It takes a lot of time to resume training after running evaluation [FSDP]. What can be the reason behind this? #20257
Unanswered
psr-ai
asked this question in
DDP / multi-GPU / multi-node
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
-
My training processes runs as follows:
----- 1. Sanity Check Validation ------
----- 2. Training ---------------------
----- 3. Validation -------------------
----- 4. Training ---------------------
My model training transitions from Sanity Check Validation to Training in no amount of time (Step 1 to 2), but from Validation to Training (Step 3 to 4) it takes a lot of time. I am training a large language model under FSDP strategy. What can be the reason?
Beta Was this translation helpful? Give feedback.
All reactions