LLM engineering quickly becomes a sizing problem. A model that looks reasonable on paper can fail because weight memory, KV cache, sequence length, batch size, quantization format, and runtime overhead do not fit the target accelerator. This collection groups local tools that help with early inference planning and low-level explanation.
Start with the LLM / NPU memory calculator to estimate weight and KV cache pressure. Use the FP quantization visualizer to explain how reduced precision changes representable values. When the work touches hardware registers, bitfields, clock plans, or timing notes, the related hardware tools help keep those calculations in the same local workflow.
These pages are deliberately lightweight. They do not upload model names, private benchmark notes, or accelerator details. They are best used before a full benchmark: explore assumptions, copy a result into a planning note, then validate with the actual runtime stack.
The page is written for practical exploration rather than hype. It gives search users enough context to decide which calculator to open, why KV cache differs from weight memory, and how quantization visuals relate to deployment trade-offs. That content depth helps the tool collection become a reference that people return to during sizing discussions.