omp packages

@sipphq/lemon-pi0.1.8

Run local .uma models in Pi through the Sipp Warp runtime

constannnnnt Published 5 days ago Apache-2.0 ~1.3K downloads/mo

Untested on omp

This is a pi package. No one has reported whether it works on omp yet. If you try it, add your result below.

Did it work for you?

Result

One report per person. A new report replaces your old one.

Reports 0

No reports yet.

Readme

lemon-pi

Run local .uma models in Pi on your own GPU. lemon-pi adds a Warp model provider and a /warp command: find a model on Hugging Face, download it, and chat with it in Pi — no API key, nothing leaves your machine.

Requirements

Platform GPU Also needed
Windows x64 NVIDIA, driver 580 or newer Nothing: installing lemon-pi downloads the CUDA 13 libraries
Linux x64 (glibc) NVIDIA, driver 580 or newer Nothing: installing lemon-pi downloads the CUDA 13 libraries
macOS, Apple Silicon Built in (Metal) Nothing

On Windows and Linux, installing lemon-pi checks for CUDA 13. When none is installed, it downloads NVIDIA's CUDA 13 libraries (about 400 MB, once) into ~/.sipp/cuda, so the install takes a few minutes longer. An installed CUDA 13 toolkit is used instead when there is one. CUDA 12 or older cannot stand in for it.

You also need Pi 1.0 or a later 1.x release (running on Node.js 22.19 or newer), and disk space for models (the current Qwen build is about 12.5 GB).

Install

pi install npm:@sipphq/lemon-pi

npm downloads only the runtime for your platform. To install for the current project only, add -l; to try it for one session without installing, run pi -e npm:@sipphq/lemon-pi.

Your first model

  1. Start pi.
  2. Type /warp hub and press Enter. Pick a repository, then the variant for your machine: cuda on Windows and Linux, metal on a Mac.
  3. lemon-pi reads the model's details (a few MB, not the weights) and shows whether it fits your GPU. Confirm to download. The download runs in the background with progress in Pi's footer; you can keep working.
  4. Type /warp models and pick the downloaded file. lemon-pi fits the context window to your GPU (it loads the model once to measure) and switches Pi to it.
  5. Chat. The model loads in the background as soon as it is selected; the footer shows Warp · <model> · loaded when it is ready.

Next time Pi starts, pick the model with /warp models or /model, or make it the default with /warp config default.

Everyday commands

Type /warp for a menu. Every subcommand without arguments opens its own menu, and Tab completes names as you type.

Command What it does
/warp models List your models and downloaded files; pick one to use it
/warp hub Browse .uma models on Hugging Face
/warp hub <org/repo> Download a model (/warp hub cancel stops it; run it again to resume)
/warp config Settings of the model in use: context window, GPU memory share, presets
/warp config load / unload Load the model now, or free its GPU memory
/warp config default Make it the model Pi starts with
/warp info What is loaded, context use and the last request's speed
/warp models unlist <model> Take a model off the list; its file stays on disk
/warp models delete <model> Delete a downloaded model from disk
/warp prefs Where models are stored, and when models preload
/warp help The full command reference

<model> can be a list number, a name, a file name, or a unique part of one.

Where things are kept

  • Downloaded models: ~/.sipp/models/<org>/<repo>/<file>.uma. Change the folder with /warp prefs models-dir <path> or the SIPP_MODELS_DIR environment variable.
  • Model settings: ~/.pi/agent/extensions/warp.json (/warp info prints the exact path).
  • Compaction limits: lemon-pi keeps compaction.modelOverrides for each Warp model in ~/.pi/agent/settings.json sized to its context window, so Pi can compact a long conversation even in a small window. When a /warp command changes a window, Pi reloads its settings after the command.
  • Set up a file from elsewhere: /warp models add <path/to/model.uma>.

Update and uninstall

pi update npm:@sipphq/lemon-pi
pi remove npm:@sipphq/lemon-pi

Removing lemon-pi keeps your downloaded models. Delete them first with /warp models delete, or remove the models folder yourself.

Troubleshooting

  • "Warp needs the CUDA 13 runtime" (Windows) or "Warp needs the CUDA 13 libraries" (Linux): the install did not download them. The message says why, for example a missing or old NVIDIA driver. The download runs from npm's install script, which pnpm, bun and npm's ignore-scripts setting skip: reinstall with npm (pi install npm:@sipphq/lemon-pi). Or install the CUDA 13 toolkit: lemon-pi finds it under the default install folder (C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA, or /usr/local/cuda-13.*) or CUDA_PATH. Anywhere else on Linux, add it to the loader path, for example export LD_LIBRARY_PATH=/opt/cuda-13/lib64:$LD_LIBRARY_PATH, then restart Pi. CUDA 12 or older does not work: each CUDA major version has its own libraries.
  • Installing behind a proxy or mirror: set LEMON_PI_CUDA_BASE_URL to a mirror of the CUDA download, or LEMON_PI_SKIP_CUDA_DOWNLOAD=1 to skip it and install the CUDA 13 toolkit yourself.
  • A model "needs attention": /warp models says why and which command fixes it — usually a moved or deleted file that /warp hub can download again.
  • Out of GPU memory: close other GPU programs, or give the model a smaller share with /warp config preset context or /warp config vram 0.5.
  • Hugging Face rate limits or a mirror: set HF_ENDPOINT to the mirror's URL.

/warp info shows anything that blocks loading.

License

Apache-2.0.