TornadoGCM

TornadoGCM is a pure-PyTorch reimplementation of NeuralGCM (Google Research) featuring a novel Precision-Zoned Hybrid Architecture (PZHA) with automated mixed-precision scheduling for GPU-accelerated global weather forecasting.

Disclaimer: This project is an independent research port. It is not affiliated with or endorsed by Google Research or DeepMind. All trademarks belong to their respective owners.

Features

• Pure PyTorch — full GPU support, DDP/FSDP distributed training, torch.compile compatibility
• PZHA mixed-precision — four precision zones (Z0 FP64, Z1 TF32, Z2 FP64, Z3 BF16) matched to physical/neural computation roles
• SDA control plane — Self-adaptive Dynamic Accuracy scheduler automatically adjusts precision zones during training
• JAX checkpoint loader — load pre-trained NeuralGCM JAX weights directly, no manual conversion script needed
• Triton acceleration — optional fused spherical-harmonic transform kernels for A100/H100
• Multiple resolutions — tested configs for 0.7°, 1.4°, and 2.8° grids

Architecture Overview

┌─────────────────────────────────────────────────────────────────────────┐
│                         AtmosphereModel (facade)                        │
├───────────────┬─────────────────────────┬───────────────────────────────┤
│  Encoder      │   NeuralGCMModel (step)  │   Decoder                    │
│  Z3 (BF16)    │  ┌──────────────────┐   │   Z3 (BF16)                  │
│               │  │ Dycore  Z1 (TF32)│   │                              │
│               │  │ Neural  Z3 (BF16)│   │                              │
│               │  │ Fixer   Z2 (FP64)│   │                              │
│               │  └──────────────────┘   │                              │
└───────────────┴─────────────────────────┴───────────────────────────────┘
                              ↕ SDA Controller (precision scheduler)

Installation

# Core (CPU/GPU)
pip install tornado-gcm

# With data I/O (xarray, zarr)
pip install "tornado-gcm[data]"

# With Triton GPU kernels
pip install "tornado-gcm[gpu]"

# Development
pip install "tornado-gcm[dev]"
pre-commit install

Requirements: Python ≥ 3.9, PyTorch ≥ 2.1, CUDA 11.8+ (for GPU).

Quick Start (5 minutes)

1. Load a pre-trained JAX checkpoint and run a forecast

import torch
from tornado_gcm.model.jax_checkpoint_loader import load_jax_checkpoint
from tornado_gcm.inference.runner import MPInference
from tornado_gcm.precision.sda import SDAConfig

# Load SDA config (choose resolution)
sda_cfg = SDAConfig.from_yaml("configs/sda_neuralgcm_2.8deg_full.yaml")

# Convert JAX weights → PyTorch
model = load_jax_checkpoint("path/to/neuralgcm_checkpoint", device="cuda")

# Build mixed-precision inference runner
runner = MPInference(model=model, device="cuda")

# Run 10-day (40-step × 6h) forecast
# initial_state: a State dataclass on GPU
trajectory = runner.rollout(initial_state, steps=40)

2. CLI

# Inference
tornado-gcm infer \
  --config configs/sda_neuralgcm_1.4deg.yaml \
  --checkpoint model.pt \
  --input era5_20200101.zarr \
  --output forecast.zarr \
  --steps 40

# Benchmark
tornado-gcm benchmark \
  --config configs/sda_neuralgcm_2.8deg_full.yaml \
  --steps 20 \
  --report bench.txt

# Training (requires ERA5 data)
tornado-gcm train \
  --config configs/sda_neuralgcm_1.4deg.yaml \
  --jax-weights neuralgcm_checkpoint \
  --steps 50000 \
  --checkpoint output/run1.pt

Performance

Tested on a single NVIDIA A100 80 GB (CUDA 12.1, PyTorch 2.2):

Note: JAX numbers from original paper on TPU v4-8 pod.

Project Structure

tornado_gcm/
├── core/          # Spectral dynamics, coordinates, time integration (Z0/Z1)
├── neural/        # NN parameterizations, encoders/decoders, towers (Z3)
├── model/         # Model assembly, JAX checkpoint loader
├── precision/     # PZHA policy, SDA controller, Triton accelerator
├── training/      # Trainers, losses, evaluation, data loading
├── inference/     # MPInference, SDAInference, production runner
├── distributed/   # DDP / DTensor sharding
├── configs/       # Ready-to-use SDA YAML configs (0.7°/1.4°/2.8°)
└── cli.py         # Click CLI entry point

Attribution

This project is a PyTorch port of NeuralGCM by Kochkov et al.:

Kochkov, D., Yuval, J., Langmore, I., Norgaard, P., Smith, J., Mooers, G., ... & Hoyer, S. (2024). Neural general circulation models for weather and climate. Nature, 632, 1060–1066. https://doi.org/10.1038/s41586-024-07744-y

Original JAX code: https://github.com/google-research/neuralgcm (Apache 2.0)

The PZHA mixed-precision framework, SDA control plane, and PyTorch-specific optimizations are original contributions of this project.

Contributing

See CONTRIBUTING.md.

License

• Source code: Apache License 2.0
• Pre-trained weights (if distributed): CC BY-SA 4.0