Kernel Orbit
Run CUDA/Triton kernels and Jupyter notebooks on Modal GPUs from VS Code.
What it does
- Kernel files: Run
.cu (CUDA) and .py (Triton) files on remote GPUs with benchmarking, profiling, and detailed metrics.
- Jupyter notebooks: Select "Modal GPU (Python)" as your notebook kernel and run cells on a remote GPU. State persists across cells; the container stays warm for 15 minutes between runs.
- GPU picker: Choose from T4, L4, A10G, A100, L40S, H100, H200, B200.
- Sessions sidebar: See active GPU sessions and kill them from the Kernel Orbit panel.
How to use
Setup
- Install uv
uv add modal- Get API tokens from modal.com/settings
- Create a
.env in your project root:
MODAL_TOKEN_ID=ak-...
MODAL_TOKEN_SECRET=as-...
Kernels (.cu / .py)
- Open a
.cu or .py file
Cmd+Shift+R (Mac) / Ctrl+Shift+R (Win/Linux)- Results appear in the Results Panel
Notebooks (.ipynb)
- Open a
.ipynb file
- Select Modal GPU (Python) from the kernel picker — the container starts warming immediately
- Run cells normally — they execute on the remote GPU
- State persists across cells; container stays warm for 15 min of inactivity
GPU Selection
Click the GPU name in the status bar or run Kernel Orbit: Select GPU Type from the command palette.
Configuration
Open VS Code Settings and search for "Modal Kernel" to customize:
- Timeout (default: 3600s = 1 hour): Maximum time a single notebook cell can run
- Increase for long training runs (max: 86400s = 24 hours)
- Decrease for quick experiments to catch runaway cells
- Default GPU (default: T4): GPU type to use by default
- GPU Count (default: 1): Number of GPUs to attach (1-8)
- Warmup/Benchmark Runs: Configure profiling behavior for kernel files
- Sync Files (default: true): Sync local workspace files to the remote GPU container
- Max Sync File Size MB (default: 100): Files larger than this are skipped with a warning
- Sync Exclude Patterns: Additional glob patterns to exclude from sync
Notebook Features
File sync
Local workspace files are synced to the remote container before each cell execution. Files ≤ 100 MB are synced automatically; larger files trigger a warning with alternatives (!wget on GPU or Modal Volume).
Package installation
!pip install, %pip install, and !pip3 install are automatically proxied through uv pip install --system for faster installs. You can also use %uv pip install directly.
Supported magics
%time, %timeit, %%time, %%timeit, %matplotlib, %pip, %uv, ! shell commands.
Roadmap
- [ ]
input() support (input_request / input_reply)
- [ ] Top-level
await (async cell execution)
- [ ] More IPython magics (
%cd, %pwd, %env, %who)
- [ ]
tqdm progress bar support
- [ ]
ipywidgets support
- [ ]
IPython.display.clear_output()
- [ ]
%%bash, %%html, %%javascript cell magics
Development
git clone https://github.com/kashifulhaque/kernel-orbit
cd kernel-orbit
npm install
# Press F5 in VS Code to launch the extension development host
