FlashJet

FlashJet groups particles into jets. It accepts a NumPy event or a padded batch of PyTorch tensors. CUDA tensors can stay on the GPU during clustering.

Use it when a training loop needs to cluster jets repeatedly. The particle assignment is discrete. Summing each jet’s four-momentum remains differentiable with respect to its assigned particles.

This site follows the published main branch. Install FlashJet from Git using the installation guide. The package version is 0.1.0; a development snapshot is not a tagged release. PyPI publishing is deferred.

See Git distribution and releases for CI coverage, downloadable artifacts, and the checks required before a release.

• Installation
  • Install from Git
  • Wheels from CI
  • Install from a checkout
  • Native CPU kernel
• Quickstart
  • One NumPy event
  • A padded tensor batch
  • Algorithms and gradients
• flashjet, in plain English
  • 1. The problem
  • 2. Why this library exists
  • 3. Five minutes to your first jets
  • 4. Which engine runs, and why you can ignore it
  • 5. Reading the jet’s internal structure
  • 6. Feeding it real data
  • 7. How fast is it
  • 8. Things that will trip you up
  • 9. Where to look next
• Backends and numerical behavior
  • Precision and reproducibility
  • History layout
• Substructure and grooming
• Ragged events and GPU batches
• API reference
  • Clustering
  • NumPy reference
  • History helpers
  • Data helpers
• Performance and benchmarks
  • Optional tuning
• Troubleshooting
  • Triton backend unavailable
  • CPU install works but clustering is slow
  • Invalid input or unexpected jets
  • Results differ slightly
  • GPU memory is exhausted
  • A test skips
• Development
• Git distribution and release checklist
  • Install from Git
  • CI coverage
  • Download CI artifacts
  • Create a tagged GitHub release
  • GPU validation
  • Documentation deployment
• Credits and license

Source · GitHub releases · Issues