Skip to main content

dinosaur-torch

dinosaur-torch

PyPI Python 3.11+ PyTorch GitHub CI License: Apache 2.0

An idiomatic-PyTorch rewrite of Dinosaur, the spectral dynamical core behind NeuralGCM. Rather than a line-for-line JAX→PyTorch translation, this package is written the way a PyTorch library would be written from scratch, while staying numerically equivalent to the original.

Numerically validated against the original JAX Dinosaur.

Design

  • Tensors in, tensors out. Functions and modules operate on torch.Tensors. NumPy appears only at I/O and construction boundaries (grid/quadrature setup, xarray conversion). There is no asarray promotion at every call site, no global default-device convention, and no host-constant device cache.
  • Precomputed constants live in nn.Module buffers. Objects that hold tensors (spectral transforms, the dycore) are torch.nn.Modules with non-persistent buffers, so .to(device) / .float() work the standard way and state_dict() contains only learned parameters (none, for the dycore).
  • Static metadata is separate from tensors. GridSpec, SigmaCoordinates etc. are frozen dataclasses — hashable, comparable, cheap — used to construct the tensor-holding modules.
  • States are torch pytrees. Model state (State, diagnostics, …) is a plain dataclass registered via torch.utils._pytree.register_dataclass, so it composes natively with torch.compile, torch.func, and CUDA graphs. No custom pytree registry.
  • Standard test style: plain pytest with parametrization (no absl/parameterized).
  • Scope: the primitive-equations path used by NeuralGCM (transforms, sigma coordinates, primitive equations, IMEX time integration, filtering, vertical/horizontal interpolation, data utilities). Shallow-water and Held–Suarez model families are intentionally not ported: no published NeuralGCM checkpoint uses them.

Layout

module contents
associated_legendre.py, fourier.py basis construction (NumPy, at setup time)
spherical_harmonic.py GridSpec (static), RealSphericalHarmonics / FastSphericalHarmonics transforms, Grid (nn.Module: transforms + spectral operators)
sigma_coordinates.py SigmaCoordinates (static) + SigmaLevels (nn.Module: vertical finite-difference / integral operators)
coordinate_systems.py CoordinateSystem (nn.Module: horizontal × vertical), spectral up/downsampling
primitive_equations.py State (torch-pytree dataclass), PrimitiveEquations (nn.Module IMEX ODE, dry/moist/cloud variants), Geopotential
time_integration.py IMEX Runge-Kutta steppers (SIL3, CN-RK2/3/4, Euler), step filters, trajectories (plain loops), digital filter initialization
filtering.py exponential / horizontal-diffusion spectral filters
vertical_interpolation.py PressureCoordinates / PressureLevels, pressure ↔ sigma regridding (batched searchsorted/gather, no vmap)
horizontal_interpolation.py conservative / bilinear / nearest lat-lon regridders (weights precomputed as buffers)
radiation.py top-of-atmosphere incident solar radiation (SolarRadiation module)
scales.py, units.py unit handling / nondimensionalization (NumPy + pint)
xarray_utils.py ERA5-style dataset preparation: regrid_horizontal, fill_nan_with_nearest, selective_temporal_shift, grid_spec_from_dataset
pytree.py tiny helpers over torch.utils._pytree

Both spherical-harmonics layouts are implemented because published NeuralGCM checkpoints use both: RealSphericalHarmonics (modal shape (2M-1, L), the 2.8° deterministic checkpoint) and FastSphericalHarmonics (zero-imag layout, modal shape (2M, L), e.g. the TL63 stochastic checkpoint; named RealSphericalHarmonicsWithZeroImag upstream).

Status

The dycore and data path are complete and numerically validated against the original JAX implementation — transforms, operators, the full primitive-equations step (dry and moist, including a 10-step baroclinic-wave trajectory), vertical/horizontal regridding, and solar radiation all match to 1e-5–1e-4 of each field's range — alongside 141 unit tests (pytest). A full SIL3 time step compiles with torch.compile(fullgraph=True) out of the box — no shim rework, no graph breaks.

Not ported: shallow water, Held–Suarez, hybrid coordinates, and leapfrog steppers (intentionally out of scope — no published NeuralGCM checkpoint uses them).

Metadata

Release files for dinosaur-torch 0.1.1

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for dinosaur-torch 0.1.1
File Size Uploaded Provenance Actions
dinosaur_torch-0.1.1.tar.gz 73.1 kB Without provenance Details for dinosaur_torch-0.1.1.tar.gz

Built distribution (wheel)

Table of built distributions (wheels) for dinosaur-torch 0.1.1
File Interpreter ABI Platform Actions
dinosaur_torch-0.1.1-py3-none-any.whl Python 3 none any Details for dinosaur_torch-0.1.1-py3-none-any.whl

Total release size: 138.3 kB

Release files / dinosaur_torch-0.1.1.tar.gz

Download URL dinosaur_torch-0.1.1.tar.gz
Size 73.1 kB
Tags Source
SHA-256 checksum
How to use checksums
e76b10615c7132194305f7844e4f56b74894fcc53fc664856a21ebe5c842a163
BLAKE2b-256 checksum
How to use checksums
410ed862dbcdc803175d2e094a6d4203eee6477546fe7836469278589fdcfa72
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via uv/0.11.19 {"installer":{"name":"uv","version":"0.11.19","subcommand":["publish"]},"python":null,"implementation":{"name":null,"version":null},"distro":{"name":"Ubuntu","version":"24.04","id":"noble","libc":null},"system":{"name":null,"release":null},"cpu":null,"openssl_version":null,"setuptools_version":null,"rustc_version":null,"ci":null}

Release files / dinosaur_torch-0.1.1-py3-none-any.whl

Download URL dinosaur_torch-0.1.1-py3-none-any.whl
Size 65.2 kB
Tags Python 3
SHA-256 checksum
How to use checksums
c282e9c49dad284b805d449c93a2e59cd005272620071846fbb3346ca0bab0fe
BLAKE2b-256 checksum
How to use checksums
ce3e47fbec417725bf96938ce5dff410fe387f08ce1ac87ac4aca119c680eedb
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via uv/0.11.19 {"installer":{"name":"uv","version":"0.11.19","subcommand":["publish"]},"python":null,"implementation":{"name":null,"version":null},"distro":{"name":"Ubuntu","version":"24.04","id":"noble","libc":null},"system":{"name":null,"release":null},"cpu":null,"openssl_version":null,"setuptools_version":null,"rustc_version":null,"ci":null}

Security for dinosaur-torch 0.1.1

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

No provenance found

None of the files in this release were uploaded with provenance metadata.

A note on package security

PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.

To learn more about protecting yourself from malicious actors, visit our package security documentation.

Release history Release notifications | RSS feed

This release

0.1.1 This release

2 release files No provenance

Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page