what is llm-inspector fr?

LLM Inspector is a command line tool that shows exactly what is happening inside a running AI model's memory on your GPU. Most monitoring tools tell you a process is using a certain amount of GPU memory, but not why. LLM Inspector breaks that number down into categories such as model weights, the KV cache, workspace memory, and other overhead, so you can see where every gigabyte is actually going. Beyond just showing current usage, it also projects what would happen if you applied different optimization strategies, starting with quantization, before you actually make any changes. For example, it can estimate how many gigabytes an FP8 or AWQ quantization scheme would save on model weights, and it will also tell you plainly when a technique will not help, such as noting that shrinking weights will not fix a bottleneck caused by a large KV cache. The README is clear that the live process, hardware, model, memory, and runtime numbers are measured directly, while the optimization projections are calculated estimates that never actually change the running model. It works with common inference setups including Ollama, vLLM, HuggingFace Transformers, FastAPI-based services, and custom PyTorch code, and runs on macOS, Linux, or any machine with an NVIDIA GPU, including DGX systems. A basic install through pip gives you the process level view, while installing the optional torch extra lets it attach directly inside the same Python environment as the model for deeper metrics like weights, KV cache, and activations. This is aimed at people who run or deploy large language models locally or on their own servers and want to understand and reduce GPU memory usage rather than guess at it. It is released under the MIT license.