Quick start

Install

Install Popcorn from PyPI with uv:

uv pip install popcorn           # first-party kernels
uv pip install "popcorn[liger]"  # Liger backends
uv pip install "popcorn[fla]"    # FLA backends
uv pip install "popcorn[cudnn]"  # cuDNN attention
uv pip install "popcorn[transformer_engine]"  # Transformer Engine softmax

[!NOTE] Add --torch-backend=auto to let uv select a PyTorch build for your hardware.

[!WARNING] pip install popcorn is not generally supported. If required, install the appropriate PyTorch build before installing Popcorn.

Install one extra for each required backend. Some backends have incompatible dependency requirements, so there is no combined extra.

Installation fetches the pinned benchmark data from popcorn-reports. Set POPCORN_SKIP_REPORTS=1 to install without reports. Automatic dispatch then uses the reference implementation. Refresh the reports with:

uv run popcorn bench pull

Reports apply only to the hardware on which they were recorded. Use popcorn bench fill to measure missing cases on the current device. See Benchmarking for details.

A backend is eligible only when a supported package version is installed. Automatic dispatch skips unavailable backends. Forcing an unavailable backend raises an installation or version error. First-party kernels under the popcorn backend are always available.

Upgrade Popcorn and its pinned reports with:

uv pip install --upgrade popcorn

Call a kernel

Import and call a kernel. Popcorn selects an eligible implementation using benchmark data, with the registered reference as the fallback.

import torch
from popcorn.kernels import rms_norm

x = torch.randn(2, 512, 4096, device="cuda", dtype=torch.bfloat16)
weight = torch.ones(4096, device="cuda", dtype=torch.bfloat16)

output = rms_norm(x, weight)

To select a backend explicitly:

output = rms_norm["liger"](x, weight)
# same as:
output = rms_norm(x, weight, backend="liger")

Next steps

• Dispatching: Implementation selection
• Validation: Correctness checks
• Benchmarking: Performance measurement
• Kernel explorer: Kernels, backends, and results