FAST HIRE: Run an open-source Python tool on a rented GPU and log where you got stuck

We publish an open-source tool for measuring what LLM inference costs on different GPUs. It is about to go out to a set of engineering teams, and I want to find where the documentation fails before they do. This is a usability test, not a development job. The deliverable is a written log of every place you got confused. Bugs are welcome but they are not the point. The point is the pauses. What you would do 1. Rent a GPU of your choosing. An A10, L4, L40S, A100 or similar is fine, on RunPod, Lambda, Vast or wherever you already have an account. Reimbursed on receipt, typically 2 to 4 dollars. 2. Start a vLLM or SGLang server with any small open-weight model. 3. Install our tool and follow its documentation until you have produced a measurement and an error report. I will send you the install command and the docs URL when I hire you. 4. Write down every point where you paused, guessed, backtracked or went looking for something the docs did not tell you. Rules Do not contact me while you work. If you get stuck, write down the timestamp and what confused you, then either work around it or stop. A message asking me a question destroys the data I am paying for. Stopping early is a valid result. If you abandon it after forty minutes, say where and why. That is more useful to me than a success you had to fight for. Do not read the source to answer a question the docs should have answered. If you find yourself opening the code to work out what a flag does, log that as a documentation failure first, then do whatever you like. Deliverable A plain text or markdown file containing: Every point of confusion, with a rough timestamp and what you expected versus what happened Total elapsed time, and how much of it was spent stuck Whether the final number looked plausible to you, and why or why not The three things you would change about the documentation, in priority order The trace file and error report the tool produced Terminal output, screenshots and a screen recording are welcome but optional. Who this suits Someone who has rented a GPU before and started an inference server before. You should be comfortable in a Linux shell. You do not need to know anything about our tool, and it is better if you do not: I am testing whether a competent stranger can get to a number using only what is published. Prior contact with this project disqualifies you for this task. What this is not Not a code review. Not a request to fix anything. Not an evaluation of whether the numbers are correct, which is our problem and not yours. To apply Two sentences on the last time you stood up an inference server, what you ran it on, and what tripped you up. I am not looking for a cover letter.