Open conceptofmind opened 18 hours ago
Indeed. This is something @kevinzwang is working on stabilizing our solution for.
You can actually try it out with the environment variable: DAFT_ENABLE_ACTOR_POOL_PROJECTIONS =1
The fix is currently available for the local executor. We have some local branches available for the Ray Runner as well and are in the process of testing/stabilizing both. Let us know if you try it!
I see you're using a GPU as well in your UDF. We will probably want to correctly assign the CUDA_VISBLE_DEVICES
appropriately for each instance of your UDF which isn't yet being done on the PyRunner.
Indeed. This is something @kevinzwang is working on stabilizing our solution for.
You can actually try it out with the environment variable:
DAFT_ENABLE_ACTOR_POOL_PROJECTIONS =1
The fix is currently available for the local executor. We have some local branches available for the Ray Runner as well and are in the process of testing/stabilizing both. Let us know if you try it!
Hi @jaychia ,
I will test out setting the environment variable for cuda visible devices and actor pool.
Thank you!
Made a PR for an initial attempt at doing CUDA_VISIBLE_DEVICES
: https://github.com/Eventual-Inc/Daft/pull/2882
You'll likely need that if running multi-GPU on a single node + PyRunner!
Given the embedding udf below, the model re-loads and is reinitializing after each write and completed partition / parquet file:
Model reloads after completing and writing one partition/file:
Any input would be greatly appreciated.