PreTensor

Runs PreDoctor tensor workloads in a local NVIDIA CUDA host.

PreTensor connects supported PreDoctor web pages to an NVIDIA CUDA inference runtime running on your own Windows computer. The extension provides a narrow browser-to-native bridge for user-requested tensor, text, and vision inference. Prompts, token data, images, and inference results are processed locally and are not transmitted by the extension to PreDoctor, advertisers, or analytics providers. Features: - local NVIDIA CUDA tensor and model execution; - text prefill and token-by-token decoding; - local vision encoding; - GPU, CUDA, and runtime readiness diagnostics; and - access restricted to PreDoctor and local-development pages. Requirements: - Windows 10 or Windows 11, x64; - a compatible NVIDIA GPU and driver; and - the separately installed PreTensor companion runtime. The extension does not include a model or native executable. Without the companion runtime, the popup reports that the local runtime is unavailable. Reference performance: in an end-to-end local Qwen3.8 27B test with a 16,383-token prompt, PreTensor measured approximately 3,980 prefill tokens/second and 37.7 decoding tokens/second on Windows, Chrome 153, and an NVIDIA RTX PRO 6000 Blackwell Max-Q GPU. Actual performance varies with the GPU, driver, model, context length, available VRAM, and system configuration.