Community for the Daft project and all things distributed data

Distributed Data Community

hi, i had a fairly noob question
for the UDFs when specifying `num_gpus=` is there a way I can spawn 8 processes to utilize all 8 GPUs I have on 1 H100 node without starting a local ray cluster?
(by spawing local processes instead perhaps? not sure what the underlying implementation is)
when i tried with:
> @daft.udf(return_dtype=daft.DataType.python(), num_gpus=8)
my `nvidia-smi` showed only 1 GPU being used

Hi Sammie, the num_gpus argument is actually to specify the number of GPUs to use per UDF task. So if youd like for the execution of your UDF to each use 1 GPU, you should specify num_gpus=1 instead and we’ll run an instance of the UDF on each of your 8 available nodes!

Also, if you’d like UDFs to run on separate processes or initialize only once, you should look into <https://www.getdaft.io/projects/docs/en/stable/api_docs/udf.html#daft.udf.UDF.with_concurrency|with_concurrency>