VRAM.cpp runs llama.cpp fit in browser

This is a smart answer to a real pain point: instead of approximating memory from model size, it reuses the same fitting logic the runtime depends on, which should make estimates much more credible.

• –The core advantage is fidelity: as llama.cpp’s fit algorithm evolves, the estimator inherits those improvements without a separate rules engine to keep in sync.
• –That makes it more useful than generic VRAM calculators for edge cases like Q3 variants, hybrid GPU+RAM fits, and newer model families.
• –The project still admits weak spots in multi-GPU plus host-memory splits and MoE fitting, so the hardest configurations are exactly where users should be most cautious.
• –As an open-source browser app, it lowers the friction for quick pre-flight checks before downloading huge GGUFs or running trial fits locally.