tocmo0nlord/axolotl
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

fix flaky tests; should be using train loss from final step rather than final avg train loss

Browse Source
This commit is contained in:
Wing Lian
2026-03-22 10:38:46 -04:00
parent 5b2e3f00ce
commit 6130e40c37
21 changed files with 37 additions and 41 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
tests/e2e/multigpu/test_fp8_fsdp2.py
View File

@@ -37,7 +37,7 @@ def verify_fp8_training_success(temp_dir):
event_file = os.path.join(tb_log_path, event_files[0])
reader = SummaryReader(event_file)
df = reader.scalars
train_loss_df = df[df.tag == "train/train_loss"]
train_loss_df = df[df.tag == "train/loss"]
if len(train_loss_df) > 0:
final_loss = train_loss_df.value.values[-1]
assert not torch.isnan(torch.tensor(final_loss)), (