuan3D-2... More information

Interactive world models for HartsyInference — action-conditioned real-time streaming sessions, camera/Plücker conditioning, FOV memory retrieval, and Matrix-Game pipelines. (Phase 10)

The HartsyInference service layer: the single entry point for loading models and generating across every modality. Consumed by the CLI, the HTTP API, and the SwarmUI backend extension.

Shared GPU backend layer for HartsyInference: device-residency caching, op profiling and the backend base every GPU backend derives from. Vendor-neutral — contains no CUDA, Vulkan, ROCm or Metal interop.

Opt-in LLM tool calling for HartsyInference: incremental tool-call parsers per model format (Hermes/Qwen, Llama-3, Gemma, Mistral), the ITextStreamFilter that plugs them into the engine, a tool registry with... More information

Opt-in real-time voice agent for HartsyInference: a per-call session that turns streamed caller audio into spoken replies. RNNoise and Silero endpointing and barge-in run on a dedicated audio thread, speech... More information

CLIP, T5, Qwen3, and Whisper tokenizers for HartsyInference diffusion and audio pipelines. Ships canonical OpenAI-CLIP / T5-XXL / Qwen3-4B vocabularies as embedded resources so consumers can construct... More information

Model loading for HartsyInference — safetensors, GGUF, model registry, and HuggingFace Hub integration.

Interactive world models for HartsyInference — action-conditioned real-time streaming sessions, camera/Plücker conditioning, FOV memory retrieval, and Matrix-Game pipelines. (Phase 10)