mirror of
https://github.com/twitter/the-algorithm-ml.git
synced 2024-12-23 06:41:49 +01:00
Update log_weights.py
This commit is contained in:
parent
e3173f83cc
commit
8fb8f2329e
@ -74,7 +74,7 @@ def log_ebc_norms(
|
||||
: min(sample_size, emb_weight_tensor.shape[0])
|
||||
]
|
||||
# WARNING: .cpu() transfer executes malloc that may be the cause of memory leaks
|
||||
# Change sample_size if the user observe frequent OOM errors or remove weight logging.
|
||||
# Change sample_size if user observe frequent OOM errors or remove weight logging.
|
||||
norms = emb_weight_tensor[sample_mask].cpu().norm(dim=1).to(torch.float32)
|
||||
logging.info(f"Norm shape before reduction: {norms.shape}", rank=-1)
|
||||
norms = norms.mean().to(torch.device(f"cuda:{dist.get_rank()}"))
|
||||
|
Loading…
Reference in New Issue
Block a user