Remote Graphcast

Run graphcast on a runpod GPU server. Output is saved to s3. Shouldn't cost more than $0.6 USD for a 10 day forcast.

Order of operations

1. Your input is validated
2. A secure runpod GPU pod is spun up on your runpod account
3. Graphcast is installed into that gpu and forecasts of your chosen length are generated for each timestamp, this takes around 10 minutes for a 10 day forcast
4. These forecasts are saved to your chosen s3 bucket, roughly 6.5GB for 10 days of forecast
5. The runpod pod is terminated
6. The program exits

Requirements

• python 3.8+ and pip
• cds.climate credentials
• an AWS s3 bucket, free tier should be fine
• S3 credentials to go with the bucket
• Runpod credentials

Installation

pip install remote-graphcast

Example Code

from remote_graphcast import remote_cast

remote_cast(
	aws_access_key_id='YOUR_AWS_ACCESS_KEY_ID',
	aws_secret_access_key='YOUR_AWS_SECRET_ACCESS_KEY',
	aws_bucket='YOUR_AWS_BUCKET_NAME',
	cds_url='https://cds.climate.copernicus.eu/api/v2', # this is probably your CDS URL 
	cds_key='YOUR_CDS_KEY',
	forcast_list="[{'start': '2023122518', 'hours_to_forcast': 48}]", 
	# dates to forcast from, note the weird quasi-JSON format, of this string, use single quotes instead of double quotes
	# select a date in the future and it will raise an error without spinning up anything
	runpod_key='YOUR_RUNPOD_KEY',
	gpu_type_id="NVIDIA A100-SXM4-80GB", # graphcast needs at least 61GB GPU ram (unless you want to quantize)
	container_disk_in_gb=50, # you'll need around 40GB per 10 day forcast + a healthy 10GB buffer
)

# internally this function will keep polling the pod it spins up and the s3 bucket until it sees that all forcasts 
# are complete, then it will return

Warning

In order to make predictions graphcast must request ERA5 reanalysis data from the European Center for Medium Range Weather Forcasts (ECMWF). Usually the download completes in < 2 minutes. However, if their servers are busy your request will be put in a queue. You can view all your open requests here. Until your request is granted is granted, the graphcast runpod server will be waiting idly (costing you money). This process has taken me > 1 hour in the past.

You can also check your progress by viewing the runpod logs.