The Htop for LLM Inference

LLM Inspector inspects live inference processes and shows exactly how GPU memory is being used, what model is running, how the runtime is configured, and where every reported value comes from.

Unlike traditional monitoring tools, it doesn't stop at inspection. It also analyzes the running workload and projects optimization opportunities—starting with quantization—to help you understand how different strategies would impact GPU memory before making any changes.

Existing tools tell you that your GPU is using 18 GB.

LLM Inspector tells you why:

Weights       7 GB
KV Cache      8 GB
Workspace     1 GB
Other         2 GB

Then it tells you what would happen if you optimized:

FP8 would save ~3 GB of weights.
AWQ would save ~5 GB of weights.
Weight quantization won't fix an 8 GB KV Cache bottleneck.

That is the difference between a GPU monitor and an inference advisor.

Most tools stop here:

GPU
└── Process A
    └── 17.3 GB

LLM Inspector goes one level deeper:

GPU
└── Process A
    ├── Weights
    ├── KV Cache
    ├── Workspace
    ├── Activations
    └── Optimization Analysis (Projected)

That bridge—system observability + model internals, with an htop

-style CLI—is the innovation. Deep metrics come from optional embedded attach()

inside the inference process (see the install guide).

llminspect inspect <pid> --verbose   # show provenance for every field

Works on any NVIDIA GPU machine — laptop, workstation, cloud VM, bare-metal server, or DGX. Docker is optional.

pip install llm-inspector

For embedded deep metrics (Weights / KV / Activations) in the same Python env as the model:

pip install "llm-inspector[torch]"

From source (contributors):

git clone https://github.com/helasaoudi/llm-inspector
cd llm-inspector
python -m venv .venv && source .venv/bin/activate
pip install -e ".[torch]"
llminspect ps
llminspect inspect <pid>
llminspect inspect <pid> --verbose

With Ollama on the host (no Docker):

ollama run llama3
llminspect inspect $(pgrep -f "ollama serve") --verbose

Ollama · vLLM · HuggingFace Transformers · FastAPI · custom PyTorch

macOS · Linux · any NVIDIA GPU server (including DGX)

MIT