LTX-LLM

LTX-LLM runs batches of performance tests against locally hosted language models on backends including llama.cpp, LM Studio and Ollama, reporting metrics such as time to first token, throughput, total run time and memory usage. It supports testing by predefined scenario, sweeping context-length settings, running a whole suite of models in one pass, and a separate reasoning-focused test mode with a looser output budget for models that show reasoning traces.

Each run writes a per-model CSV of results, and separate scripts generate comparison charts across models or across scenarios. It scans a configurable model directory, falling back through a command-line flag, an environment variable, or LM Studio's own default model folder, and skips a model suite item whose complete results already exist. It installs its Python dependencies from a requirements file and, in llama.cpp mode, starts the local llama-server automatically where possible.