File size: 1,601 Bytes
4c95a77 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 | ---
license: mit
library_name: pytorch
tags:
- weather
- meteorology
- diffusion
- forecast
- hrrr
---
# HRRRCast V3 Torch Port
This repository contains converted PyTorch checkpoints for the HRRRCast V3 diffusion model ported from the NOAA GSL Keras checkpoint.
## What this is
- Converted Torch checkpoints for the HRRRCast diffusion network
- Intended to be used together with the GitHub code repository
- Supports `fp32`, `fp16`, and `bf16` checkpoints
## What this is not
- Not a `transformers` model
- Not a `diffusers` pipeline
- Not a browser widget model
- Not a lightweight consumer inference package
## Required code
Use these checkpoints with the corresponding GitHub repository that contains:
- `torch_port/`
- `src/`
- `net-diffusion/model.config.json`
The loader reconstructs the Torch graph from the extracted Keras `config.json` plus the uploaded Torch state dict.
## Example
```bash
PYTHONPATH=. ./.venv/bin/python -m torch_port.forecast \
converted/hrrrcast_diffusion_bf16.pt \
2026-03-26T19 \
6 \
--members 0 \
--base_dir /path/to/preprocessed_npz \
--output_dir /path/to/output \
--device cuda \
--dtype bf16 \
--tile_size 96,96 \
--tile_halo 32
```
## Runtime notes
- Full CONUS HRRR grid: `1059 x 1799`
- Diffusion sampling is expensive
- Tiled inference is required on typical single-GPU setups
- Single-member tiled `bf16` inference can still take minutes per forecast hour
## Intended use
Research use, port validation, inference experiments, quantization work, and downstream model distillation.
## License
MIT. See `LICENSE`.
|