A free VRAM calculator for AI models. Not just for LLMs: text generation, embeddings, vision, multimodal, image diffusion, video, audio, and tabular workloads, across inference, LoRA/QLoRA fine-tuning, and full training. Every calculation runs in the browser. Demo app to simultaneously build frontend harness.

MCP server for Wan2GP video generation with GPU detection and VRAM management

Connects remote Ollama servers to local clients over LAN with auto-discovery and VRAM management.

Local AI platform: WebGPU/WGSL browser inference engine + HuggingFace Transformers + Ollama. TurboQuant KV cache compression, GPTQ INT4 fused dequant, mixed-precision BF16/INT4 for hybrid SSM+attention models. 9B parameters in a browser, 8GB VRAM.

Discrete-event simulation of multi-LoRA adapter serving strategies on a single GPU — comparing naive swap, hot-set preloading, and batch-by-adapter under variable VRAM pressure and arrival rates.

Warm KV-transfer llama.cpp swaps under a sandboxed TypeScript policy, with a per-GPU VRAM budget and load-driven autoscaling.

[LEGACY PoC] A sovereign, local-first AI reasoning runtime and VRAM orchestration engine built for constrained consumer hardware.